Scolaris Content Display Scolaris Content Display

Pharmacological interventions for the treatment of delirium in critically ill adults

Collapse all Expand all

Background

Although delirium is typically an acute reversible cognitive impairment, its presence is associated with devastating impact on both short‐term and long‐term outcomes for critically ill patients. Advances in our understanding of the negative impact of delirium on patient outcomes have prompted trials evaluating multiple pharmacological interventions. However, considerable uncertainty surrounds the relative benefits and safety of available pharmacological interventions for this population.

Objectives

Primary objective

1. To assess the effects of pharmacological interventions for treatment of delirium on duration of delirium in critically ill adults with confirmed or documented high risk of delirium

Secondary objectives

To assess the following:

1. effects of pharmacological interventions on delirium‐free and coma‐free days; days with coma; delirium relapse; duration of mechanical ventilation; intensive care unit (ICU) and hospital length of stay; mortality; and long‐term outcomes (e.g. cognitive; discharge disposition; health‐related quality of life); and

2. the safety of such treatments for critically ill adult patients.

Search methods

We searched the following databases from their inception date to 21 March 2019: Ovid MEDLINE®, Ovid MEDLINE® In‐Process & Other Non‐Indexed Citations, Embase Classic+Embase, and PsycINFO using the Ovid platform. We also searched the Cochrane Library on Wiley, the International Prospective Register of Systematic Reviews (PROSPERO) (http://www.crd.york.ac.uk/PROSPERO/), the Cumulative Index to Nursing and Allied Health Literature (CINAHL), and Web of Science. We performed a grey literature search of relevant databases and websites using the resources listed in Grey Matters developed by the Canadian Agency for Drugs and Technologies in Health (CADTH). We also searched trial registries and abstracts from annual scientific critical care and delirium society meetings.

Selection criteria

We sought randomized controlled trials (RCTs), including quasi‐RCTs, of any pharmacological (drug) for treatment of delirium in critically ill adults. The drug intervention was to be compared to another active drug treatment, placebo, or a non‐pharmacological intervention (e.g. mobilization). We did not apply any restrictions in terms of drug class, dose, route of administration, or duration of delirium or drug exposure. We defined critically ill patients as those treated in an ICU of any specialty (e.g. burn, cardiac, medical, surgical, trauma) or high‐dependency unit.

Data collection and analysis

Two review authors independently identified studies from the search results; four review authors (in pairs) performed data extraction and assessed risk of bias independently. We performed data synthesis through pairwise meta‐analysis and network meta‐analysis (NMA). Our hypothetical network structure was designed to be analysed at the drug class level and illustrated a network diagram of 'nodes' (i.e. drug classes) and 'edges' (i.e. comparisons between different drug classes from existing trials), thus describing a treatment network of all possible comparisons between drug classes. We assessed the quality of the body of evidence according to GRADE, as very low, low, moderate, or high.

Main results

We screened 7674 citations, from which 14 trials with 1844 participants met our inclusion criteria. Ten RCTs were placebo‐controlled, and four reported comparisons of different drugs. Drugs examined in these trials were the following: antipsychotics (n = 10), alpha2 agonists (n = 3; all dexmedetomidine), statins (n = 2), opioids (n = 1; morphine), serotonin antagonists (n = 1; ondansetron), and cholinesterase (CHE) inhibitors (n = 1; rivastigmine). Only one of these trials consistently used non‐pharmacological interventions that are known to improve patient outcomes in both intervention and control groups.

Eleven studies (n = 1153 participants) contributed to analysis of the primary outcome. Results of the NMA showed that the intervention with the smallest ratio of means (RoM) (i.e. most preferred) compared with placebo was the alpha2 agonist dexmedetomidine (0.58; 95% credible interval (CrI) 0.26 to 1.27; surface under the cumulative ranking curve (SUCRA) 0.895; moderate‐quality evidence). In order of descending SUCRA values (best to worst), the next best interventions were atypical antipsychotics (RoM 0.80, 95% CrI 0.50 to 1.11; SUCRA 0.738; moderate‐quality evidence), opioids (RoM 0.88, 95% CrI 0.37 to 2.01; SUCRA 0.578; very‐low quality evidence), and typical antipsychotics (RoM 0.96, 95% CrI 0.64 to1.36; SUCRA 0.468; high‐quality evidence).

The NMAs of multiple secondary outcomes revealed that only the alpha2 agonist dexmedetomidine was associated with a shorter duration of mechanical ventilation (RoM 0.55, 95% CrI 0.34 to 0.89; moderate‐quality evidence), and the CHE inhibitor rivastigmine was associated with a longer ICU stay (RoM 2.19, 95% CrI 1.47 to 3.27; moderate‐quality evidence). Adverse events often were not reported in these trials or, when reported, were rare; pair‐wise analysis of QTc prolongation in seven studies did not show significant differences between antipsychotics, ondansetron, dexmedetomidine, and placebo.

Authors' conclusions

We identified trials of varying quality that examined six different drug classes for treatment of delirium in critically ill adults. We found evidence that the alpha2 agonist dexmedetomidine may shorten delirium duration, although this small effect (compared with placebo) was seen in pairwise analyses based on a single study and was not seen in the NMA results. Alpha2 agonists also ranked best for duration of mechanical ventilation and length of ICU stay, whereas the CHE inhibitor rivastigmine was associated with longer ICU stay. We found no evidence of a difference between placebo and any drug in terms of delirium‐free and coma‐free days, days with coma, physical restraint use, length of stay, long‐term cognitive outcomes, or mortality. No studies reported delirium relapse, resolution of symptoms, or quality of life. The ten ongoing studies and the six studies awaiting classification that we identified, once published and assessed, may alter the conclusions of the review.

PICOs

Population
Intervention
Comparison
Outcome

The PICO model is widely used and taught in evidence-based health care as a strategy for formulating questions and search strategies and for characterizing clinical studies or meta-analyses. PICO stands for four different potential components of a clinical question: Patient, Population or Problem; Intervention; Comparison; Outcome.

See more on using PICO in the Cochrane Handbook.

Medicines to treat delirium in critically ill adult patients

Review question

We reviewed the evidence from randomized controlled trials for the benefits and safety of all prescription medicines used to treat critically ill adult patients with delirium in the intensive care units (ICUs) of hospitals.

Background

Delirium is commonly associated with surgery, infection, or critical illness. It is experienced as new‐onset, generally short‐term inability to think clearly. Patients with delirium shift between periods of clear thinking and periods of agitation and/or great sleepiness and confusion. Lack of sleep, pain, a noisy environment, physical restraint, and the use of sedatives and strong analgesics are some of the contributing factors. Delirium affects both immediate and longer‐term health outcomes of critically ill patients as it can increase the length of time a breathing machine is required, time spent in the ICU and in hospital, and the chance of functional weakening and death. The odds of a poor outcome with delirium are increased with frail patients and those of advanced age and already present cognitive difficulties. Frequently, delirious ICU patients are given medicines to help treat symptoms such as agitation.

Study characteristics

This review is current to 21 March 2019. We found 14 randomized controlled studies that enrolled a total of 1844 adult participants. Six different classes of medicines were tested. These were antipsychotic drugs used as tranquillizers in ten studies; the sedative alpha2 agonist dexmedetomidine in three studies; statins that reduce cholesterol in two studies; opioids as part of pain management in one study; serotonin antagonists for nausea and vomiting in one study; and cholinesterase inhibitors, which are medicines for Alzheimer's disease, in one study. Ten studies compared medicine to placebo ‐ an inactive medicine also known as a sugar pill; four studies compared different drugs. Eleven studies with 1153 participants reported on the main outcome of this review ‐ duration of delirium.

Key findings

When drug classes were directly compared with placebo, only the alpha2 agonist dexmedetomidine was found to reduce the duration of delirium, and the cholinesterase inhibitor rivastigmine was found to prolong the duration of delirium. Each of these results is based on findings from a single small study. The other drugs when compared to placebo did not change delirium duration. The Review authors used the statistical method of network meta‐analysis to compare the six different drug classes. Dexmedetomidine was ranked most effective in reducing delirium duration, followed by atypical antipsychotics. However, network meta‐analysis of delirium duration failed to rule out the possibility of no difference for all six drug classes compared to placebo. Using this method, we did not find that any drug improved the duration of coma, length of stay, long‐term cognitive outcomes, or death. The alpha2 agonist dexmedetomidine shortened time spent on a breathing machine. Adverse events often were not reported in these trials or were rare when reported. An analysis of reported events showed that events were similar to those reported with placebo. We found 10 ongoing studies and six studies awaiting classification that, once published and assessed, may change the conclusions of this review.

Quality of the evidence

Most of the included studies were small but of good design. Nine of the 14 studies were considered to have low risk of bias.

Authors' conclusions

available in

Implications for practice

In clinical practice, pharmacological interventions are commonly administered to critically ill patients to manage their symptoms of delirium (Burry 2017). We found evidence that the alpha2 agonist dexmedetomidine may have some role in shortening delirium duration, although this small effect was seen in pairwise analyses based on a single small study compared with placebo, and was not seen in the NMA results. No other pharmacological intervention including antipsychotics, the most commonly prescribed drug for delirium treatment, had any effect on delirium duration nor on any of our a priori selected secondary outcomes. It is also important to note that the cholinesterase inhibitor rivastigmine was associated with harm, and as such, guidelines suggest against its use for treatment of ICU delirium. The 10 ongoing studies and the six studies awaiting classification, once published and assessed, may alter the conclusions of this review; therefore, their results are much anticipated. The frequency of prescribing these drug classes for critically ill adults with delirium and the non‐significant findings of our review should be considered at the bedside and should be incorporated into future pain, agitation, and delirium guidelines.

Implications for research

We identified 10 ongoing studies, of which seven have a large target enrolment number (100 to 1000 participants), suggesting growing interest in the treatment of ICU delirium. These RCTs should strengthen our results and may potentially alter the direction of our findings. For example, five ongoing trials are examining antipsychotics and three are examining the alpha2 agonist dexmedetomidine ‐ the drug classes found most promising in our analysis ‐ each trial with large target enrolment.

We note the promise of many new treatment trials on the horizon; however, we must acknowledge the need to standardize outcome reporting in ICU delirium trials to permit maximum pooling and interpretation of results. We found far greater variability in the definitions of study outcomes used than we had anticipated, which led us to modify our primary outcome and to limit pooling for some outcomes (e.g. mortality). We found no reporting on some clinically important outcomes such as symptom management (e.g. treating agitation, stopping treatment interferences) and long‐term cognitive outcomes, and we found new outcomes not listed in our protocol (e.g. number of days in coma) in multiple new RCTs and ongoing trials. The Del‐COrS ("Developmnt of core outcome sets for effectiveness trial of interventions to prevent and/or treat delirium") Group is leading the development of international consensus on outcomes for trials of intervention to prevent and treat delirium in multiple patient populations (Rose 2017). Findings from this group should be used to guide future ICU delirium trials.

We also found that RCTs in this review rarely reported on the use of non‐pharmacological strategies. Among the trials that we identified, all but one showed poor utilization of non‐pharmacological strategies. For example, early mobilization has been shown to reduce the duration of delirium (Barr 2013), and its use in practice is encouraged. Therefore, future trials should clearly describe the use of such strategies in their methods and should report compliance in their results. We also found poor reporting on the use of physical restraints ‐ a non‐pharmacological intervention associated with delirium and prolonged duration of delirium (Rose 2016).

Summary of findings

Open in table viewer
Summary of findings for the main comparison. Duration of delirium

Outcome: duration of delirium

Patient or population: critically ill adults with confirmed or at high risk of delirium
Settings: intensive care units in Australia and New Zealand, Canada, Egypt, Netherlands, Turkey, USA, UK
Intervention: any pharmacological intervention

Control: placebo or active comparator

Comparisons

Illustrative comparative risks* (95% CrI)

Ratio of means (RoM) based on log RoM estimates from meta‐analysis

(IV, random, 95% CI)

Number of participants
(studies)

Quality of the evidence
(GRADE) based on NMA

NMA results (assuming consistency equations)

Assumed risk

Corresponding risk based on NMA estimates

Placebo/Comparator

Intervention drug

Typical antipsychotic vs placebo

Median duration of delirium: 3 to 5 days for placebo

3.86 days of delirium (95% CrI 2.57 to
5.46) corresponding to 4 days in the placebo group

RoM: exp(0.02) = 1.02 (95% CI 0.91 to 1.14); log RoM: 0.02 (‐0.09 to 0.13); I² = 0%

608
(4 studies)

⊕⊕⊕⊕
High

RoM (95% CrI): 0.96
(0.64 to 1.36),
SUCRA = 0.468, mean Pr(best) = 0.010, mean rank = 4.19

Atypical antipsychotic vs placebo

Median duration of delirium: 3 to 5 days for placebo

3.22 days of delirium (95% CrI 2.01 to
4.43) corresponding to 4 days in the placebo group

RoM: exp(‐0.31) = 0.73 (95% CI 0.49 to 1.11); log RoM: ‐0.31 (‐0.71 to 0.10); I² = 82%

500
(4 studies)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 0.80
(0.50 to 1.11),
SUCRA = 0.738, mean Pr(best) = 0.114, mean rank = 2.57

Statin (HMG‐CoA)

vs placebo

Mean duration of delirium: 6.8 to 8.68 days for placebo

4.20 days of delirium (95% CrI 2.44 to
7.09) corresponding to 4 days in the placebo group

RoM: exp(0.07) = 1.07 (95% CI 0.91 to 1.25); log RoM: 0.07 (‐0.09 to 0.22); I² = 0%

414
(2 studies)

⊕⊕⊕⊝
Moderateb

RoM (95% CrI): 1.05
(0.61 to 1.77),
SUCRA = 0.365, mean Pr(best) = 0.023, mean rank = 4.81

Alpha2 agonist

vs placebo

Median duration of delirium: 2.583 days for placebo

2.31 days of delirium (95% CrI 1.06 to
5.06) corresponding to 4 days in the placebo group

RoM: exp(‐0.55) = 0.58 (95% CI 0.43 to 0.79); log RoM: ‐0.55 (‐0.85 to ‐0.24); I² not applicable

71
(1 study)

⊕⊕⊕⊝
Moderateb

RoM (95% CrI): 0.58
(0.26 to 1.27),
SUCRA = 0.895, mean Pr(best) = 0.717, mean rank = 1.63

Cholinesterase inhibitor

vs placebo

Median duration of delirium: 3 days for placebo

7.37 days of delirium (95% CrI 3.26 to
16.38) corresponding to 4 days in the placebo group

RoM: exp(0.61) = 1.84 (95% CI 1.25 to 2.69); log RoM: 0.61 (0.22 to 0.99); I² not applicable

104
(1 study)

⊕⊕⊕⊝
Moderateb

RoM (95% CrI): 1.84
(0.82 to 4.10),
SUCRA = 0.054, mean Pr(best) = 0.006, mean rank = 6.68

Opioid

vs placebo

No study reported this comparison

3.53 days of delirium (95% CrI 1.46 to

8.05) corresponding to 4 days in the placebo group

Pairwise meta‐analysis not performed

0
(0 studies)

⊕⊕⊝⊝
Very lowb,c

RoM (95% CrI): 0.88
(0.37 to 2.01),
SUCRA = 0.578, mean Pr(best) = 0.129, mean rank = 3.53

*The basis for the assumed risk (e.g. the median control group risk across studies). The corresponding risk (and its 95% CrI) is calculated as the assumed risk multiplied by the ratio of means (and its 95% CrI) based on NMA.

Abbreviations: CI: confidence interval; CrI: credible interval; HMG‐CoA: 5‐hydroxy‐3‐methylglutaryl‐coenzyme A reductase inhibitor; NMA: network meta‐analysis; Pr(best): probability(best); RoM: ratio of means; SUCRA: surface under the cumulative ranking curve.

GRADE Working Group grades of evidence.
High quality: further research is very unlikely to change our confidence in the estimate of effect.
Moderate quality: further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate.
Low quality: further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate.
Very low quality: we are very uncertain about the estimate.

aDowngraded one level for heterogeneity (I² > 75% considered as large heterogeneity).
bDowngraded one level for imprecision (wide credible interval).
cDowngraded two levels for only indirect evidence available and risk of bias of a single trial informing opioid vs typical antipsychotic.

Open in table viewer
Summary of findings 2. Days with coma

Outcome: days with coma

Patient or population: critically ill adult with confirmed or at high risk of delirium
Settings: intensive care units in Australia and New Zealand, Canada, Egypt, Netherlands, Turkey, USA, UK
Intervention: any pharmacological intervention

Control: placebo or active comparator

Comparisons

Illustrative comparative risks* (95% CI)

Ratio of means (RoM) based on log RoM estimates from meta‐analysis

(IV, random, 95% CI)

Number of participants
(studies)

Quality of the evidence
(GRADE) based on NMA

NMA results
(assuming
consistency equations)

Assumed risk

Corresponding risk based on NMA estimates

Placebo/Comparator

Intervention drug

Typical antipsychotic vs placebo

Median number of days with coma: 1 to 2 days for placebo

1.53 days with coma (95% CrI 0.86 to
2.57) corresponding to 2 days in the placebo group

RoM: exp(‐0.29) = 0.75 (95% CI 0.49 to 1.13);
log RoM: ‐0.29 (‐0.71 to 0.12); I² = 74%

588
(3 studies)

⊕⊕⊝⊝
Lowa,b

RoM (95% CrI): 0.77
(0.43 to 1.29),
SUCRA = 0.820, mean Pr(best) = 0.620, mean rank = 1.54

Atypical antipsychotic

vs placebo

Median number of days with coma: 1 to 2 days for placebo

1.88 days with coma (95% CrI 0.96 to
3.43) corresponding to 2 days in the placebo group

RoM: exp(0.06) = 1.06 (95% CI 0.88 to 1.30);
log RoM: 0.06 (‐0.13 to 0.26); I² = 0%

440
(2 studies)

⊕⊕⊕⊝
Moderateb

RoM (95% CrI): 0.94
(0.48 to 1.72),
SUCRA = 0.422, mean Pr(best) = 0.132, mean rank = 2.73

Statin (HMG‐CoA)

vs placebo

Mean number of days with coma: 1.1 to 4.2 days for placebo

1.84 days with coma (95% CrI 0.98 to
3.59) corresponding to 2 days in the placebo group

RoM: exp(‐0.10) = 0.90 (95% CI 0.73 to 1.12);
log RoM: ‐0.10 (‐0.32 to 0.11); I² = 0%

414
(2 studies)

⊕⊕⊕⊝
Moderateb

RoM (95% CrI): 0.92
(0.49 to 1.80),
SUCRA = 0.481, mean Pr(best) = 0.222, mean rank = 2.56

*The basis for the assumed risk (e.g. the median control group risk across studies). The corresponding risk (and its 95% CrI) is calculated as the assumed risk multiples the ratio of means (and its 95% CrI) based on NMA.
CI: confidence interval; CrI: credible interval; HMG‐CoA: 5‐hydroxy‐3‐methylglutaryl‐coenzyme A reductase inhibitor; NMA: network meta‐analysis; Pr(best): probability(best); RoM: ratio of means; SUCRA: surface under the cumulative ranking curve.

GRADE Working Group grades of evidence.
High quality: further research is very unlikely to change our confidence in the estimate of effect.
Moderate quality: further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate.
Low quality: further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate.
Very low quality: we are very uncertain about the estimate.

aDowngraded one level for heterogeneity (I² of 50% to 75%, > 75% considered as medium and large heterogeneity).
bDowngraded one level for imprecision (wide credible interval).

Open in table viewer
Summary of findings 3. Duration of mechanical ventilation

Outcome: duration of mechanical ventilation

Patient or population: critically ill adult with confirmed or at high risk of delirium
Settings: intensive care units in Australia and New Zealand, Canada, Egypt, Netherlands, Turkey, USA, UK
Intervention: any pharmacological intervention

Control: placebo or active comparator

Comparisons

Illustrative comparative risks* (95% CI)

Ratio of means (RoM) based on log RoM estimates from meta‐analysis

(IV, random, 95% CI)

Number of participants
(studies)

Quality of the evidence
(GRADE) based on NMA

NMA results (assuming
consistency equations)

Assumed risk

Corresponding risk based on NMA estimates

Placebo/Comparator

Intervention drug

Typical antipsychotics
vs placebo

Median duration of mechanical ventilation: 3 to 5 days for placebo

3.71 days of mechanical ventilation (95% CrI 2.89 to 4.94) corresponding to 4 days in the placebo group

RoM: exp(‐0.08) = 0.92 (95% CI 0.79 to 1.06); log RoM: ‐0.08 (‐0.23 to 0.06); I² = 0%

515
(3 studies)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 0.93
(0.72 to 1.24),
SUCRA = 0.576, mean Pr(best) = 0.009, mean rank = 3.12

Atypical antipsychotics
vs placebo

Median duration of mechanical ventilation: 3 to 11 days for placebo

3.91 days of mechanical ventilation (95% CrI 2.85 to 5.10) corresponding to 4 days in the placebo group

RoM: exp(‐0.02) = 0.98 (95% CI 0.84 to 1.34); log RoM: ‐0.02 (‐0.17 to 0.14); I² = 0%

476
(3 studies)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 0.98
(0.71 to 1.28),
SUCRA = 0.440, mean Pr(best) = 0.012, mean rank = 3.80

Statin (HMG‐CoA)
vs placebo

Mean duration of mechanical ventilation: 11 days for placebo

4.38 days of mechanical ventilation (95% CrI 2.82 to 6.77) corresponding to 4 days in the placebo group

RoM: exp(0.09) = 1.09 (95% CI 0.90 to 1.34); log RoM: 0.09 (‐0.11 to 0.29); I² not applicable

272
(1 study)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 1.10
(0.71 to 1.69),
SUCRA = 0.223, mean Pr(best) = 0.014, mean rank = 4.88

Alpha2 agonist
vs placebo

Median duration of mechanical ventilation: 1.846 days for placebo

2.21 days of mechanical ventilation (95% CrI 1.36 to 3.58) corresponding to 4 days in the placebo group

RoM: exp(‐0.59) = 0.55 (95% CI 0.41 to 0.75); log RoM: ‐0.59 (‐0.89 to ‐0.29); I² not applicable

71
(1 study)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 0.55
(0.34 to 0.89),
SUCRA = 0.974, mean Pr(best) = 0.931, mean rank = 1.13

Opioid
vs placebo

No study reported this comparison

3.96 days of mechanical ventilation (95% CrI 2.32 to 7.02) corresponding to 4 days in the opioid group

Pairwise meta‐analysis not performed

0

(0 studies)

⊕⊕⊝⊝
Very lowa,b

RoM (95% CrI): 0.99
(0.58 to 1.76),
SUCRA = 0.410, mean Pr(best) = 0.033, mean rank = 3.95

*The basis for the assumed risk (e.g. the median control group risk across studies). The corresponding risk (and its 95% CrI) is calculated as the assumed risk multiples the ratio of means (and its 95% CrI) based on NMA.
CI: confidence interval; CrI: credible interval; HMG‐CoA: 5‐hydroxy‐3‐methylglutaryl‐coenzyme A reductase inhibitor; NMA: network meta‐analysis; Pr(best): probability(best); RoM: ratio of means; SUCRA: surface under the cumulative ranking curve.

GRADE Working Group grades of evidence.
High quality: further research is very unlikely to change our confidence in the estimate of effect.
Moderate quality: further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate.
Low quality: further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate.
Very low quality: we are very uncertain about the estimate.

aDowngraded one level for imprecision (wide credible interval).
bDowngraded two levels for only indirect evidence available and risk of bias of a single trial informing opioid vs typical antipsychotic.

Open in table viewer
Summary of findings 4. Length of ICU stay

Outcome: length of ICU stay

Patient or population: critically ill adult with confirmed or at high risk of delirium
Settings: intensive care units in Australia and New Zealand, Canada, Egypt, Netherlands, Turkey, USA, UK
Intervention: any pharmacological intervention

Control: placebo or active comparator

Comparisons

Illustrative comparative risks* (95% CI)

Ratio of means (RoM) based on log RoM estimates from meta‐analysis

(IV, random, 95% CI)

Number of participants
(studies)

Quality of the evidence
(GRADE) based on NMA

NMA results (assuming
consistency equations)

Assumed risk

Corresponding risk based on NMA estimates

Placebo/Comparator

Intervention drug

Typical antipsychotic
vs placebo

Median length of ICU
stay: 5 to 9 days for placebo

7.92 days of ICU stay (95% CrI 6.79 to
9.37) corresponding to 8 days in the placebo group

RoM: exp(0.01) = 1.01 (95% CI 0.90 to 1.14); log RoM: 0.01 (‐0.11 to 0.13); I² = 0%

618
(4 studies)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 0.99
(0.85 to 1.17),
SUCRA = 0.496, mean Pr(best) = 0.014, mean rank = 4.02

Atypical antipsychotic
vs placebo

Median length of ICU
stay: 3 to 16 days for placebo

7.40 days of ICU stay (95% CrI 6.37 to
8.66) corresponding to 8 days in the placebo group

RoM: exp(‐0.09) = 0.91 (95% CI 0.84 to 1.00); log RoM: ‐0.09 (‐0.18 to ‐0.00); I² = 0%

577
(4 studies)

⊕⊕⊕⊕
High

RoM (95% CrI): 0.92
(0.80 to 1.08),
SUCRA = 0.709, mean Pr(best) = 0.106, mean rank = 2.75

Statin (HMG‐CoA)
vs placebo

Mean length of ICU
stay: 13 days for placebo

8.54 days of ICU stay (95% CrI 6.46 to
11.25) corresponding to 8 days in the placebo group

RoM: exp(0.06) = 1.06 (95% CI 0.91 to 1.23); log RoM: 0.06 (‐0.09 to 0.21); I² not applicable

272
(1 study)

⊕⊕⊝⊝
Lowa,b

RoM (95% CrI): 1.07
(0.81 to 1.41),
SUCRA = 0.344, mean Pr(best) = 0.030, mean rank = 4.93

Alpha2 agonist
vs placebo

Median length of ICU
stay: 7.5 days for placebo

6.43 days of ICU stay (95% CrI 4.42 to
9.33) corresponding to 8 days in the placebo group

RoM: exp(‐0.22) = 0.80 (95% CI 0.59 to 1.08); log RoM: ‐0.22 (‐0.53 to 0.08); I² not applicable

71
(1 study)

⊕⊕⊝⊝
Lowa,b

RoM (95% CrI): 0.80
(0.55 to 1.17),
SUCRA = 0.853, mean Pr(best) = 0.608, mean rank = 1.88

Cholinesterase inhibitor
vs placebo

Median length of ICU
stay: 8 days for placebo

17.53 days of ICU stay (95% CrI 11.76 to
26.14) corresponding to 8 days in the placebo group

RoM: exp(0.78) = 2.18 (95% CI 1.58 to 3.03); log RoM: 0.78 (0.46 to 1.11); I² not applicable

104
(1 study)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 2.19
(1.47 to 3.27),
SUCRA = 0.002, mean Pr(best) = 0, mean rank = 6.99

Opioid
vs placebo

No study reported this comparison

7.40 days of ICU stay (95% CrI 4.95 to
11.24) corresponding to 8 days in the opioid group

Pairwise meta‐analysis not performed

0

(0 studies)

⊕⊕⊝⊝
Very lowa,c

RoM (95% CrI): 0.92
(0.62 to 1.40),
SUCRA = 0.639, mean Pr(best) = 0.238, mean rank = 3.17

*The basis for the assumed risk (e.g. the median control group risk across studies). The corresponding risk (and its 95% CrI) is calculated as the assumed risk multiples the ratio of means (and its 95% CrI) based on NMA.
CI: confidence interval; CrI: credible interval; HMG‐CoA: 5‐hydroxy‐3‐methylglutaryl‐coenzyme A reductase inhibitor; NMA: network meta‐analysis; Pr(best): probability(best); RoM: ratio of means; SUCRA: surface under the cumulative ranking curve.

GRADE Working Group grades of evidence.
High quality: further research is very unlikely to change our confidence in the estimate of effect.
Moderate quality: further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate.
Low quality: further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate.
Very low quality: we are very uncertain about the estimate.

aDowngraded one level for imprecision (wide credible interval).
bDowngraded one level for single trial with risk of bias and indirectness.

cDowngraded two levels for only indirect evidence available and risk of bias of a single trial informing opioid vs typical antipsychotic.

Open in table viewer
Summary of findings 5. Length of hospital stay

Outcome: length of hospital stay

Patient or population: critically ill adult with confirmed or at high risk of delirium
Settings: intensive care units in Australia and New Zealand, Canada, Egypt, Netherlands, Turkey, USA, UK
Intervention: any pharmacological intervention

Control: placebo or active comparator

Outcomes

Illustrative comparative risks* (95% CI)

Ratio of means (RoM) based on log RoM estimates from meta‐analysis

(IV, random, 95% CI)

Number of participants
(studies)

Quality of the evidence
(GRADE) based on NMA

NMA results (assuming
consistency equations)

Assumed risk

Corresponding risk based on NMA estimates

Placebo/Comparator

Intervention drug

Typical AP
vs placebo

Median length of hospital stay: 13 to 26 days for placebo

16.48 days of hospital stay (95% CrI 11.74 to 21.29) corresponding to 18 days in the placebo group

RoM: exp(‐0.12) = 0.89 (95% CI 0.68 to 1.15); log RoM: ‐0.12 (‐0.38 to 0.14); I² = 72%

479
(2 studies)

⊕⊕⊝⊝
Lowa,b

RoM (95% CrI): 0.92
(0.65 to 1.18),
SUCRA = 0.722, mean Pr(best) = 0.235, mean rank = 2.67

Atypical AP
vs placebo

Median length of hospital stay: 6 to 26 days for placebo

16.69 days of hospital stay (95% CrI 12.47 to 20.79) corresponding to 18 days in the placebo group

RoM: exp(‐0.04) = 0.96 (95% CI 0.88 to 1.05); log RoM: ‐0.04 (‐0.13 to 0.05); I² = 0%

511
(3 studies)

⊕⊕⊕⊝

Moderateb

RoM (95% CrI): 0.93
(0.69 to 1.16),
SUCRA = 0.693, mean Pr(best) = 0.218, mean rank = 2.84

Statin (HMG‐CoA)
vs placebo

Mean length of hospital stay: 22 to 23.1 days for placebo

17.55 days of hospital stay (95% CrI 12.45 to 23.47) corresponding to 18 days in the placebo group

RoM: exp(‐0.01) = 0.99 (95% CI 0.88 to 1.13); log RoM: ‐0.01 (‐0.13 to 0.12); I² = 0%

369

(2 studies)

⊕⊕⊕⊝

Moderateb

RoM (95% CrI): 0.98
(0.69 to 1.30),
SUCRA = 0.537, mean Pr(best) = 0.147, mean rank = 3.78

Alpha2 agonist
vs placebo

Median length of hospital stay: 12.5 days for placebo

19.80 days of hospital stay (95% CrI 12.37 to 31.52) corresponding to 18 days in the placebo group

RoM: exp(0.09) = 1.09 (95% CI 0.84 to 1.42); log RoM: 0.09 (‐0.17 to 0.35); I² not applicable

71

(1 study)

⊕⊕⊕⊝

Moderatec

RoM (95% CrI): 1.10
(0.69 to 1.75),
SUCRA = 0.301, mean Pr(best) = 0.090, mean rank = 5.19

Cholinesterase Inhibitor
vs placebo

Median length of hospital stay: 25 days for placebo

20.00 days of hospital stay (95% CrI 12.64 to 31.93) corresponding to 18 days in the placebo group

RoM: exp(0.11) = 1.12 (95% CI 0.86 to 1.43); log RoM: 0.11 (‐0.15 to 0.36); I² not applicable

104

(1 study)

⊕⊕⊕⊝

Moderatec

RoM (95% CrI): 1.11
(0.70 to 1.77),
SUCRA = 0.280, mean Pr(best) = 0.078, mean rank = 5.32

Opioid
vs placebo

No study reported this comparison

17.51 days of hospital stay (95% CrI 9.89 to 28.78) corresponding to 18 days in the opioid group

Pairwise meta‐analysis not performed

0

(0 studies)

⊕⊕⊝⊝
Very lowb,d

RoM (95% CrI): 0.97
(0.55 to 1.60),
SUCRA = 0.532, mean Pr(best) = 0.225, mean rank = 3.81

*The basis for the assumed risk (e.g. the median control group risk across studies). The corresponding risk (and its 95% CrI) is calculated as the assumed risk multiples the ratio of means (and its 95% CrI) based on NMA.
CI: confidence interval; CrI: credible interval; HMG‐CoA: 5‐hydroxy‐3‐methylglutaryl‐coenzyme A reductase inhibitor; NMA: network meta‐analysis; Pr(best): probability(best); RoM: ratio of means; SUCRA: surface under the cumulative ranking curve.

GRADE Working Group grades of evidence.
High quality: further research is very unlikely to change our confidence in the estimate of effect.
Moderate quality: further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate.
Low quality: further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate.
Very low quality: we are very uncertain about the estimate.

aDowngraded one level for heterogeneity (I² of 50% to 75%, > 75% considered as medium and large heterogeneity).
bDowngraded one level for imprecision (wide credible interval).
cDowngraded one level for single small trial with risk of bias and indirectness.
dDowngraded two levels for only indirect evidence available and risk of bias of a single trial informing opioid vs typical antipsychotic.

Open in table viewer
Summary of findings 6. QTc prolongation

Outcome: QTc prolongation

Patient or population: critically ill adult with confirmed or at high risk of delirium
Settings: intensive care units in Australia and New Zealand, Canada, Egypt, Netherlands, Turkey, USA, UK
Intervention: any pharmacological intervention

Control: placebo or active comparator

Comparisons

Illustrative comparative risks* (95% CI)

Relative effect

OR (95% CI)

Absolute effect

(auto calculation using
GRADEpro GDT)

Number of
participants
(studies)

Quality of the evidence
(GRADE)

Assumed risk

Corresponding risk

Placebo/Comparator

Intervention drug

Typical antipsychotic vs placebo

62 per 1000

78 per 1000

1.26 (0.68 to 2.34)
I² = 0%

15 more per 1000
(from 19 fewer to
72 more)

656
(4 studies)

⊕⊕⊕⊕
High

Atypical antipsychotic vs placebo

90 per 1000

118 per 1000

1.28 (0.45 to 3.66)
I² = 56%

22 more per 1000
(from 48 fewer to
176 more)

577
(4 studies)

⊕⊕⊕⊝
Moderatea

Typical antipsychotic
vs atypical antipsychotic

114 per 1000

66 per 1000

0.55 (0.28 to 1.08)
I² = 0%

48 fewer per 1000
(from 79 fewer to
8 more)

447
(2 studies)

⊕⊕⊕⊕
High

Alpha2 agonist vs
typical antipsychotic

400 per 1000

400 per 1000

1.00 (0.17 to 5.98)
I² not applicable

0 fewer per 1000
(from 298 fewer to
399 more)

20
(1 study)

⊕⊕⊝⊝
Lowb

Alpha2 agonist
vs 5HT3 inhibitor

0 per 1000

0 per 1000

OR not estimable
I² not applicable

Not estimable

64
(1 study)

⊕⊕⊝⊝
Lowb

*The basis for the assumed risk (e.g. the median control group risk across studies). The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI).

CI: confidence interval; OR: odds ratio.

GRADE Working Group grades of evidence.
High quality: further research is very unlikely to change our confidence in the estimate of effect.
Moderate quality: further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate.
Low quality: further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate.
Very low quality: we are very uncertain about the estimate.

aDowngraded one level for heterogeneity (I² of 50% to 75%, > 75% considered as medium and large heterogeneity).

bDowngraded two levels for imprecision (wide confidence interval, single small trial with risk of bias).

Background

available in

Description of the condition

Delirium is a reversible, non‐specific syndrome of cognitive impairment commonly associated with surgery, infection, or critical illness (APA 2013). In the intensive care unit (ICU), this acute brain dysfunction is reported in 40% to 60% of non‐ventilated patients, and in 50% to 80% of mechanically ventilated patients (Ely 2001a; Ely 2001b; Ely 2007; Hipp 2012; Inouye 2014). Delirium is challenging to detect, as symptoms are highly variable, with either hyperactivity or hypoactivity, or even a mixed picture, and symptoms fluctuate with periods of lucidity (Inouye 2014). Delirium may be detected by psychiatric assessment based on the Diagnostic and Statistical Manual (DSM) criteria (APA 2013), or by use of a validated screening tool (Bergeron 2001; Ely 2001a; Neelon 1996); however, assessment in the ICU is predicated on the patient being awake and able to communicate, and delirium is said to be "unable to be assessed" when the patient does not respond to verbal communication. In the ICU, commonly used sedatives and opioids impair consciousness, thereby making identification of delirium challenging (Patel 2014). Drug exposure should be considered when ICU delirium is assessed, and if possible, assessments should be co‐ordinated with periods of wakefulness or should be conducted during a sedation interruption (Patel 2014).

Over the past decade, we have acquired a greater understanding of the effects of delirium on patients, their families, and the healthcare system. Clinically important outcomes of delirious critically ill patients include prolonged duration of mechanical ventilation and ICU and hospital stay, as well as long‐term cognitive impairment, increased likelihood of transfer to long‐term care facilities, and mortality (Black 2011; Ely 2001b; Ely 2004; Girard 2010b; Jackson 2004; Lin 2004; Milbrant 2004; Pisani 2009; Van den Boogaard 2012). The odds of a poor outcome with delirium are increased by patient frailty, advanced age (> 75 years), pre‐existing cognitive impairment, and visual or hearing impairment (Andrew 2006; Inouye 2006a). Precipitating factors are numerous and include sleep deprivation, pain, environmental insults (e.g. noise, physical restraint use, catheters), and psychoactive drug exposure (e.g. sedatives) (Burry 2017; Fraser 2013; Inouye 2006a; Rose 2016; Zaal 2015).

Description of the intervention

Pharmacological interventions for delirium treatment have focused on alterations in neurotransmitter pathways, in particular dopaminergic and cholinergic pathways. At present, the pathophysiology of delirium is not fully understood (Gunther 2008; Reade 2014). Hypotheses currently include abnormalities in cerebral oxidative metabolism, direct neurotoxic effects of inflammatory cytokines, such as those released during sepsis and septic shock, and alterations in neurotransmitters that modulate cognition, behaviour, and mood (e.g. cholinergic, dopaminergic, serotonergic, gamma‐aminobutyric acid (GABA) pathways) (Cerejeira 2011; de Rooji 2007; Ebersoldt 2007; Flacker 1999; Gunther 2008; Inouye 2006b; Rudolph 2008; White 2002). These pathophysiological mechanisms are not thought to be mutually exclusive and are likely to act together.

In the light of these different proposed mechanisms, it is not surprising that numerous pharmacological strategies for delirium have been investigated, including alpha2 agonists, antidepressants, antipsychotic drugs (either typical or atypical agents), benzodiazepines, cholinesterase inhibitors, melatonin and melatonin agonists, and opioids (Devlin 2010; Girard 2010a; Maldonado 2009; Ohta 2013; Reade 2009; Rubino 2010; van Eijk 2010). In considering these agents, it is important to note that critical care guidelines first recommend the use of non‐pharmacological strategies in both prevention and management of delirium (Barr 2013). These non‐pharmacological strategies include early mobilization and re‐orientation, risk factor assessment and modification (e.g. drugs, medical devices), and normalization of the sleep‐wake cycle (e.g. noise reduction, use of ear plugs) (Inouye 2006a; Schweickert 2009). Guidelines suggest that when delirium is suspected or identified, patients should be closely evaluated for identification of underlying cause(s), allowing for exposure to be removed or corrected whenever possible; pharmacological interventions are to be used only when non‐pharmacological methods have failed to control symptoms (Barr 2013).

How the intervention might work

Given the multiple neurotransmitters linked to development of delirium, pharmacological strategies have investigated target suspected neurotransmitter imbalances or attempts to control distressing cognitive (e.g. hallucinations) or dangerous behaviours (e.g. agitation, interference with medical devices). Pharmacological strategies may target pain control (e.g. opioids) or the dopaminergic (e.g. antipsychotics), cholinergic (e.g. cholinesterase inhibitors), GABA (e.g. benzodiazepines), N‐methyl‐D‐aspartate (NMDA) (e.g. ketamine), serotonergic (e.g. antidepressants, antinauseants, melatonin), and alpha2 (e.g. clonidine, dexmedetomidine) pathways (Devlin 2010; Girard 2010a; Maldonado 2009; Ohta 2013; Reade 2009; Rubino 2010; van Eijk 2010). The specific therapeutic effects of such agents are unknown, but effects may be mediated through their ability to affect sedation and behavioural symptoms.

Despite conflicting evidence for the benefits of various pharmacological interventions, many of these agents are routinely used to treat ICU delirium, or to at least manage symptoms (e.g. agitation), and they are often continued after hospital discharge (Bell 2007; MacSweeney 2009). Of the available pharmacological strategies, antipsychotics represent the most common treatment for ICU delirium, despite limited evidence regarding their benefit and studies in non‐critically ill patients identifying significant adverse effects, including sudden death (Barr 2013; Briskman 2010; Burry 2014; Gill 2007; MacSweeney 2009; Tropea 2009; Wang 2005).

Why it is important to do this review

ICU delirium is associated with prolonged duration of mechanical ventilation and ICU and hospital stay, as well as increased mortality (Ely 2001b; Ely 2004; Girard 2010b; Jackson 2004; Lin 2004; Milbrant 2004; Pisani 2009; Van den Boogaard 2012). ICU delirium initiates a cascade of events that can include functional decline and long‐term cognitive impairment, with resultant caregiver burden (Girard 2010b; Jackson 2004; Van den Boogaard 2012). The geriatric and oncological literature shows that delirium is traumatic for both patients and family members, and it can lead to long‐term psychological sequelae (Bruera 2009; Morita 2004; Partridge 2013; Rosenbloom‐Brunton 2010). The economic burden of delirium is also significant; each additional day spent in a delirious state is associated with a 20% increased risk of prolonged hospitalization, translating to an average of more than 10 additional hospital days per patient. The annual cost of delirium is estimated to be greater than USD 164 billion in the USA, and greater than EUR 182 billion as estimated across 18 European countries (Leslie 2008; OECD 2012; WHO Regional Office 2012). Furthermore, delirium is considered a substantial public health concern that has garnered the attention of patient safety institutes; it is now included as an indicator of quality care for the elderly (IHI 2015).

Advances in detection of ICU delirium and improved understanding of its impact on patient outcomes have prompted trials comparing different treatment options (both pharmacological and non‐pharmacological), either against each other or versus placebo. However, there remains considerable uncertainty regarding the relative benefits and safety of pharmacological interventions for the ICU population, and trials have shown benefit (Devlin 2010; Pandharipande 2007; Reade 2009), indeterminate outcomes (Girard 2010a; Page 2013), or harm (van Eijk 2010). A previous Cochrane Review on antipsychotics for delirium did not specifically address the ICU population (Lonergan 2007); numerous ICU‐specific trials have been published since this review was completed. A recent systematic review of ICU delirium included both prevention and treatment studies (Al‐Qadheeb 2014), as well as randomized controlled trials (RCTs) evaluating sedation strategies, in which delirium was evaluated as a secondary endpoint when the study population considered was not restricted to patients with confirmed delirium. As a Cochrane Review protocol by Herling and colleagues will provide data on delirium prevention trials in critically ill adult patients (Herling 2018), our review focuses on delirium treatment trials in critically ill adult patients.

Given the availability of numerous strategies to treat ICU delirium in clinical practice, and the existence of many trials yielding conflicting results, we planned this systematic review to include a network meta‐analysis (NMA) to determine the comparative benefits and harms of all published pharmacological interventions for treatment of delirium based on available direct and indirect evidence of relevance. An NMA, also known as a multiple treatment comparison meta‐analysis, is a statistical method used to assess the comparative effectiveness of multiple different interventions among similar patient populations that have not been compared directly in an RCT. In contrast to conventional pairwise meta‐analysis (e.g. RCTs comparing treatment A vs treatment B), NMAs can provide estimates of relative efficacy between all interventions, even though some have never been compared head‐to‐head via indirect evidence (i.e. comparing results from two or more studies that have one treatment in common).

Objectives

available in

Primary objective

  1. To assess the effects of pharmacological interventions for treatment of delirium on duration of delirium in critically ill adults with confirmed or documented high risk of delirium

Secondary objectives

To assess the following:

  1. effects of pharmacological interventions on delirium‐free and coma‐free days; days with coma; delirium relapse; duration of mechanical ventilation; ICU and hospital length of stay; mortality; and long‐term outcomes (e.g. cognitive; discharge disposition; health‐related quality of life); and

  2. the safety of such treatments for critically ill adult patients.

Methods

available in

Criteria for considering studies for this review

Types of studies

We sought randomized controlled trials (RCTs), including quasi‐RCTs (i.e. when the method of allocation was not strictly random, such as by alternation, date of birth, or case record number), and RCTs with an open‐label study design. We excluded non‐RCT study designs due to their potential for bias and the anticipated availability of RCTs.

Types of participants

We sought RCTs designed to examine pharmacological interventions for treatment of delirium in critically ill adults. We defined critically ill patients as those treated in an ICU of any specialty (e.g. burn, cardiac, medical, surgical, trauma) or high‐dependency unit. We included trials in which a trained individual (e.g. psychiatrist) evaluated participants for delirium using the Diagnostic and Statistical Manual of Mental Disorders (DSM) criteria (APA 2013), or using a validated delirium assessment tool (e.g. Confusion Assessment Method for the ICU (CAM‐ICU), Intensive Care Delirium Screening Checklist (ICDSC), Neelon and Champagne (NEECHAM) Confusion Scale, Delirium Rating Scale, or Delirium Rating Scale‐revised‐98) (Bergeron 2001; Ely 2001b; Neelon 1996; Trzepacz 2001). We also included RCTs that treated subsyndromal delirium (i.e. some features of delirium), as these patients are considered to be at high risk of transitioning to delirium and are often included in ICU delirium treatment studies.

Types of interventions

We sought delirium treatment RCTs that compared use of any pharmacological (drug) to treat delirium including alpha2 agonists (e.g. clonidine, dexmedetomidine), antidepressants (e.g. fluoxetine), antipsychotics (either typical (e.g. haloperidol) or atypical agents (e.g. quetiapine)), benzodiazepines (e.g. lorazepam), cholinesterase (CHE) inhibitors (e.g. rivastigmine), N‐methyl‐D‐aspartate (NMDA) receptor antagonist (e.g. ketamine), melatonin and melatonin agonists (e.g. ramelteon), opioids (e.g. morphine), propofol, serotonin receptor antagonists (e.g. ondansetron), and statins (e.g. atorvastatin) versus another active drug treatment, a placebo, or a non‐pharmacological intervention (e.g. mobilization). We did not apply any restrictions in terms of drug class, dose, route of administration, or duration of delirium or drug exposure.

Our hypothetical network structure published in the protocol was designed to be analysed at the drug class level and illustrated a network diagram of 'nodes' (i.e. drug classes) and 'edges' (i.e. comparisons between different drug classes from existing trials) (Burry 2015), thus describing a treatment network of all possible comparisons between drug classes. The extent to which trial data are available along the 'edges' for each outcome will depend upon the search results.

Types of outcome measures

Primary outcomes

  1. Duration of delirium (defined as the time from which delirium was identified or the patient was randomized until resolution (i.e. screened negative as defined by study authors)), measured in days

Secondary outcomes

  1. Delirium‐free and coma‐free days (to 14, 21, 28 days) and days with coma (reported in days)

  2. Relapse of delirium (reported as a proportion)

  3. Resolution of delirium symptoms (e.g. hallucinations, agitation)

  4. Duration of mechanical ventilation (days)

  5. Length of stay (ICU and hospital) (days)

  6. Mortality (e.g. 30‐day, 60‐day, 90‐day, ICU, hospital, following hospital discharge, and one year as reported by study authors)

  7. Use of physical restraint

  8. Hospital discharge disposition (e.g. chronic care facility, home)

  9. Long‐term cognitive outcomes (e.g. change in Mini Mental Status Exam) as reported by study authors

  10. Health‐related quality of life (as reported by study authors)

  11. Adverse drug events (e.g. akathisia, arrhythmias, extrapyramidal side effects, seizures)

Search methods for identification of studies

We sought to identify all eligible trials regardless of publication status through systematic and sensitive search strategies as outlined in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). We did not impose any language or publication restrictions.

Electronic searches

Our electronic search strategies were developed and tested through an iterative process with an experienced medical information specialist (Appendix 1; Appendix 2; Appendix 3; Appendix 4). The search strategies utilized a combination of controlled vocabulary terms (e.g. ICU, delirium) and keywords (e.g. ICU, acute brain dysfunction). We used a validated RCT filter and a filter that limited studies to humans. We searched the following electronic databases from their inception date to 21 March 2019: Ovid MEDLINE ALL®, Embase Classic+Embase, and PsycINFO using OVID platform. We also searched the Cochrane Library on Wiley, the International Prospective Register of Systematic Reviews (PROSPERO; http://www.crd.york.ac.uk/PROSPERO/), the Cumulative Index to Nursing and Allied Health Literature (CINAHL), and Web of Science. We adjusted search vocabulary and syntax for each database. The core strategy was reviewed prior to execution by another senior information specialist using the Peer Review for Electronic Search Strategies (PRESS) template (Sampson 2009).

We performed a separate search for published systematic reviews to identify additional published or unpublished trials. We performed a grey literature search of relevant databases and websites using resources listed in Grey Matters (http://www.cadth.ca/en/resources/finding‐evidence‐is/grey‐matters) developed by the Canadian Agency for Drugs and Technologies in Health (CADTH). Last, we scanned the reference lists of all included studies and any relevant reviews on delirium treatment to identify additional studies.

Searching other resources

We hand searched the citations of all included studies and any systematic reviews identified. We searched abstracts from annual scientific meetings of the Society of Critical Care Medicine, the European Society of Intensive Care Medicine, the International Symposium on Intensive Care and Emergency Medicine, the American Delirium Society, the American Thoracic Society, Chest, and the Australian and New Zealand Intensive Care Society from 2011 to 2019 to identify studies not yet published in full. We also searched for unpublished and ongoing trials on the following websites using the term "delirium".

  1. www.clinicaltrials.gov/

  2. www.who.int/trialsearch

Data collection and analysis

Selection of studies

Two review authors (LB, LR) independently screened all retrieved titles and abstracts using the selection criteria described in the protocol (Burry 2015). Next, these two review authors (LB, LR) independently reviewed selected full‐text articles to determine inclusion. We resolved disagreements by discussion, without the need to refer to the assigned independent arbiter (EWE). References were managed in the software package EndNote (Endnote Version X6, Thomson Reuters, Carlsbad, CA, USA), and we documented the reasons for exclusion in the notes field. We documented the process of study selection using a PRISMA flow diagram (Moher 2009).

Data extraction and management

We extracted data from the included trials using a standardized electronic form (Microsoft Corporation, Redmond, WA, USA). Four review authors (DW, SM, NA, IE) worked independently to extract data; two review authors were assigned to each study. Data extractors were not blinded to the identity of study authors. We extracted data related to publication (e.g. journal reference, study authors, year of publication), study design (e.g. number of centres, country, methods of enrolment, randomization, allocation concealment, blinding), patient demographics (e.g. age, sex, severity of illness score, reasons for admission), interventions (e.g. drug, mode of administration, dose, how titrated, who administered, use of rescue medications for agitation), delirium and sedation assessment (e.g. method, who assessed), co‐interventions that might alter duration delirium, stay or mechanical ventilation (e.g. ventilator weaning strategies, type of sedative or analgesic, early mobilization), and our selected outcomes. We also extracted data on management of missing data, reporting of outcomes, type of analysis performed (e.g. intention to treat), and other potential sources of bias (e.g. funding source, referral bias). When necessary, we (LB) contacted the study corresponding author to clarify issues related to data reporting or to obtain further study details. Data extraction was confirmed and discrepancies between review author pairs resolved by an arbiter (LB). Checked data were then entered into Review Manager 5 by one review author (WC) and were double‐checked by two review authors (BH, LB) (Review Manager 2014).

Assessment of risk of bias in included studies

Each data extractor (DW, SM, NA, IE) independently assessed risk of bias for his/her assigned studies. A third review author (LB) verified each assessment. Risk of bias was determined via a domain‐based evaluation that was included in the data extraction form, and as recommended by Cochrane (Higgins 2011). The domains were as follows.

  1. Random sequence generation (i.e. selection bias).

  2. Allocation concealment (i.e. selection bias).

  3. Blinding of participants and personnel (i.e. performance bias).

  4. Blinding of outcomes assessment (i.e. detection bias).

  5. Incomplete outcome data (i.e. attrition bias).

  6. Selective reporting.

  7. Other bias (e.g. study source of funding, role of the sponsor, referral bias).

For each domain, we explicitly judged the risk of bias as high, low, or unclear. We assigned domains 'unclear' if detail was insufficient to determine risk, or if risk of bias was unclear or unknown. We judged incomplete outcome data as low risk of bias when causes of dropout were similar and numbers were balanced between study groups and less than 15%. We generated a risk of bias graph and summary upon completion of assessment.

Measures of treatment effect

For all continuous outcomes (duration of delirium, duration of ventilation, hospital length of stay, ICU length of stay, delirium‐free and coma‐free days, coma days), more than half of the included studies reported medians and interquartile ranges (IQRs) as opposed to means and standard deviations (SDs), standard errors (SEs) or confidence intervals (CIs). We converted medians and IQRs to means and SDs according to methods described elsewhere (Wan 2014). Due to the skewed nature of these outcomes, we transformed means and SDs to the log scale using methods outlined previously (Higgins 2008). For continuous outcomes, the mean difference (MD) between two interventions on the log scale equals the log ratio of means (log RoM); after exponentiation, estimates can be interpreted as the RoM of two interventions. Evidence synthesis on the log RoM scale allows continuous outcomes measured within various lengths of time windows across studies. Findings for binary outcomes were expressed in terms of odds ratios (ORs).

Based on mean and SD values following transformation, fixed‐effect and random‐effects NMA models with Normal Likelihood and the identify link were fit to the data (Dias 2011b). We present comparisons between interventions in terms of RoM (RoM: mean[expt]/mean[ctrl]) with 95% credible intervals (CrI). Values of RoM < 1 favour the active intervention, whereas values of RoM > 1 favour the placebo or comparator for all continuous outcomes except for delirium‐free and coma‐free days. For dichotomous outcome measures, both fixed‐effect and random‐effects NMA models with binomial likelihood were fit to the data, with comparisons between interventions expressed in terms of ORs with 95% CrI.

For each outcome, NMA enabled us to calculate the probability for each intervention to be at each possible rank. The Surface Under the Cumulative RAnking curve (SUCRA) value, the mean rankings (with 2.5% and 97.5% quantiles) of each intervention, and the probability of each intervention to be the best (referred to hereafter as 'Pr(best)') were also estimated (Salanti 2011). Pr(best) and SUCRA values range between 0 and 1, with values nearer 1 indicative of preferred treatments. Values of smaller mean rank also suggest preferred treatments. Further details regarding the methods and implementation of NMA are provided in the published protocol (Burry 2015).

Unit of analysis issues

We used individual study participants in each trial arm as the unit of analysis. We included all interventions relevant to this review. If a trial involves multiple arms of the same drug class (e.g. multiple atypical antipsychotics) compared to a control group, we planned to merge data from the same drug class for pairwise comparisons. Neither cluster‐randomized trials nor cross‐over trials were identified through the literature search. We did not anticipate cross‐over trials to evaluate delirium in the ICU, as this study design is not typically used in the ICU.

Dealing with missing data

We conducted meta‐analyses based on data available from our included studies. For missing SDs associated with continuous outcomes, we first contacted study authors for more information; we made a maximum of three attempts.

Assessment of heterogeneity

An important aspect of NMA is examining included studies to determine if they are sufficiently similar in terms of study design and patient population. We describe each included trial in the Characteristics of included studies tables. Within a treatment network involving multiple interventions, heterogeneity can be the result of an uneven distribution of important clinical and methodological effect modifiers across studies or across comparisons. We assessed the presence of statistical heterogeneity by visual inspection of forest plots and by calculation of the I² statistic (Higgins 2003), as well as by the Chi² test for homogeneity (P < 0.10 deemed significant). If the I² statistic was > 50%, we assessed the types and sources of heterogeneity (clinical and methodological). We qualitatively assessed clinical heterogeneity by examining additional delirium management strategies used in each trial (e.g. use of rescue medications or physical restraints to manage severe agitation, non‐drug strategies such as noise reduction or early mobilization). We also assessed clinical heterogeneity by examining factors that may influence delirium and sedation practices (for example, types of sedatives and analgesics used, use of drugs known to increase the risk of delirium, e.g. benzodiazepines, and definitions of outcomes assessed).

Assessment of reporting biases

Reporting biases can occur due to an increased likelihood of positive (demonstration of effect) trials (large or small) being published compared to negative (no effect demonstrated) trials. It is difficult to estimate the number of unpublished delirium trials. For direct comparisons in the network where a minimum of 10 studies were available, we reviewed comparison‐adjusted funnel plots to assess for small‐study effects as signals of publication bias (Salanti 2014).

Data synthesis

Methods for direct treatment comparisons

We performed conventional pairwise meta‐analyses in Review Manager 5.3 for all outcomes and comparisons that had at least two studies available (Review Manager 2014). A variation of the inverse‐variance random‐effects model was applied to continuous outcomes (DerSimonian 1986), whereas the Mantel‐Haenszel random‐effects model was applied to binary outcomes (DeMets 1987), allowing for variation within and between studies.

Methods for network meta‐analysis (mixed treatment comparisons)

NMA is a method of synthesizing evidence from trials addressing the same question but involving multiple different interventions. NMA combines direct and indirect evidence across a network of RCTs into a single effect size for each pair of interventions. For a given comparison (e.g. A vs B), direct evidence was provided by studies that compared two treatments head‐to‐head. Indirect evidence for this comparison was provided by studies that compared A versus C and B versus C (Caldwell 2005; Higgins 1996).

We followed established procedures to assess the validity of the assumptions of homogeneity, similarity, and consistency (Donegan 2013). We performed NMAs within a Bayesian framework, assuming a common between‐study variance parameter across all comparisons and accounting for correlations in multi‐arm studies (Lu 2006; Salanti 2011). A vague prior distribution for the between‐study variance parameter (specifically, Uniform (0, 3)) and vague prior distribution for log ratio of means between each intervention compared with placebo (specifically, Normal (0, 100)) were used for all analyses. We reported findings when using the most recent PRISMA Extension Statement for NMA (Hutton 2015). Two review authors (WC, BH) performed NMAs with OpenBUGS software (version 3.2.3, MRC Biostatistics Unit, Cambridge, UK) (Lunn 2000; Spiegelhalter 2014). We expressed findings for continuous outcomes in terms of RoMs and findings for binary outcomes in terms of ORs with corresponding 95% CrI (Dias 2011a; Dias 2011b; Dias 2013). Network diagrams were drawn to depict the evidence for each outcome. In the network diagrams, the size of the treatment nodes reflects the number of participants randomized to each treatment, and the thickness of the edges reflects the number of studies informing each comparison.

We evaluated the adequacy of model fit by comparing the total residual deviance to the number of unconstrained data points (i.e. the total number of study arms); fit was adequate if these quantities were close. Based on mean and SD values following transformation, fixed‐effect and random‐effects NMA models with Normal Likelihood and the identity link were fit to the data (Dias 2011b). Both fixed‐effect (FE) and random‐effects (RE) consistency models were fit, and we compared these models using the Deviance Information Criterion (DIC), with lower value indicating better model fit (Spiegelhalter 2002). We considered a difference of five points or more indicative of an important difference. We also fit unrelated means models to the data and compared DIC values and posterior mean deviance contributions with those from consistency models to detect violations of the consistency assumption. We assessed model convergence with established methods including inspection of Gelman‐Rubin‐Brooks diagnostics and potential scale reduction factors (Brooks 1998; Gelman 1996). As described earlier, we also estimated SUCRA values, mean rankings, and Pr(best) values for each intervention (Salanti 2011). For additional analyses, we planned to explore the impact of certain study characteristics through subgroup analyses or meta‐regression.

Subgroup analysis and investigation of heterogeneity

We planned to explore subgroup analyses or meta‐regression analyses or both, to assess the impact of covariates on findings to establish their robustness, if sufficient studies were available; specifically:

  1. age (< 65 years, ≥ 65 years);

  2. different ICU populations (e.g. medical only, surgical only);

  3. delirium subtype (e.g. hyperactive, hypoactive, mixed); or

  4. use of co‐interventions with non‐drug approaches (e.g. noise reduction, music therapy, early mobilization).

Sensitivity analysis

We planned to consider sensitivity analyses involving alternative geometries of the network. Planned re‐formulations of the network included:

  1. excluding studies with high risk of bias;

  2. collapsing atypical and typical antipsychotics into one node;

  3. splitting each node to reflect ‘low dose’ and ‘high dose’, based on the median dose reported in trials; and

  4. splitting each node to reflect fixed dosing and PRN (pro re nata or as needed) only dosing.

We explored additional analyses after excluding studies that focused on subsyndromal delirium.

'Summary of findings' tables and GRADE

In the 'Summary of findings' tables, we present the specific review outcomes duration of delirium, delirium‐free and coma‐free days, days with coma, duration of mechanical ventilation, length of ICU, and length of hospital stay, as recommended by Cochrane (Higgins 2011; Schunemann 2011; Yepes‐Nunez 2019). We used the GRADE approach (https://gradepro.org/) to assess the quality of the evidence for comparisons based on NMA. We graded the quality of evidence for each outcome as 'high', 'moderate', 'low', or 'very low' using GRADEPro software (GRADEpro GDT), after considering trial limitations (randomization, allocation concealment, and blinded outcome assessment), within‐study directness of evidence, heterogeneity, precision of effect estimates, and indirectness. We did not assess risk of publication bias/small‐study effects through funnel plots given the small number of studies available for any pairwise comparison. When we identified an issue that we considered to be serious for each of the GRADE criteria, we downgraded the quality of evidence and justified our decision in the table footnotes. We assessed the extent of heterogeneity (i.e. I² statistic) and examined imprecision based on the width of the CI for treatment effect estimates.

Results

Description of studies

Results of the search

See Characteristics of included studies,Characteristics of excluded studies,Characteristics of studies awaiting classification, and Characteristics of ongoing studies.

The results of our search are outlined in Figure 1. The electronic database search yielded 7658 citations, and we identified an additional 16 records through other sources. After we removed duplicate items, 4461 unique citations remained. We excluded 4076 studies based on title and abstract, and we assessed the remaining 385 papers as full text. Fourteen studies met our inclusion criteria (Al‐Qadheeb 2016; Atalan 2013; Bakri 2015; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Page 2017; Reade 2009; Reade 2016; Skrobik 2004; van Eijk 2010). Six studies await classification (NCT02366299; NCT00429676; Emerson 2014; Peters 2015; Schoeffler 2012; ISRCTN33122761) ‐ three as conference abstracts (Emerson 2014; Peters 2015; Schoeffler 2012), and three as trial registrations (NCT02366299; NCT00429676; ISRCTN33122761). Ten studies are ongoing (NCT01811459; NCT03317067; NCT02807467; NCT02216266; NCT02343575; NCT00351299; NCT03628391; IRCT20121231011956N10; IRCT20180911040998N1; NCT03392376), two of which have published protocols (Louis 2018; Hollinger 2017).


Study flow diagram.

Study flow diagram.

Included studies

See the Characteristics of included studies table.

Study population

The 14 included studies recruited 1844 adult participants, with sample sizes ranging from 20 in Reade 2009 to 566 in Girard 2018. Seven studies enrolled more than 100 participants (Girard 2010a; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Page 2017; van Eijk 2010). Twelve studies enrolled a mix of medical and surgical participants; two enrolled cardiovascular surgery participants only (Atalan 2013; Hakim 2012).

Eight studies used the CAM‐ICU to screen for delirium (Atalan 2013; Girard 2010a; Girard 2018; Needham 2016; Page 2013; Page 2017; Reade 2016; van Eijk 2010); the remaining six used the ICDSC (Al‐Qadheeb 2016; Bakri 2015; Devlin 2010; Hakim 2012; Reade 2009; Skrobik 2004). Five studies permitted inclusion of patients at high risk of developing delirium (i.e. delirium status not confirmed at study enrolment) (Al‐Qadheeb 2016; Girard 2010a; Hakim 2012; Page 2013; Reade 2009). Of these, two trials enrolled participants with subsyndromal delirium (Al‐Qadheeb 2016; Hakim 2012). One study enrolled mechanically ventilated participants with specifically agitated delirium (Reade 2009). Through written communication with the principal investigator, we confirmed that all participants had at a minimum subsyndromal delirium at enrolment, with 40% confirmed as delirious (i.e. ICDSC > 4). The remaining studies enrolled a combination of delirious and comatose participants (Girard 2010a; Page 2013), or investigators confirmed delirium status before enrolment (Girard 2018). These trials all examined delirium during ICU stay (and not thereafter).

Study design and setting

All trials but one were randomized (Skrobik 2004). Six trials were multi‐centre studies (Devlin 2010; Girard 2010a; Girard 2018; Needham 2016; Reade 2016; van Eijk 2010), and eight were single‐centre studies (Al‐Qadheeb 2016; Atalan 2013; Bakri 2015; Hakim 2012; Page 2013; Page 2017; Reade 2009; Skrobik 2004). Six studies were conducted in North America ‐ four exclusively in the USA (Al‐Qadheeb 2016; Girard 2010a; Girard 2018; Needham 2016), one exclusively in Canada (Skrobik 2004), and one in both Canada and the USA (Devlin 2010). The other studies took place in Australia and New Zealand (Reade 2009; Reade 2016), Egypt (Bakri 2015; Hakim 2012), the Netherlands (van Eijk 2010), Turkey (Atalan 2013), and the UK (Page 2013; Page 2017).

Interventions and comparators

Ten trials were placebo‐controlled (Al‐Qadheeb 2016; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Page 2017; Reade 2016; van Eijk 2010). Four were head‐to‐head comparisons of different drugs (Atalan 2013; Bakri 2015; Reade 2009; Skrobik 2004). Three included three study groups (Bakri 2015; Girard 2010a; Girard 2018). Ten studied an antipsychotic intervention, predominantly haloperidol (Al‐Qadheeb 2016; Atalan 2013; Bakri 2015; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Page 2013; Reade 2009; Skrobik 2004). Three studied alpha2 agonists (all used dexmedetomidine) (Bakri 2015; Reade 2009; Reade 2016). Two trials studied a statin (Needham 2016; Page 2017). The remaining trials evaluated morphine (Atalan 2013), ondansetron (Bakri 2015), or rivastigmine (van Eijk 2010). Ten trials titrated the study drug based on symptoms or response (Atalan 2013; Bakri 2015; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Reade 2009; Reade 2016; Skrobik 2004; van Eijk 2010); four used fixed drug regimens (Al‐Qadheeb 2016; Needham 2016; Page 2013; Page 2017).

The extent to which study medication was given also varied, with some studies continuing drug for a fixed duration irrespective of whether delirium had resolved and others protocolizing discontinuation of the study drug once the patient was no longer delirious (Devlin 2010; Girard 2018; Page 2013). The duration of study drug exposure varied across trials including maximum of 28 days (Needham 2016; Page 2017), 14 days (Girard 2010a; Girard 2018; Page 2013), 10 days (Al‐Qadheeb 2016; Atalan 2013; Devlin 2010), seven days (Reade 2016), five days (Skrobik 2004), three days (Bakri 2015), as long as deemed medically necessary (Reade 2009), until delirium resolution or hospital discharge (van Eijk 2010), or for 24 hours after ICDSC was zero (Hakim 2012).

Eleven trials allowed use of an additional drug for management of breakthrough delirium symptoms or agitation (e.g. sedative, antipsychotic) (Atalan 2013; Bakri 2015; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Page 2013; Reade 2009; Reade 2016; Skrobik 2004; van Eijk 2010).

Outcomes

Outcomes varied in terms of measurement and reporting. All but two studies reported median (IQR) or mean (SD) for delirium duration (Bakri 2015; Skrobik 2004). The planned primary outcome defined as time from which delirium wasfirst identified to when it was first resolved was rarely reported (Devlin 2010). Most trials reported duration of delirium with variable definitions of resolved delirium (e.g. one negative score, two consecutive days with negative score, no definition provided). Therefore we chose to pool the results as duration of delirium as reported by study authors. Five studies reported median (IQR) or mean (SD) for number of days with coma (Girard 2010a; Girard 2018; Needham 2016; Page 2013; Page 2017); four reported median (IQR) or mean (SD) for number of days alive without delirium or coma (Girard 2010a; Girard 2018; Page 2013; Page 2017); eight reported median (IQR) or mean (SD) for mechanical ventilation duration (Al‐Qadheeb 2016; Atalan 2013; Devlin 2010; Girard 2010a; Girard 2018; Needham 2016; Reade 2009; Reade 2016); 11 reported median (IQR) or mean (SD) for ICU length of stay (Al‐Qadheeb 2016; Atalan 2013; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Reade 2009; Reade 2016; van Eijk 2010); nine reported median (IQR) or mean (SD) for hospital length of stay (Atalan 2013; Devlin 2010; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Page 2017; Reade 2016; van Eijk 2010); and 11 reported mortality at various time points (Al‐Qadheeb 2016; Atalan 2013; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Page 2017; Reade 2009; van Eijk 2010). Three studies reported discharge disposition (Al‐Qadheeb 2016; Devlin 2010; Reade 2016). Reported adverse events included arrhythmias (Girard 2010a; Girard 2018; Page 2013; Reade 2009), extrapyramidal symptoms (Al‐Qadheeb 2016; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Page 2013; Skrobik 2004), use of physical restraints (Reade 2009; van Eijk 2010), unintentional device removal (Al‐Qadheeb 2016; Devlin 2010; Page 2013; Reade 2009; Reade 2016), and QTc prolongation (Al‐Qadheeb 2016; Bakri 2015; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Page 2013; Reade 2009).

For meta‐analysis and network meta‐analysis, we removed one open‐label trial from syntheses given what were judged to be special features in the study population (i.e. cardiovascular surgery, commonly associated with short ICU stays) and differences in baseline characteristics between dexmedetomidine and haloperidol arms despite randomization (Reade 2009).

Excluded studies

See Characteristics of excluded studies.

We excluded seven studies for the following reasons (Eremenko 2014; Khan 2019; Mailhot 2014; Pandharipande 2007; Riker 2009; Tagarakis 2012;Waszynski 2018): study design (Eremenko 2014; Pandharipande 2007; Riker 2009); no pharmacological intervention (Khan 2019; Mailhot 2014; Waszynski 2018); and no validated method to determine delirium (Tagarakis 2012).

Studies awaiting classification

See Characteristics of studies awaiting classification.

Six studies available as abstract ‐ Emerson 2014,Peters 2015, and Schoeffler 2012 ‐ or as trial registration ‐ NCT00429676,ISRCTN33122761, and NCT02366299 ‐ await classification due to insufficient information. These studies evaluate an antipsychotic (NCT00429676), clonidine (Schoeffler 2012), physostigmine (ISRCTN33122761), dexmedetomidine and propofol (NCT02366299), a multi‐component delirium management strategy (Emerson 2014), and intranasal insulin aspart (Peters 2015).

Ongoing studies

See Characteristics of ongoing studies.

Ten studies classified as ongoing studies will be monitored for incorporation into future updates of this review (IRCT20121231011956N10; IRCT20180911040998N1; NCT03392376; NCT01811459; NCT03317067; NCT02807467; NCT02216266; NCT02343575; NCT00351299; NCT03628391). Interventions include antipsychotics (NCT01811459; NCT03628391; IRCT20121231011956N10; IRCT20180911040998N1; NCT03392376), dexmedetomidine (NCT03317067; NCT02807467, NCT00351299), physostigmine (NCT02216266), and valproic acid (NCT02343575).

Risk of bias in included studies

We summarize risk of bias data in Figure 2 and Figure 3. Nine trials scored low risk of bias across all domains (Al‐Qadheeb 2016; Devlin 2010; Girard 2010a; Girard 2018; Needham 2016; Page 2013; Page 2017; Reade 2016; van Eijk 2010).


Risk of bias graph: review authors' judgements about each risk of bias item presented as percentages across all included studies.

Risk of bias graph: review authors' judgements about each risk of bias item presented as percentages across all included studies.


Risk of bias summary: review authors' judgements about each risk of bias item for each included study.

Risk of bias summary: review authors' judgements about each risk of bias item for each included study.

Allocation

We judged all studies but two ‐ Atalan 2013 and Skrobik 2004 ‐ to have low risk of selection bias due to random sequence generation. Skrobik 2004 performed quasi‐randomization (i.e. even/odd enrolment day), and Atalan 2013 did not report the method of sequence generation. Twelve studies used computer‐generated randomization tables (Al‐Qadheeb 2016; Bakri 2015; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Page 2013; Needham 2016; Page 2017; Reade 2009; Reade 2016; van Eijk 2010). We judged eleven studies to have adequate allocation concealment via web‐based programs or sealed opaque envelopes (Al‐Qadheeb 2016; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Page 2017; Reade 2009; Reade 2016; van Eijk 2010).

Blinding

Eleven studies have low risk of performance bias given blind design and explicit discussion of blinded study participants, clinicians, or study personnel (including outcome assessors) (Al‐Qadheeb 2016; Bakri 2015; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Page 2017; Reade 2016; van Eijk 2010). We judged one study to have unclear risk of blinding bias as no details of blinding were available (Atalan 2013). We judged two studies to have high risk of bias as these trials lacked blinding (Reade 2009; Skrobik 2004). All trials but one had blinded outcome assessment (Reade 2009).

Incomplete outcome data

We judged all studies but two to have low risk of attrition bias as they accounted for all screened, enrolled, and randomized participants (Atalan 2013; Bakri 2015), and all except one employed an intention‐to‐treat principle in their analyses (Skrobik 2004). Two studies used a modified intention‐to‐treat analysis (e.g. modification permitted to account for post‐randomization circumstances that prevented use of data from certain participants) (Reade 2016; van Eijk 2010). We judged two studies to have unclear risk of attrition bias because they did not include figures, tables, or text outlining the numbers of participants who were screened, enrolled, and randomized, and/or who successfully completed the study protocol (Atalan 2013; Bakri 2015).

Selective reporting

We judged eleven studies to have low risk of reporting bias based on examination of their respective trial registration or published protocols (Al‐Qadheeb 2016; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Page 2017; Reade 2009; Reade 2016; van Eijk 2010). The remaining trials were deemed at unclear risk, as trial registrations or protocols were not available to confirm outcome reporting.

Other potential sources of bias

We judged all studies but two to have low risk of other potential sources of bias (Atalan 2013; Bakri 2015). All studies cited funding sources, except Atalan 2013, which provided no funding details. Two studies were conducted without external funding (Bakri 2015; Hakim 2012). Study support for a pharmaceutical company was declared in seven studies; however all stated that these companies had no involvement in study design, data collection, analysis, or data reporting (Devlin 2010; Girard 2010a; Needham 2016; Reade 2009; Reade 2016; Skrobik 2004; van Eijk 2010).

Effects of interventions

See: Summary of findings for the main comparison Duration of delirium; Summary of findings 2 Days with coma; Summary of findings 3 Duration of mechanical ventilation; Summary of findings 4 Length of ICU stay; Summary of findings 5 Length of hospital stay; Summary of findings 6 QTc prolongation

See Summary of findings tables (summary of findings Table for the main comparison; summary of findings Table 2; summary of findings Table 3; summary of findings Table 4; summary of findings Table 5; summary of findings Table 6). The 'Summary of findings' tables provide overall estimates of treatment effects compared with placebo. We summarize the quality of evidence for delirium duration, delirium‐free and coma‐free days, duration of mechanical ventilation, and length of ICU stay obtained through pairwise comparisons and NMA.

Geometry of evidence networks by endpoints

For all outcomes, most trials compared one active intervention (drug) against placebo; trials involving comparisons between active interventions were rare. The number of participants enrolled for each active therapy was small in the networks compared to the number enrolled for placebo comparisons. Figure 4 (panel A to F) presents the network diagrams that indicate corresponding eligible regimens in the evidence network for each outcome. Each line links treatments directly compared across studies. Head‐to‐head trials were available for 7/21 (33%) of the pairwise comparisons for delirium duration, 6/15 (40%) for duration of mechanical ventilation, 7/21 (33%) for hospital length of stay, 7/21 (33%) for ICU length of stay, 4/6 (67%) for delirium‐free and coma‐free days, and 4/6 (67%) for days in coma.


Network diagrams of pairwise comparisons for the six outcomes with network meta‐analyses.

Network diagrams of pairwise comparisons for the six outcomes with network meta‐analyses.

Primary outcome
1. Duration of delirium

Pairwise meta‐analysis (direct comparisons)

Data from a total of 11 trials (n = 1530 participants) contributed to the analysis of duration of delirium (Al‐Qadheeb 2016; Atalan 2013; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Page 2017; Reade 2016; van Eijk 2010); ten RCTs (n = 1477 participants) were placebo‐controlled. Treatment effect estimates from pairwise meta‐analyses are reported in Analysis 1.1 and Figure 5. Pairwise meta‐analyses showed that the alpha2 agonist dexmedetomidine may be associated with a shorter duration of delirium (ratio of means (RoM) 0.58, 95% confidence interval (CI) 0.43 to 0.79; 71 participants; 1 study), and the cholinesterase inhibitor rivastigmine may be associated with a longer duration of delirium (RoM 1.84, 95% CI 1.25 to 2.69; 104 participants; 1 study) compared to placebo. The pairwise meta‐analyses showed no effect on the duration of delirium for typical antipsychotics (RoM 1.02, 95% CI 0.91 to 1.14; 608 participants; 4 studies), atypical antipsychotics (RoM 0.73, 95% CI 0.49 to 1.11; 500 participants; 4 studies), or statins (RoM 1.07, 95% CI 0.91 to 1.25; 414 participants; 2 studies) compared to placebo.


Forest plot of comparison: 1 Duration of delirium (log units), outcome: 1.1 Duration of delirium (log units).

Forest plot of comparison: 1 Duration of delirium (log units), outcome: 1.1 Duration of delirium (log units).

NMA (combinations of direct and indirect comparisons)

The random‐effects consistency model was an adequate fit, with posterior total residual deviance of 27.33 (compared to 24 unconstrained data points). The forest plot in Figure 6 presents the ratio of means (RoM) estimates for each intervention compared to placebo derived from the random‐effects consistency model, along with 95% credible intervals (CrIs). For all interventions compared to placebo, 95% CrIs were wide and failed to rule out the possibility of no difference. The intervention with the smallest RoM (i.e. most preferred) was the alpha2 agonist dexmedetomidine (RoM 0.58, 95% CrI 0.26 to 1.27; SUCRA 0.895; moderate‐quality evidence) (Table 1). In order of descending surface under the cumulative ranking curve (SUCRA) values (best to worst; Table 2), the next best interventions were atypical antipsychotics (RoM vs placebo 0.80, 95% CrI 0.50 to 1.11; SUCRA 0.738; moderate‐quality evidence), opioids (RoM vs placebo 0.88, 95% CrI 0.37 to 2.01; SUCRA 0.578; very low‐quality evidence), typical antipsychotics (RoM vs placebo 0.96, 95% CrI 0.64 to 1.36; SUCRA 0.468; high‐quality evidence), placebo (SUCRA 0.403), statins (RoM vs placebo 1.05, 95% CrI 0.61 to 1.77; SUCRA 0.365; moderate‐quality evidence), and the cholinesterase inhibitor rivastigmine (RoM vs placebo 1.84, 95% CrI 0.82 to 4.10; SUCRA 0.054; moderate‐quality evidence). In addition to comparisons versus placebo, Table 1 shows the comparisons between active interventions. As an example of interpretation, the RoM estimate of 0.58 (95% CrI 0.26 to 1.27) in the lower triangle suggests a 42% reduction in the mean duration of delirium with alpha2 agonists compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 93.8% that alpha2 agonists are better than placebo in terms of duration of delirium. The between‐study SD, as a measure of heterogeneity, was estimated to be 0.29 (95% CrI 0.03 to 0.75). Comparison of DIC values between the random‐effects consistency model (‐14.66) and the corresponding random‐effects unrelated means model (DIC ‐14.55), as well as inspection of a scatterplot of posterior mean deviance contributions from both models (Figure 7), suggested no violation of the consistency assumption. Our inspection of Gelman‐Rubin‐Brooks diagnostics and potential scale reduction factors for all NMAs confirmed convergence with 200,000 iterations in all cases (among which 100,000 were burn‐in).


Findings from network meta‐analysis: duration of delirium, delirium‐free and coma‐free days, and days with coma.

Findings from network meta‐analysis: duration of delirium, delirium‐free and coma‐free days, and days with coma.


Consistency assumption check: posterior mean deviance contribution plots for RE consistency model vs unrelated means model; they did not suggest violation of the consistency assumption.

Consistency assumption check: posterior mean deviance contribution plots for RE consistency model vs unrelated means model; they did not suggest violation of the consistency assumption.

Open in table viewer
Table 1. Duration of delirium: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)

Alpha2
agonist

0.807

0.819

0.907

0.922

0.977

0.938

0.72
(0.33, 1.87)

Atypical

antipsychotic

0.610

0.858

0.861

0.975

0.927

0.66
(0.21, 2.13)

0.92
(0.36, 2.11)

Opioid

0.627

0.685

0.924

0.660

0.60
(0.26, 1.48)

0.84
(0.50, 1.25)

0.92
(0.42, 1.97)

Typical

antipsychotic

0.646

0.945

0.607

0.55
(0.22, 1.43)

0.77
(0.37, 1.38)

0.84
(0.30, 2.25)

0.92
(0.47, 1.73)

Statin

0.909

0.396

0.31
(0.10, 0.97)

0.44
(0.17, 1.00)

0.48
(0.14, 1.51)

0.52
(0.21, 1.25)

0.57
(0.22, 1.49)

CHE

Inhibitor

0.054

0.58
(0.26, 1.27)

0.80
(0.50, 1.11)

0.88
(0.37, 2.01)

0.96
(0.64, 1.36)

1.05
(0.61, 1.77)

1.84
(0.82, 4.10)

Placebo

CHE: cholinesterase.

Crl: credible interval.

RoM: ratio of means.

A complete summary of estimates for efficacy from the random‐effects (RE) consistency model assuming vague priors is displayed.

Treatments other than placebo are in the order of decreasing surface under the cumulative ranking curve (SUCRA) value from upper left to lower right. For each comparison, the lower/right‐most treatment is the reference treatment. For example, the RoM estimate of 0.58 (95% CrI 0.26 to 1.27) in the lower triangle suggests a 42% reduction in the mean duration of delirium with alpha2 agonists compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 93.8% that alpha2 agonists are better than placebo in terms of duration of delirium. Estimates which ruled out the possibility of no difference based on pairwise RoM estimates are shown in bold font.

Open in table viewer
Table 2. Duration of delirium: mean SUCRA value, mean probability to be the best, and mean rank for each treatment

RE consistency model

Mean SUCRA

Mean Pr(best)

Mean ranka

Alpha2 agonist

0.895

0.717

1.63 (1 to 6)

Atypical antipsychotic

0.738

0.114

2.57 (1 to 5)

Opioid

0.578

0.129

3.53 (1 to 7)

Typical antipsychotic

0.468

0.010

4.19 (2 to 6)

Placebo

0.403

0.001

4.58 (3 to 6)

Statin

0.365

0.023

4.81 (2 to 7)

CHE inhibitor

0.054

0.006

6.68 (3 to 7)

aMean rank with 2.5% and 97.5% quantiles in parentheses.

CHE: cholinesterase.

Pr: probability.

RE: random‐effects.

SUCRA: surface under the cumulative ranking curve.

Secondary outcomes
1. a) Delirium‐free and coma‐free days

Pairwise meta‐analysis (direct comparisons)

Data from a total of four trials (n = 950 participants) contributed to the analysis of delirium‐free and coma‐free days (Girard 2010a; Girard 2018; Page 2013; Page 2017). Studies were placebo‐controlled with antipsychotics ‐ in Girard 2010a,Girard 2018, and Page 2013 ‐ or statins ‐ in Page 2017 ‐ as the intervention. No pairwise comparison resulted in fewer delirium‐free and coma‐free days (Analysis 2.1; Figure 8).


Forest plot of comparison: 2 Delirium‐free and coma‐free days (log units), outcome: 2.1 Delirium‐free and coma‐free days.

Forest plot of comparison: 2 Delirium‐free and coma‐free days (log units), outcome: 2.1 Delirium‐free and coma‐free days.

NMA (combinations of direct and indirect comparisons)

The forest plot in Figure 6 presents estimates for all interventions compared to placebo from the random‐effects consistency model. The random‐effects consistency model was an adequate fit, with posterior total residual deviance of 10.59 (compared to 10 unconstrained data points). The intervention with the largest RoM was atypical antipsychotics (RoM vs placebo 1.31, 95% CrI 0.69 to 2.83; SUCRA 0.845; moderate‐quality evidence). In order of descending SUCRA values, the next best interventions were typical antipsychotics (RoM vs placebo 1.14, 95% CrI 0.64 to 2.16; SUCRA 0.589; moderate‐quality evidence), placebo (SUCRA 0.327), and statins (RoM vs placebo 0.90, 95% CrI 0.32 to 2.52; SUCRA 0.239; moderate‐quality evidence). In addition to comparisons versus placebo, comparisons between active interventions are provided in Table 3, and secondary measures of effect are presented in Table 4. In all cases, 95% CrIs were wide and failed to rule out the possibility of no difference. The between‐study SD, as a measure of heterogeneity, was estimated to be 0.37 (95% CrI 0.02 to 1.42). Comparison of DIC values between the random‐effects consistency model (‐11.23) and the corresponding random‐effects unrelated means model (DIC ‐11.26), as well as inspection of a scatterplot of posterior mean deviance contributions from both models (Figure 7), did not suggest violation of the consistency assumption.

Open in table viewer
Table 3. Delirium‐ and coma‐free days: league table of posterior median pairwise RoM and 95% (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)

Atypical
antipsychotic

0.784

0.854

0.898

1.15
(0.58, 2.40)

Typical

antipsychotic

0.768

0.781

1.46
(0.45, 5.41)

1.27
(0.40, 4.30)

Statin

0.340

1.31
(0.69, 2.83)

1.14
(0.64, 2.16)

0.90
(0.32, 2.52)

Placebo

RoM: ratio of means.

A complete summary of estimates for efficacy from the random‐effects (RE) consistency model assuming vague priors is displayed.

Treatments other than placebo are in the order of decreasing surface under the cumulative ranking curve (SUCRA) value from upper left to lower right. For each comparison, the lower/right‐most treatment is the reference treatment. For example, the RoM estimate of 1.31 (95% credible interval (CrI) 0.69 to 2.83) in the lower triangle suggests a 31% increase in mean delirium‐ and coma‐free days with atypical antipsychotics compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 89.8% that atypical antipsychotics are better than placebo in terms of delirium‐ and coma‐free days.

Open in table viewer
Table 4. Delirium‐ and coma‐free days: mean SUCRA value, mean probability to be the best, and mean rank for each treatment

RE consistency model

Mean SUCRA

Mean Pr(best)

Mean ranka

Atypical antipsychotic

0.845

0.690

1.46 (1 to 4)

Typical antipsychotic

0.589

0.160

2.23 (1 to 4)

Placebo

0.327

0.033

3.02 (1 to 4)

Statin

0.239

0.116

3.28 (1 to 4)

aMean rank with 2.5% and 97.5% quantiles in parentheses.

Pr: probability.

RE: random‐effects.

SUCRA: surface under the cumulative ranking curve.

1. b) Days with coma

Pairwise meta‐analysis (direct comparisons)

Data from a total of five trials (n = 1222 participants) contributed to the analysis of days with coma (Girard 2010a; Girard 2018; Needham 2016; Page 2013; Page 2017). All five studies were placebo‐controlled, with antipsychotics (in Girard 2010a,Girard 2018, and Page 2013) or statins (in Needham 2016 and Page 2017) as the intervention group. No pairwise comparison resulted in fewer days with coma (Analysis 3.1; Figure 9).


Forest plot of comparison: 3 Days with coma (log units), outcome: 3.1 Days with coma (log units).

Forest plot of comparison: 3 Days with coma (log units), outcome: 3.1 Days with coma (log units).

NMA (combinations of direct and indirect comparisons)

The forest plot in Figure 6 presents estimates for each intervention compared to placebo from the random‐effects consistency model. The random‐effects consistency model was an adequate fit, with a posterior total residual deviance of 12.34 (compared to 12 unconstrained data points). The intervention with the smallest RoM versus placebo was typical antipsychotics (RoM vs placebo 0.77, 95% CrI 0.43 to 1.29; SUCRA 0.820; low‐quality evidence). In order of descending SUCRA values, the next best interventions were statins (RoM vs placebo 0.92, 95% CrI 0.49 to 1.80; SUCRA 0.481; moderate‐quality evidence), atypical antipsychotics (RoM vs placebo 0.94, 95% CrI 0.48 to 1.72; SUCRA 0.422; moderate‐quality evidence), and placebo (SUCRA 0.278). In addition to comparisons versus placebo, comparisons between active interventions are provided in Table 5, and secondary measures of effect are presented in Table 6. In all cases, 95% CrIs were wide and failed to rule out the possibility of no difference. The between‐study SD, as a measure of heterogeneity, was estimated to be 0.34 (95% CrI 0.03 to 1.08). Comparison of DIC values between the random‐effects consistency model (‐5.32) and the corresponding random‐effects unrelated means model (‐5.33), as well as inspection of a scatterplot of posterior mean deviance contributions from both models (Figure 7), did not suggest violation of the consistency assumption.

Open in table viewer
Table 5. Days with coma: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)

Typical

antipsychotic

0.740

0.815

0.905

0.83
(0.34, 1.87)

Statin

0.532

0.651

0.81
(0.44, 1.56)

0.98
(0.41, 2.58)

Atypical

antipsychotic

0.612

0.77
(0.43, 1.29)

0.92
(0.49, 1.80)

0.94
(0.48, 1.72)

Placebo

Crl: credible interval.

RoM: ratio of means.

A complete summary of estimates for efficacy from the random‐effects (RE) consistency model assuming vague priors is displayed.

Treatments other than placebo are in the order of decreasing surface under the cumulative ranking curve (SUCRA) value from upper left to lower right. For each comparison, the lower/right‐most treatment is the reference treatment. For example, the RoM estimate of 0.77 (95% CrI 0.43 to 1.29) in the lower triangle suggests a 23% reduction in mean coma days with typical antipsychotics compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 90.5% that typical antipsychotics are better than placebo in terms of days with coma.

Open in table viewer
Table 6. Days with coma: mean SUCRA values, mean probability to be the best, and mean rank for each treatment

RE consistency model

Mean SUCRA

Mean Pr(best)

Mean ranka

Typical antipsychotic

0.820

0.620

1.54 (1 to 4)

Statin

0.481

0.222

2.56 (1 to 4)

Atypical antipsychotic

0.422

0.132

2.73 (1 to 4)

Placebo

0.278

0.026

3.17 (1 to 4)

aMean rank with 2.5% and 97.5% quantiles in parentheses.

Pr: probability.

RE: random‐effects.

SUCRA: surface under the cumulative ranking curve

2. Relapse of delirium (% patients)

No study reported data on this outcome.

3. Resolution of delirium symptoms (e.g. hallucinations, agitation)

No study reported data on resolution of delirium symptoms as a specific outcome. Agitation was reported in three studies (Al‐Qadheeb 2016; Devlin 2010; Page 2013). Al‐Qadheeb 2016 found that the haloperidol group spent fewer hours per study day agitated (Sedation Agitation Scale ≥ 5) compared to the placebo group (median 0 vs 2; P = 0.008). Similarly, Devlin 2010 found that quetiapine was associated with fewer hours of agitation (SAS ≥ 5) compared to placebo (6 vs 36; P = 0.02). Page 2013 found that a smaller proportion of participants had agitated Richmond Agitation‐Sedation Scale (RASS) scores (RASS > 2+) in the first 14 days of the study in the haloperidol group compared to the placebo group (median 13% vs 20%; P = 0.0075).

4. Duration of mechanical ventilation

Pairwise meta‐analysis (direct comparisons)

Data from a total of seven trials (n = 1167 participants) contributed to the analysis of duration of mechanical ventilation (Al‐Qadheeb 2016; Atalan 2013; Devlin 2010; Girard 2010a; Girard 2018; Needham 2016; Reade 2016); all but one study was placebo‐controlled (Atalan 2013). Trials evaluated dexmedetomidine (Reade 2016), antipsychotics (Al‐Qadheeb 2016, Devlin 2010; Girard 2010a; Girard 2018), opioids (Atalan 2013), and statins (Needham 2016). For meta‐analysis and network meta‐analysis, we could not include Page 2013 or Reade 2009 in the syntheses of mechanical ventilation duration. We excluded Page 2013 due to missing SD. We excluded Reade 2009 as the addition of this trial resulted in problems with the consistency equation. We judged there to be important differences in study populations (i.e. cardiovascular surgery commonly associated with short ICU stays) that explained the disruption of the consistency equation. Amongst the pairwise comparisons versus placebo (Analysis 4.1; Figure 10), dexmedetomidine was associated with a reduced duration of mechanical ventilation (RoM 0.55, 95% CI 0.41 to 0.75; 71 participants; 1 study), and typical antipsychotics (RoM 0.92, 95% CI 0.79 to 1.06; 515 participants; 3 studies), atypical antipsychotics (RoM 0.98, 95% CI 0.84 to 1.34; 476 participants; 3 studies), and statins (RoM 1.09, 95% CI 0.90 to 1.34; 272 participants; 1 study) did not.


Forest plot of comparison: 4 Duration of mechanical ventilation (log units), outcome: 4.1 Duration of mechanical ventilation (log units).

Forest plot of comparison: 4 Duration of mechanical ventilation (log units), outcome: 4.1 Duration of mechanical ventilation (log units).

NMA (combinations of direct and indirect comparisons)

The forest plot in Figure 11 presents RoM estimates for each intervention compared to placebo from the random‐effects consistency model. The random‐effects consistency model was an adequate fit, with posterior total residual deviance of 14.13 (compared to 16 unconstrained data points). The intervention with the smallest RoM versus placebo was dexmedetomidine (RoM 0.55, 95% CrI 0.34 to 0.89; SUCRA 0.974; moderate‐quality evidence). In order of descending SUCRA values, the next best interventions were typical antipsychotics (RoM 0.93, 95% CI 0.72 to 1.24; SUCRA 0.576; moderate‐quality evidence), atypical antipsychotics (RoM 0.98, 95% CI 0.71 to 1.28; SUCRA 0.440; moderate‐quality evidence), opioids (RoM vs placebo 0.99, 95% CrI 0.58 to 1.76; SUCRA 0.410; very low‐quality evidence), placebo (SUCRA 0.377), and statins (RoM vs placebo 1.10, 95% CrI 0.71 to 1.69; SUCRA 0.223; moderate‐quality evidence). Comparisons between active interventions are provided in Table 7, and secondary measures of effect are presented in Table 8. The between‐study SD, as a measure of heterogeneity, was estimated to be 0.14 (95% CrI 0.005 to 0.53). Comparison of DIC values between the random‐effects consistency model (‐15.16) and the corresponding random‐effects unrelated means model (‐15.24), as well as inspection of a scatterplot of posterior mean deviance contributions from both models (Figure 7), did not suggest violation of the consistency assumption.


Findings from network meta‐analysis: duration of mechanical ventilation, length of ICU and hospital stay.

Findings from network meta‐analysis: duration of mechanical ventilation, length of ICU and hospital stay.

Open in table viewer
Table 7. Duration of mechanical ventilation: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)

Alpha2
agonist

0.973

0.973

0.958

0.978

0.986

0.59
(0.34, 1.01)

Typical

antipsychotic

0.665

0.628

0.805

0.754

0.57
(0.33, 1.02)

0.95
(0.72, 1.35)

Atypical

antipsychotic

0.527

0.729

0.582

0.56
(0.26, 1.14)

0.94
(0.58, 1.52)

0.98
(0.53, 1.70)

Opioid

0.645

0.517

0.50
(0.26, 0.97)

0.85
(0.52, 1.45)

0.89
(0.52, 1.47)

0.90
(0.46, 1.87)

Statin

0.274

0.55
(0.34, 0.89)

0.93
(0.72, 1.24)

0.98
(0.71, 1.28)

0.99
(0.58, 1.76)

1.10
(0.71, 1.69)

Placebo

Crl: credible interval.

RoM: ratio of means.

A complete summary of estimates for efficacy from the random‐effects (RE) consistency model assuming vague priors is displayed.

Treatments other than placebo are in the order of decreasing surface under the cumulative ranking curve (SUCRA) value from upper left to lower right. For each comparison, the lower/right‐most treatment is the reference treatment. For example, the RoM estimate of 0.55 (95% CrI 0.34 to 0.89) in the lower triangle suggests a 45% reduction in the mean duration of mechanical ventilation with alpha2 agonists compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 98.6% that alpha2 agonists are better than placebo for the duration of mechanical ventilation. Estimates which ruled out the possibility of no difference based on pairwise RoM estimates are shown in bold font.

Open in table viewer
Table 8. Duration of mechanical ventilation: mean SUCRA values, mean probability to be the best, and mean rank for each treatment

RE consistency model

Mean SUCRA

Mean Pr(best)

Mean ranka

Alpha2 agonists

0.974

0.931

1.13 (1 to 3)

Typical antipsychotic

0.576

0.009

3.12 (2 to 6)

Atypical antipsychotic

0.440

0.012

3.80 (2 to 6)

Opioid

0.410

0.033

3.95 (1 to 6)

Placebo

0.377

0.001

4.11 (2 to 6)

Statin

0.223

0.014

4.88 (2 to 6)

aMean rank with 2.5% and 97.5% quantiles in parentheses.

Pr: probability.

RE: random‐effects.

SUCRA: surface under the cumulative ranking curve.

5. a) Length of ICU stay

Pairwise meta‐analysis (direct comparisons)

Data from a total of 10 trials (n = 1475 participants) contributed to the analysis of length of ICU stay (Al‐Qadheeb 2016; Atalan 2013; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Reade 2016; van Eijk 2010); all but one trial were placebo‐controlled (Atalan 2013). Atypical antipsychotics were associated with significantly reduced length of ICU stay (RoM 0.91, 95% CI 0.84 to 1.00; 577 participants; 4 studies), and the cholinesterase inhibitor rivastigmine was associated with significantly increased length of ICU stay (RoM 2.18, 95% CI 1.58 to 3.03; 104 participants; 1 study) compared to placebo (Analysis 5.1;Figure 12). No difference was found for typical antipsychotics (RoM 1.01, 95% CI 0.90 to 1.14; 618 participants; 4 studies), statins (RoM 1.06, 95% CI 0.91 to 1.23; 272 participants; 1 study), or alpha2 agonists (RoM 0.80, 95% CI 0.59 to 1.08; 71 participants; 1 study) compared to placebo.


Forest plot of comparison: 5 Length of ICU stay (log units), outcome: 5.1 Length of ICU stay (log units).

Forest plot of comparison: 5 Length of ICU stay (log units), outcome: 5.1 Length of ICU stay (log units).

NMA (combinations of direct and indirect comparisons)

The forest plot in Figure 11 presents RoM estimates for each intervention compared to placebo from the random‐effects consistency model. The random‐effects consistency model was an adequate fit, with posterior total residual deviance of 20.45 (compared to 22 unconstrained data points). The cholinesterase inhibitor rivastigmine was found to have longer length of ICU stay compared to placebo, and all remaining comparisons showed wide 95% CrIs that failed to rule out the possibility of no difference. The intervention with the smallest RoM versus placebo was dexmedetomidine (RoM 0.80, 95% CrI 0.55 to 1.17; SUCRA 0.853; low‐quality evidence). In order of descending SUCRA values, the next best interventions were atypical antipsychotics (RoM vs placebo 0.92, 95% CrI 0.80 to 1.08; SUCRA 0.709; high‐quality evidence), opioids (RoM vs placebo 0.92, 95% CrI 0.62 to 1.40; SUCRA 0.639; very low‐quality evidence), typical antipsychotics (RoM vs placebo 0.99, 95% CrI 0.85 to 1.17; SUCRA 0.496; moderate‐quality evidence), placebo (SUCRA 0.457), statins (RoM vs placebo 1.07, 95% CrI 0.81 to 1.41; SUCRA 0.344; low‐quality evidence), and cholinesterase inhibitors (RoM vs placebo 2.19, 95% CrI 1.47 to 3.27; SUCRA 0.002; moderate‐quality evidence). Comparisons between active interventions are provided in Table 9, and secondary measures of effect are presented in Table 10. The between‐study SD, as a measure of heterogeneity, was estimated to be 0.09 (95% CrI 0.003 to 0.28). Comparison of DIC values between the random‐effects consistency model (‐27.94) and the corresponding random‐effects unrelated means model (‐27.97), as well as inspection of a scatterplot of posterior mean deviance contributions from both models (Figure 7), did not suggest violation of the consistency assumption.

Open in table viewer
Table 9. Length of ICU stay: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)

Alpha2 agonist

0.766

0.705

0.858

0.902

0.999

0.886

0.87
(0.58, 1.29)

Atypical

antipsychotic

0.499

0.792

0.856

0.999

0.874

0.87
(0.50, 1.49)

1.00
(0.65, 1.51)

Opioid

0.647

0.736

0.996

0.658

0.81
(0.54, 1.21)

0.93
(0.77, 1.12)

0.93
(0.64, 1.36)

Typical

antipsychotic

0.714

0.998

0.559

0.75
(0.47, 1.20)

0.87
(0.64, 1.20)

0.87
(0.53, 1.44)

0.93
(0.68, 1.29)

Statin

0.995

0.278

0.37
(0.21, 0.63)

0.42
(0.28, 0.65)

0.42
(0.24, 0.75)

0.45
(0.30, 0.70)

0.49
(0.30, 0.79)

CHE inhibitor

0.001

0.80
(0.55, 1.17)

0.92
(0.80, 1.08)

0.92
(0.62, 1.40)

0.99
(0.85, 1.17)

1.07
(0.81, 1.41)

2.19
(1.47, 3.27)

Placebo

CHE: cholinesterase.

Crl: credible interval.

ICU: intensive care unit.

RoM: ratio of means.

A complete summary of estimates for efficacy from the random‐effects (RE) consistency model assuming vague priors is displayed.

Treatments other than placebo are in the order of decreasing surface under the cumulative ranking curve (SUCRA) value from upper left to lower right. For each comparison, the lower/right‐most treatment is the reference treatment. For example, the RoM estimate of 0.80 (95% CrI 0.55 to 1.17) in the lower triangle suggests a 20% reduction in mean length of ICU stay with alpha2 agonists compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 88.6% that alpha2 agonists are better than placebo for the length of ICU stay. Estimates which ruled out the possibility of no difference based on pairwise RoM estimates are shown in bold font.

Open in table viewer
Table 10. Length of ICU stay: mean SUCRA value, mean probability to be the best, and mean rank for each treatment

RE consistency model

Mean SUCRA

Mean Pr(best)

Mean ranka

Alpha2 agonists

0.853

0.608

1.88 (1 to 6)

Atypical antipsychotic

0.709

0.106

2.75 (1 to 5)

Opioid

0.639

0.238

3.17 (1 to 6)

Typical antipsychotic

0.496

0.014

4.02 (2 to 6)

Placebo

0.457

0.004

4.26 (2 to 6)

Statin

0.344

0.030

4.93 (1 to 6)

CHE inhibitor

0.002

0.000

6.99 (7 to 7)

aMean rank with 2.5% and 97.5% quantiles in parentheses.

CHE: cholinesterase.

ICU: intensive care unit.

Pr: probability.

RE: random‐effects.

SUCRA: surface under the cumulative ranking curve.

5. b) Length of hospital stay

Pairwise meta‐analysis (direct comparisons)

Data from a total of nine trials (n = 1403 participants) contributed to the analysis of length of hospital stay (Atalan 2013; Devlin 2010; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Page 2017; Reade 2016; van Eijk 2010); eight studies were placebo‐controlled. No pairwise comparison was statistically significant (Analysis 6.1;Figure 13).


Forest plot of comparison: 6 Length of hospital stay (log units), outcome: 6.1 Length of hospital stay (log units).

Forest plot of comparison: 6 Length of hospital stay (log units), outcome: 6.1 Length of hospital stay (log units).

NMA (combination of direct and indirect comparisons)

The forest plot in Figure 11 presents RoM estimates for each intervention compared to placebo from the random‐effects consistency model. The random‐effects consistency model was an adequate fit, with a posterior total residual deviance of 19.07 (compared to 19 unconstrained data points). The intervention with the smallest RoM versus placebo was typical antipsychotics (RoM 0.92, 95% CrI 0.65 to 1.18; SUCRA 0.722; low‐quality evidence). In order of descending SUCRA values, the next best interventions were atypical antipsychotics (RoM vs placebo 0.93, 95% CrI 0.69 to 1.16; SUCRA 0.693; moderate‐quality evidence), statins (RoM vs placebo 0.98, 95% CrI 0.69 to 1.30; SUCRA 0.537; moderate‐quality evidence), opioids (RoM vs placebo 0.97, 95% CrI 0.55 to 1.60; SUCRA 0.532; very low‐quality evidence), placebo (SUCRA 0.435), dexmedetomidine (RoM vs placebo 1.10, 95% CrI 0.69 to 1.75; SUCRA 0.301; moderate‐quality evidence), and rivastigmine (RoM vs placebo 1.11, 95% CrI 0.70 to 1.77; SUCRA 0.280; moderate‐quality evidence). Comparisons between active interventions are provided in Table 11, and secondary measures of effect are presented in Table 12. In all cases, 95% CrIs were wide and failed to rule out the possibility of no difference. The between‐study SD, as a measure of heterogeneity, was estimated to be 0.15 (95% CrI 0.005 to 0.53). Comparison of DIC values between the random‐effects consistency model (‐27.19), and the corresponding random‐effects unrelated means model (‐27.32), as well as inspection of a scatterplot of posterior mean deviance contributions from both models (Figure 7), did not suggest violation of the consistency assumption.

Open in table viewer
Table 11. Length of hospital stay: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)

Typical
antipsychotic

0.546

0.673

0.654

0.810

0.828

0.820

0.99
(0.70, 1.38)

Atypical

antipsychotic

0.651

0.604

0.801

0.818

0.827

0.94
(0.60, 1.43)

0.95
(0.63, 1.42)

Statin

0.498

0.716

0.736

0.596

0.94
(0.60, 1.47)

0.95
(0.55, 1.67)

1.00
(0.55, 1.88)

Opioid

0.685

0.700

0.561

0.83
(0.46, 1.39)

0.84
(0.48, 1.38)

0.88
(0.49, 1.52)

0.88
(0.42, 1.73)

Alpha2 agonist

0.515

0.305

0.82
(0.46, 1.37)

0.83
(0.47, 1.36)

0.88
(0.49, 1.49)

0.87
(0.41, 1.70)

0.99
(0.51, 1.90)

CHE inhibitor

0.278

0.92
(0.65, 1.18)

0.93
(0.69, 1.16)

0.98
(0.69, 1.30)

0.97
(0.55, 1.60)

1.10
(0.69, 1.75)

1.11
(0.70, 1.77)

Placebo

CHE: cholinesterase.

Crl: credible interval.

RoM: ratio of means.

Treatments other than placebo are in the order of decreasing surface under the cumulative ranking curve (SUCRA) value from upper left to lower right. For each comparison, the lower/right‐most treatment is the reference treatment. For example, the RoM estimate of 0.92 (95% CrI 0.65 to 1.18) in the lower triangle suggests an 8% reduction in mean length of hospital stay with typical antipsychotics compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 82% that typical antipsychotics are better than placebo for length of hospital stay.

Open in table viewer
Table 12. Length of hospital stay: mean SUCRA value, mean probability to be the best, and mean rank for each treatment

RE consistency model

Mean SUCRA

Mean Pr(best)

Mean ranka

Typical antipsychotic

0.722

0.235

2.67 (1 to 6)

Atypical antipsychotic

0.693

0.218

2.84 (1 to 6)

Statin

0.537

0.147

3.78 (1 to 7)

Opioid

0.532

0.225

3.81 (1 to 7)

Placebo

0.435

0.008

4.39 (2 to 6)

Alpha2 agonists

0.301

0.090

5.19 (1 to 7)

CHE inhibitor

0.280

0.078

5.32 (1 to 7)

aMean rank with 2.5% and 97.5% quantiles in parentheses.

CHE: cholinesterase.

Pr: probability.

RE: random‐effects.

SUCRA: surface under the cumulative ranking curve.

6. Mortality

Pairwise meta‐analysis (direct comparisons)

Data from a total of 10 trials (n = 1584 participants) contributed to the analysis of mortality (Al‐Qadheeb 2016; Atalan 2013; Devlin 2010; Girard 2010a; Girard 2018; Hakim 2012; Needham 2016; Page 2013; Page 2017; van Eijk 2010). Studies were placebo‐controlled with cholinesterase inhibitors, typical and atypical antipsychotics, and statins as the interventions assessed, except for one trial (Atalan 2013), which compared opioids with typical antipsychotics. Mortality was reported at various time points and settings (e.g. 14 day, 28 day, ICU, hospital). No comparisons were statistically significant (Analysis 7.1;Figure 14).


Forest plot of comparison: 7 Mortality, outcome: 7.1 Mortality.

Forest plot of comparison: 7 Mortality, outcome: 7.1 Mortality.

NMA (combinations of direct and indirect comparisons)

We planned to perform an NMA for mortality. However, due to variability in follow‐up duration and settings across comparisons in the network, NMA was judged by the research team to be inappropriate and thus was not pursued. The disconnected network of interventions for the setting of ICU mortality alone (or hospital mortality alone) made NMA infeasible.

7. Use of physical restraint

Two studies reported on physical restraint application but used different outcome measures that were not amenable to meta‐analysis. Reade 2009 reported that 8/10 participants in the antipsychotic group were restrained compared to 9/10 participants in the alpha2 agonist group (no statistical difference), and van Eijk 2010 reported the percentage of days on which participants were restrained (no difference between groups was observed: 1% placebo and 1% cholinesterase inhibitor, respectively).

8. Hospital discharge disposition

Three studies reported on patient discharge disposition with insufficient information for pooling (Al‐Qadheeb 2016; Devlin 2010; Reade 2016). Devlin 2010 reported the combined outcome of home or rehabilitation facility (89% quetiapine vs 56% placebo; P = 0.06). Al‐Qadheeb 2016 reported no overall statistical difference in the percentage of participants discharged home (41.2% haloperidol vs 26.5% placebo), to a rehabilitation facility (29.4% haloperidol vs 47.1% placebo), or to long‐term care (2.9% haloperidol vs 2.9% placebo). Finally, Reade 2016 reported the percentage of participants transferred to rehabilitation facilities (13.2% dexmedetomidine vs 9.7% placebo; P = 0.65).

9. Long‐term cognitive outcome

This outcome was reported for only one trial (Page 2017). Study investigators assessed cognitive outcomes at six months using the Brief Test of Adult Cognition by Telephone (BTACT) (Lachman 2008). The BTACT assesses multiple dimensions central to effective cognitive functioning (e.g. episodic memory, reasoning, executive function). They also compared the Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE) at baseline versus data at six‐month follow‐up (Jorm 1994). BTACT composite scores and differences between the IQCODE at baseline and at six‐month follow‐up did not differ between the two groups.

10. Health‐related quality of life

No study reported this outcome.

11. Adverse events ‐ a) Akathisia

Pairwise meta‐analysis (direct comparisons)

Akathisia was reported in two trials comparing antipsychotics to placebo (Girard 2010a; Page 2013). The overall number of participants was low (N = 242), as was the number of events. Akathisia was assessed subjectively with a 10‐cm visual analogue scale (Girard 2010a), or it was not specified how assessment was performed (Page 2013). We found no differences in any of the drug pairwise comparisons (Analysis 8.1; Figure 15). We assessed the evidence as low quality.


Forest plot of comparison: 8 Akathisia, outcome: 8.1 Akathisia.

Forest plot of comparison: 8 Akathisia, outcome: 8.1 Akathisia.

NMA (combination of direct and indirect comparisons)

We did not conduct an NMA for this adverse event as only trials investigating antipsychotic drugs reported on this outcome.

11. Adverse events ‐ b) Arrhythmia and QTc prolongation

Pairwise meta‐analysis (direct comparisons)

Arrhythmias were reported as an adverse event in four trials (Girard 2010a; Girard 2018; Page 2013; Reade 2009), and 828 participants were analysed. These trials compared antipsychotics ‐ in Girard 2010a,Girard 2018, and Page 2013 ‐ versus placebo or dexmedetomidine ‐ in Reade 2009. The number of included participants was small, and events were rare. Only typical antipsychotics compared with placebo were associated with significantly increased odds of arrhythmias (OR 3.09, 95% CI 1.11 to 8.62) amongst all pairwise comparisons (Analysis 9.1; Figure 16).


Forest plot of comparison: 9 Arrhythmias, outcome: 9.1 Arrhythmias.

Forest plot of comparison: 9 Arrhythmias, outcome: 9.1 Arrhythmias.

QTc prolongation, measured by electrocardiogram, was reported in seven studies (Al‐Qadheeb 2016; Bakri 2015; Devlin 2010; Girard 2010a; Girard 2018; Page 2013; Reade 2009), and 996 participants were analysed. Trials investigated antipsychotics compared to placebo, ondansetron, and dexmedetomidine. The overall number of participants was small, as was the number or reported events. No comparisons were statistically significant (Analysis 10.1; Figure 17).


Forest plot of comparison: 10 QTc prolongation, outcome: 10.1 QTc prolongation.

Forest plot of comparison: 10 QTc prolongation, outcome: 10.1 QTc prolongation.

NMA (combination of direct and indirect comparisons)

We did not conduct an NMA for this adverse event.

11. Adverse events ‐ c) Extrapyramidal side effects

Pairwise meta‐analysis (direct comparisons)

Extrapyramidal side effects were assessed in six antipsychotic trials (Al‐Qadheeb 2016; Devlin 2010; Girard 2010a; Girard 2018; Page 2013; Skrobik 2004), which included a total of 985 analysed participants. Extrapyramidal symptoms were assessed on the modified Simpson‐Angus Scale in five trials (Devlin 2010; Girard 2010a; Girard 2018; Page 2013; Skrobik 2004), and one trial did not report the assessment method used (Al‐Qadheeb 2016). Pooled results showed no significant differences compared to placebo (Analysis 11.1;Figure 18).


Forest plot of comparison: 11 Extrapyramidal symptoms, outcome: 11.1 Extrapyramidal symptoms.

Forest plot of comparison: 11 Extrapyramidal symptoms, outcome: 11.1 Extrapyramidal symptoms.

NMA (combinations of direct and indirect comparisons)

We planned to perform an NMA for extrapyramidal side effects. However, due to both the rare nature of events in this analysis and violation of the consistency assumption, NMA was judged to be inappropriate.

11. Adverse events ‐ d) Seizures

No trial reported or examined seizures as an outcome.

Subgroup analyses

We planned to explore subgroup analyses or meta‐regression analyses, or both, to address the impact of age, ICU patient population, delirium subtype, and use of non‐drug co‐interventions on our findings to establish their robustness.

Neither subgroup analyses nor meta‐regression analyses were feasible to explore the delirium subtype (e.g. hyperactive, hypoactive, mixed) and use of non‐drug co‐interventions. We did not have a well‐connected evidence network to perform subgroup analyses for studies with mean participant age ≥ 65 years. Subgroup analyses for studies with mean age < 65 years resulted in widened CIs/CrIs for typical and atypical antipsychotics and disappearance of opioids and cholinesterase inhibitors from the evidence network, but did not provide different results compared to overall analyses.

Sensitivity analyses

We explored some sensitivity analyses involving alternative geometries of the network. We did not end up with a well‐connected evidence network for each outcome once high risk of bias trials were excluded. There were insufficient trials to conduct analyses involving alternative geometries based on dose or frequency of drug administration. Planned sensitivity analyses collapsing atypical and typical antipsychotics into one node did not provide different results compared to the overall analyses.

After exclusion of studies that focused on subsyndromal delirium (Al‐Qadheeb 2016; Hakim 2012), CIs/CrIs widened for typical and atypical antipsychotics, but results were not different compared to findings of the overall analyses.

Removal from analysis of studies with patients of low illness severity eliminated alpha2 agonists (dexmedetomidine) and opioids (morphine) from the evidence network (Atalan 2013;Hakim 2012;Reade 2016), but results for the remaining interventions were not different compared to findings of the overall analyses.

Reporting bias

We did not produce a funnel plot for each pairwise comparison. To detect small‐study effects by checking asymmetry per pairwise comparison is not feasible due to the low number of identified trials.

Discussion

available in

Summary of main results

We identified 14 randomized controlled trials (RCTs) that enrolled 1844 adult participants which evaluated pharmacological treatments for delirium in the intensive care unit (ICU). These trials evaluated six different drug classes, primarily comparing one active drug versus placebo. Most trials were small, enrolling fewer than 100 participants. Nine trials scored low risk of bias across all domains; the overall quality of evidence for each outcome was assessed via the GRADE approach; quality ranged from low to high.

Pairwise meta‐analyses showed that only the alpha2 agonist dexmedetomidine (vs placebo) significantly reduced the duration of delirium in critically ill adults with delirium; this was based on a single study with < 100 participants. Network meta‐analysis shows that the smallest ratio of means (vs placebo) was associated with the alpha2 agonist dexmedetomidine, followed by atypical antipsychotics. However, effect sizes for either of the drug classes were neither statistically nor clinically significant. Among secondary outcomes, network meta‐analysis (NMA) revealed that only dexmedetomidine was associated with a shorter duration of mechanical ventilation, and that the cholinesterase (CHE) inhibitor rivastigmine was associated with longer ICU stay. Otherwise, no pharmacological intervention was found to achieve statistical or clinical significance for the secondary outcomes. Analyses of reported adverse drug events found that events were similar to those seen with placebo. The 10 ongoing studies and the six studies awaiting classification that we identified, once published and assessed, may alter the conclusions of this review.

Please notice that the 95% credible intervals from Bayesian NMA results are generally more conservative (wider) than the corresponding 95% confidence intervals from pairwise meta‐analyses. If a pairwise comparison had at least one study contributing direct evidence to NMA and resulted in a 95% credible interval that ruled out the possibility of no difference, the corresponding 95% confidence interval from pairwise meta‐analysis was also significant.

Overall completeness and applicability of evidence

When we designed the protocol (Burry 2015), for several reasons we anticipated at least 20 trials specifically investigating various pharmacological interventions for the treatment of ICU delirium. These reasons included the inclusion of delirium as a quality indicator in care of the elderly, poor outcomes associated with delirium in critically ill patients, and the number of registered trials, as well as the strong recommendations for delirium prevention and treatment provided in the Society of Critical Care Medicine's pain, agitation, and delirium guidelines (Barr 2013). Using strong literature review methods, we identified only 14 published trials that matched our review questions. However, we did identify six studies awaiting classification and 10 ongoing trials, several of which are large‐scale, multi‐centre trials. This suggests that this topic will expand greatly in the next five years. We found that most trials, and those ongoing, examined use of pharmacological interventions commonly given in clinical practice, primarily antipsychotics and the alpha2 agonist dexmedetomidine. We found sufficient data to conduct pairwise comparisons and NMA to answer our primary outcome of interest, but we could not analyse some of the secondary outcomes that we deemed clinically important, as these outcomes were not investigated in any trial (i.e. relapse, resolution of symptoms, long‐term cognitive outcomes, and health‐related quality of life). Nor did we find sufficient information to conduct our planned subgroup analyses on age, ICU population type, delirium subtype, or use of non‐pharmacological co‐interventions.

Quality of the evidence

We scored the risk of bias for each trial and used GRADEpro software to inform the generation of evidence quality statements. Among the 14 RCTs included in this review, nine trials scored low risk of bias across all domains. We judged available evidence to range from low to high quality. Evidence for the primary outcome ‐ duration of delirium ‐ was of moderate to high quality when each drug class was compared to placebo. We most commonly downgraded this evidence for imprecision.

Potential biases in the review process

This review followed Cochrane's systematic review procedures closely, with only minor amendments to the published protocol (Burry 2015). Our search was exhaustive without restrictions; therefore we believe we have evaluated the available evidence in full. The trials included in our review directly examined our chosen population and the primary outcome ‐ duration of delirium ‐ as their primary or secondary outcome. We had originally set the primary outcome to be duration of delirium, defined as time from which it wasfirst identified to when it was first resolved (i.e. screened negative as defined by study authors (e.g. first negative screen, two consecutive screenings)), and our secondary outcome to be duration of delirium (as defined by study authors). We found far more variability in the definition of the outcome used than we had anticipated; thus we ended up reporting only the duration of delirium for pooling of results. The definition applied by study authors also varied, with some using 24 hours without delirium, some 48 hours, and others not reporting the definition they applied.

For continuous outcomes, we approximated means and standard deviations (SDs) from medians and interquartile ranges (IQRs) (Wan 2014) to make use of studies that reported only medians and IQRs for some outcomes. Before all ratio of means (RoM) analyses, we transformed means and SDs to the log scale (Higgins 2008) to overcome various time windows across studies with existing pairwise meta‐analysis and NMA methods, and to make evidence synthesis possible. The first transformation may not always yield accurate RoM estimates for skewed outcomes in small studies. Despite the robust properties of the second transformation for skewed outcomes (Higgins 2008), interpretation of RoM analyses is challenging. We did not approximate means and SDs using any range‐related formulae.

We are not aware of other potential sources of bias.

Agreements and disagreements with other studies or reviews

This is the first NMA examining treatment of delirium for ICU patients, and it is the first Cochrane systematic review examining pharmacological interventions for ICU delirium. We identified two recent systematic reviews examining antipsychotics for prevention or treatment of delirium, or both, in any hospital population (i.e. ICU and non‐ICU) (Kishi 2016; Neufeld 2016). Our findings regarding antipsychotics are consistent with those of Neufeld 2016, in that antipsychotics had no effect on delirium duration when review authors pooled the results of treatment trials. Kishi 2016 conducted a review examining antipsychotics for prevention or treatment of delirium, or both, in any hospital population, including data from four studies that were unpublished or were published in abstract form only. The review by Kishi reported response rate (response rate at the study endpoint examining many different severity and global scales) and did not report on duration of delirium. Pooled results for response rate showed that antipsychotics were superior to placebo and non‐antipsychotic drugs. We identified one Cochrane systematic review on alpha2 agonists for long‐term sedation during mechanical ventilation in critically ill patients, which examined risk of delirium as a secondary outcome (Chen 2015); review authors did not report on duration of delirium nor on other delirium outcomes that we reported. The Chen review found no evidence that dexmedetomidine decreased the risk of delirium (risk ratio 0.85, 95% CI 0.63 to 1.14; seven studies; 1624 participants; low‐quality evidence) compared to traditional sedatives.

Study flow diagram.
Figures and Tables -
Figure 1

Study flow diagram.

Risk of bias graph: review authors' judgements about each risk of bias item presented as percentages across all included studies.
Figures and Tables -
Figure 2

Risk of bias graph: review authors' judgements about each risk of bias item presented as percentages across all included studies.

Risk of bias summary: review authors' judgements about each risk of bias item for each included study.
Figures and Tables -
Figure 3

Risk of bias summary: review authors' judgements about each risk of bias item for each included study.

Network diagrams of pairwise comparisons for the six outcomes with network meta‐analyses.
Figures and Tables -
Figure 4

Network diagrams of pairwise comparisons for the six outcomes with network meta‐analyses.

Forest plot of comparison: 1 Duration of delirium (log units), outcome: 1.1 Duration of delirium (log units).
Figures and Tables -
Figure 5

Forest plot of comparison: 1 Duration of delirium (log units), outcome: 1.1 Duration of delirium (log units).

Findings from network meta‐analysis: duration of delirium, delirium‐free and coma‐free days, and days with coma.
Figures and Tables -
Figure 6

Findings from network meta‐analysis: duration of delirium, delirium‐free and coma‐free days, and days with coma.

Consistency assumption check: posterior mean deviance contribution plots for RE consistency model vs unrelated means model; they did not suggest violation of the consistency assumption.
Figures and Tables -
Figure 7

Consistency assumption check: posterior mean deviance contribution plots for RE consistency model vs unrelated means model; they did not suggest violation of the consistency assumption.

Forest plot of comparison: 2 Delirium‐free and coma‐free days (log units), outcome: 2.1 Delirium‐free and coma‐free days.
Figures and Tables -
Figure 8

Forest plot of comparison: 2 Delirium‐free and coma‐free days (log units), outcome: 2.1 Delirium‐free and coma‐free days.

Forest plot of comparison: 3 Days with coma (log units), outcome: 3.1 Days with coma (log units).
Figures and Tables -
Figure 9

Forest plot of comparison: 3 Days with coma (log units), outcome: 3.1 Days with coma (log units).

Forest plot of comparison: 4 Duration of mechanical ventilation (log units), outcome: 4.1 Duration of mechanical ventilation (log units).
Figures and Tables -
Figure 10

Forest plot of comparison: 4 Duration of mechanical ventilation (log units), outcome: 4.1 Duration of mechanical ventilation (log units).

Findings from network meta‐analysis: duration of mechanical ventilation, length of ICU and hospital stay.
Figures and Tables -
Figure 11

Findings from network meta‐analysis: duration of mechanical ventilation, length of ICU and hospital stay.

Forest plot of comparison: 5 Length of ICU stay (log units), outcome: 5.1 Length of ICU stay (log units).
Figures and Tables -
Figure 12

Forest plot of comparison: 5 Length of ICU stay (log units), outcome: 5.1 Length of ICU stay (log units).

Forest plot of comparison: 6 Length of hospital stay (log units), outcome: 6.1 Length of hospital stay (log units).
Figures and Tables -
Figure 13

Forest plot of comparison: 6 Length of hospital stay (log units), outcome: 6.1 Length of hospital stay (log units).

Forest plot of comparison: 7 Mortality, outcome: 7.1 Mortality.
Figures and Tables -
Figure 14

Forest plot of comparison: 7 Mortality, outcome: 7.1 Mortality.

Forest plot of comparison: 8 Akathisia, outcome: 8.1 Akathisia.
Figures and Tables -
Figure 15

Forest plot of comparison: 8 Akathisia, outcome: 8.1 Akathisia.

Forest plot of comparison: 9 Arrhythmias, outcome: 9.1 Arrhythmias.
Figures and Tables -
Figure 16

Forest plot of comparison: 9 Arrhythmias, outcome: 9.1 Arrhythmias.

Forest plot of comparison: 10 QTc prolongation, outcome: 10.1 QTc prolongation.
Figures and Tables -
Figure 17

Forest plot of comparison: 10 QTc prolongation, outcome: 10.1 QTc prolongation.

Forest plot of comparison: 11 Extrapyramidal symptoms, outcome: 11.1 Extrapyramidal symptoms.
Figures and Tables -
Figure 18

Forest plot of comparison: 11 Extrapyramidal symptoms, outcome: 11.1 Extrapyramidal symptoms.

Comparison 1 Duration of delirium (log ratio of means scale), Outcome 1 Duration of delirium (log units).
Figures and Tables -
Analysis 1.1

Comparison 1 Duration of delirium (log ratio of means scale), Outcome 1 Duration of delirium (log units).

Comparison 2 Delirium‐free and coma‐free days (log ratio of means scale), Outcome 1 Delirium‐free and coma‐free days.
Figures and Tables -
Analysis 2.1

Comparison 2 Delirium‐free and coma‐free days (log ratio of means scale), Outcome 1 Delirium‐free and coma‐free days.

Comparison 3 Days with coma (log ratio of means scale), Outcome 1 Days with coma (log units).
Figures and Tables -
Analysis 3.1

Comparison 3 Days with coma (log ratio of means scale), Outcome 1 Days with coma (log units).

Comparison 4 Duration of mechanical ventilation (log ratio of means scale), Outcome 1 Duration of mechanical ventilation (log units).
Figures and Tables -
Analysis 4.1

Comparison 4 Duration of mechanical ventilation (log ratio of means scale), Outcome 1 Duration of mechanical ventilation (log units).

Comparison 5 Length of ICU stay (log ratio of means scale), Outcome 1 Length of ICU stay (log units).
Figures and Tables -
Analysis 5.1

Comparison 5 Length of ICU stay (log ratio of means scale), Outcome 1 Length of ICU stay (log units).

Comparison 6 Length of hospital stay (log ratio of means scale), Outcome 1 Length of hospital stay (log units).
Figures and Tables -
Analysis 6.1

Comparison 6 Length of hospital stay (log ratio of means scale), Outcome 1 Length of hospital stay (log units).

Comparison 7 Mortality, Outcome 1 Mortality.
Figures and Tables -
Analysis 7.1

Comparison 7 Mortality, Outcome 1 Mortality.

Comparison 8 Akathisia, Outcome 1 Akathisia.
Figures and Tables -
Analysis 8.1

Comparison 8 Akathisia, Outcome 1 Akathisia.

Comparison 9 Arrhythmias, Outcome 1 Arrhythmias.
Figures and Tables -
Analysis 9.1

Comparison 9 Arrhythmias, Outcome 1 Arrhythmias.

Comparison 10 QTc prolongation, Outcome 1 QTc prolongation.
Figures and Tables -
Analysis 10.1

Comparison 10 QTc prolongation, Outcome 1 QTc prolongation.

Comparison 11 Extrapyramidal symptoms, Outcome 1 Extrapyramidal symptoms.
Figures and Tables -
Analysis 11.1

Comparison 11 Extrapyramidal symptoms, Outcome 1 Extrapyramidal symptoms.

Summary of findings for the main comparison. Duration of delirium

Outcome: duration of delirium

Patient or population: critically ill adults with confirmed or at high risk of delirium
Settings: intensive care units in Australia and New Zealand, Canada, Egypt, Netherlands, Turkey, USA, UK
Intervention: any pharmacological intervention

Control: placebo or active comparator

Comparisons

Illustrative comparative risks* (95% CrI)

Ratio of means (RoM) based on log RoM estimates from meta‐analysis

(IV, random, 95% CI)

Number of participants
(studies)

Quality of the evidence
(GRADE) based on NMA

NMA results (assuming consistency equations)

Assumed risk

Corresponding risk based on NMA estimates

Placebo/Comparator

Intervention drug

Typical antipsychotic vs placebo

Median duration of delirium: 3 to 5 days for placebo

3.86 days of delirium (95% CrI 2.57 to
5.46) corresponding to 4 days in the placebo group

RoM: exp(0.02) = 1.02 (95% CI 0.91 to 1.14); log RoM: 0.02 (‐0.09 to 0.13); I² = 0%

608
(4 studies)

⊕⊕⊕⊕
High

RoM (95% CrI): 0.96
(0.64 to 1.36),
SUCRA = 0.468, mean Pr(best) = 0.010, mean rank = 4.19

Atypical antipsychotic vs placebo

Median duration of delirium: 3 to 5 days for placebo

3.22 days of delirium (95% CrI 2.01 to
4.43) corresponding to 4 days in the placebo group

RoM: exp(‐0.31) = 0.73 (95% CI 0.49 to 1.11); log RoM: ‐0.31 (‐0.71 to 0.10); I² = 82%

500
(4 studies)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 0.80
(0.50 to 1.11),
SUCRA = 0.738, mean Pr(best) = 0.114, mean rank = 2.57

Statin (HMG‐CoA)

vs placebo

Mean duration of delirium: 6.8 to 8.68 days for placebo

4.20 days of delirium (95% CrI 2.44 to
7.09) corresponding to 4 days in the placebo group

RoM: exp(0.07) = 1.07 (95% CI 0.91 to 1.25); log RoM: 0.07 (‐0.09 to 0.22); I² = 0%

414
(2 studies)

⊕⊕⊕⊝
Moderateb

RoM (95% CrI): 1.05
(0.61 to 1.77),
SUCRA = 0.365, mean Pr(best) = 0.023, mean rank = 4.81

Alpha2 agonist

vs placebo

Median duration of delirium: 2.583 days for placebo

2.31 days of delirium (95% CrI 1.06 to
5.06) corresponding to 4 days in the placebo group

RoM: exp(‐0.55) = 0.58 (95% CI 0.43 to 0.79); log RoM: ‐0.55 (‐0.85 to ‐0.24); I² not applicable

71
(1 study)

⊕⊕⊕⊝
Moderateb

RoM (95% CrI): 0.58
(0.26 to 1.27),
SUCRA = 0.895, mean Pr(best) = 0.717, mean rank = 1.63

Cholinesterase inhibitor

vs placebo

Median duration of delirium: 3 days for placebo

7.37 days of delirium (95% CrI 3.26 to
16.38) corresponding to 4 days in the placebo group

RoM: exp(0.61) = 1.84 (95% CI 1.25 to 2.69); log RoM: 0.61 (0.22 to 0.99); I² not applicable

104
(1 study)

⊕⊕⊕⊝
Moderateb

RoM (95% CrI): 1.84
(0.82 to 4.10),
SUCRA = 0.054, mean Pr(best) = 0.006, mean rank = 6.68

Opioid

vs placebo

No study reported this comparison

3.53 days of delirium (95% CrI 1.46 to

8.05) corresponding to 4 days in the placebo group

Pairwise meta‐analysis not performed

0
(0 studies)

⊕⊕⊝⊝
Very lowb,c

RoM (95% CrI): 0.88
(0.37 to 2.01),
SUCRA = 0.578, mean Pr(best) = 0.129, mean rank = 3.53

*The basis for the assumed risk (e.g. the median control group risk across studies). The corresponding risk (and its 95% CrI) is calculated as the assumed risk multiplied by the ratio of means (and its 95% CrI) based on NMA.

Abbreviations: CI: confidence interval; CrI: credible interval; HMG‐CoA: 5‐hydroxy‐3‐methylglutaryl‐coenzyme A reductase inhibitor; NMA: network meta‐analysis; Pr(best): probability(best); RoM: ratio of means; SUCRA: surface under the cumulative ranking curve.

GRADE Working Group grades of evidence.
High quality: further research is very unlikely to change our confidence in the estimate of effect.
Moderate quality: further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate.
Low quality: further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate.
Very low quality: we are very uncertain about the estimate.

aDowngraded one level for heterogeneity (I² > 75% considered as large heterogeneity).
bDowngraded one level for imprecision (wide credible interval).
cDowngraded two levels for only indirect evidence available and risk of bias of a single trial informing opioid vs typical antipsychotic.

Figures and Tables -
Summary of findings for the main comparison. Duration of delirium
Summary of findings 2. Days with coma

Outcome: days with coma

Patient or population: critically ill adult with confirmed or at high risk of delirium
Settings: intensive care units in Australia and New Zealand, Canada, Egypt, Netherlands, Turkey, USA, UK
Intervention: any pharmacological intervention

Control: placebo or active comparator

Comparisons

Illustrative comparative risks* (95% CI)

Ratio of means (RoM) based on log RoM estimates from meta‐analysis

(IV, random, 95% CI)

Number of participants
(studies)

Quality of the evidence
(GRADE) based on NMA

NMA results
(assuming
consistency equations)

Assumed risk

Corresponding risk based on NMA estimates

Placebo/Comparator

Intervention drug

Typical antipsychotic vs placebo

Median number of days with coma: 1 to 2 days for placebo

1.53 days with coma (95% CrI 0.86 to
2.57) corresponding to 2 days in the placebo group

RoM: exp(‐0.29) = 0.75 (95% CI 0.49 to 1.13);
log RoM: ‐0.29 (‐0.71 to 0.12); I² = 74%

588
(3 studies)

⊕⊕⊝⊝
Lowa,b

RoM (95% CrI): 0.77
(0.43 to 1.29),
SUCRA = 0.820, mean Pr(best) = 0.620, mean rank = 1.54

Atypical antipsychotic

vs placebo

Median number of days with coma: 1 to 2 days for placebo

1.88 days with coma (95% CrI 0.96 to
3.43) corresponding to 2 days in the placebo group

RoM: exp(0.06) = 1.06 (95% CI 0.88 to 1.30);
log RoM: 0.06 (‐0.13 to 0.26); I² = 0%

440
(2 studies)

⊕⊕⊕⊝
Moderateb

RoM (95% CrI): 0.94
(0.48 to 1.72),
SUCRA = 0.422, mean Pr(best) = 0.132, mean rank = 2.73

Statin (HMG‐CoA)

vs placebo

Mean number of days with coma: 1.1 to 4.2 days for placebo

1.84 days with coma (95% CrI 0.98 to
3.59) corresponding to 2 days in the placebo group

RoM: exp(‐0.10) = 0.90 (95% CI 0.73 to 1.12);
log RoM: ‐0.10 (‐0.32 to 0.11); I² = 0%

414
(2 studies)

⊕⊕⊕⊝
Moderateb

RoM (95% CrI): 0.92
(0.49 to 1.80),
SUCRA = 0.481, mean Pr(best) = 0.222, mean rank = 2.56

*The basis for the assumed risk (e.g. the median control group risk across studies). The corresponding risk (and its 95% CrI) is calculated as the assumed risk multiples the ratio of means (and its 95% CrI) based on NMA.
CI: confidence interval; CrI: credible interval; HMG‐CoA: 5‐hydroxy‐3‐methylglutaryl‐coenzyme A reductase inhibitor; NMA: network meta‐analysis; Pr(best): probability(best); RoM: ratio of means; SUCRA: surface under the cumulative ranking curve.

GRADE Working Group grades of evidence.
High quality: further research is very unlikely to change our confidence in the estimate of effect.
Moderate quality: further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate.
Low quality: further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate.
Very low quality: we are very uncertain about the estimate.

aDowngraded one level for heterogeneity (I² of 50% to 75%, > 75% considered as medium and large heterogeneity).
bDowngraded one level for imprecision (wide credible interval).

Figures and Tables -
Summary of findings 2. Days with coma
Summary of findings 3. Duration of mechanical ventilation

Outcome: duration of mechanical ventilation

Patient or population: critically ill adult with confirmed or at high risk of delirium
Settings: intensive care units in Australia and New Zealand, Canada, Egypt, Netherlands, Turkey, USA, UK
Intervention: any pharmacological intervention

Control: placebo or active comparator

Comparisons

Illustrative comparative risks* (95% CI)

Ratio of means (RoM) based on log RoM estimates from meta‐analysis

(IV, random, 95% CI)

Number of participants
(studies)

Quality of the evidence
(GRADE) based on NMA

NMA results (assuming
consistency equations)

Assumed risk

Corresponding risk based on NMA estimates

Placebo/Comparator

Intervention drug

Typical antipsychotics
vs placebo

Median duration of mechanical ventilation: 3 to 5 days for placebo

3.71 days of mechanical ventilation (95% CrI 2.89 to 4.94) corresponding to 4 days in the placebo group

RoM: exp(‐0.08) = 0.92 (95% CI 0.79 to 1.06); log RoM: ‐0.08 (‐0.23 to 0.06); I² = 0%

515
(3 studies)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 0.93
(0.72 to 1.24),
SUCRA = 0.576, mean Pr(best) = 0.009, mean rank = 3.12

Atypical antipsychotics
vs placebo

Median duration of mechanical ventilation: 3 to 11 days for placebo

3.91 days of mechanical ventilation (95% CrI 2.85 to 5.10) corresponding to 4 days in the placebo group

RoM: exp(‐0.02) = 0.98 (95% CI 0.84 to 1.34); log RoM: ‐0.02 (‐0.17 to 0.14); I² = 0%

476
(3 studies)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 0.98
(0.71 to 1.28),
SUCRA = 0.440, mean Pr(best) = 0.012, mean rank = 3.80

Statin (HMG‐CoA)
vs placebo

Mean duration of mechanical ventilation: 11 days for placebo

4.38 days of mechanical ventilation (95% CrI 2.82 to 6.77) corresponding to 4 days in the placebo group

RoM: exp(0.09) = 1.09 (95% CI 0.90 to 1.34); log RoM: 0.09 (‐0.11 to 0.29); I² not applicable

272
(1 study)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 1.10
(0.71 to 1.69),
SUCRA = 0.223, mean Pr(best) = 0.014, mean rank = 4.88

Alpha2 agonist
vs placebo

Median duration of mechanical ventilation: 1.846 days for placebo

2.21 days of mechanical ventilation (95% CrI 1.36 to 3.58) corresponding to 4 days in the placebo group

RoM: exp(‐0.59) = 0.55 (95% CI 0.41 to 0.75); log RoM: ‐0.59 (‐0.89 to ‐0.29); I² not applicable

71
(1 study)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 0.55
(0.34 to 0.89),
SUCRA = 0.974, mean Pr(best) = 0.931, mean rank = 1.13

Opioid
vs placebo

No study reported this comparison

3.96 days of mechanical ventilation (95% CrI 2.32 to 7.02) corresponding to 4 days in the opioid group

Pairwise meta‐analysis not performed

0

(0 studies)

⊕⊕⊝⊝
Very lowa,b

RoM (95% CrI): 0.99
(0.58 to 1.76),
SUCRA = 0.410, mean Pr(best) = 0.033, mean rank = 3.95

*The basis for the assumed risk (e.g. the median control group risk across studies). The corresponding risk (and its 95% CrI) is calculated as the assumed risk multiples the ratio of means (and its 95% CrI) based on NMA.
CI: confidence interval; CrI: credible interval; HMG‐CoA: 5‐hydroxy‐3‐methylglutaryl‐coenzyme A reductase inhibitor; NMA: network meta‐analysis; Pr(best): probability(best); RoM: ratio of means; SUCRA: surface under the cumulative ranking curve.

GRADE Working Group grades of evidence.
High quality: further research is very unlikely to change our confidence in the estimate of effect.
Moderate quality: further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate.
Low quality: further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate.
Very low quality: we are very uncertain about the estimate.

aDowngraded one level for imprecision (wide credible interval).
bDowngraded two levels for only indirect evidence available and risk of bias of a single trial informing opioid vs typical antipsychotic.

Figures and Tables -
Summary of findings 3. Duration of mechanical ventilation
Summary of findings 4. Length of ICU stay

Outcome: length of ICU stay

Patient or population: critically ill adult with confirmed or at high risk of delirium
Settings: intensive care units in Australia and New Zealand, Canada, Egypt, Netherlands, Turkey, USA, UK
Intervention: any pharmacological intervention

Control: placebo or active comparator

Comparisons

Illustrative comparative risks* (95% CI)

Ratio of means (RoM) based on log RoM estimates from meta‐analysis

(IV, random, 95% CI)

Number of participants
(studies)

Quality of the evidence
(GRADE) based on NMA

NMA results (assuming
consistency equations)

Assumed risk

Corresponding risk based on NMA estimates

Placebo/Comparator

Intervention drug

Typical antipsychotic
vs placebo

Median length of ICU
stay: 5 to 9 days for placebo

7.92 days of ICU stay (95% CrI 6.79 to
9.37) corresponding to 8 days in the placebo group

RoM: exp(0.01) = 1.01 (95% CI 0.90 to 1.14); log RoM: 0.01 (‐0.11 to 0.13); I² = 0%

618
(4 studies)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 0.99
(0.85 to 1.17),
SUCRA = 0.496, mean Pr(best) = 0.014, mean rank = 4.02

Atypical antipsychotic
vs placebo

Median length of ICU
stay: 3 to 16 days for placebo

7.40 days of ICU stay (95% CrI 6.37 to
8.66) corresponding to 8 days in the placebo group

RoM: exp(‐0.09) = 0.91 (95% CI 0.84 to 1.00); log RoM: ‐0.09 (‐0.18 to ‐0.00); I² = 0%

577
(4 studies)

⊕⊕⊕⊕
High

RoM (95% CrI): 0.92
(0.80 to 1.08),
SUCRA = 0.709, mean Pr(best) = 0.106, mean rank = 2.75

Statin (HMG‐CoA)
vs placebo

Mean length of ICU
stay: 13 days for placebo

8.54 days of ICU stay (95% CrI 6.46 to
11.25) corresponding to 8 days in the placebo group

RoM: exp(0.06) = 1.06 (95% CI 0.91 to 1.23); log RoM: 0.06 (‐0.09 to 0.21); I² not applicable

272
(1 study)

⊕⊕⊝⊝
Lowa,b

RoM (95% CrI): 1.07
(0.81 to 1.41),
SUCRA = 0.344, mean Pr(best) = 0.030, mean rank = 4.93

Alpha2 agonist
vs placebo

Median length of ICU
stay: 7.5 days for placebo

6.43 days of ICU stay (95% CrI 4.42 to
9.33) corresponding to 8 days in the placebo group

RoM: exp(‐0.22) = 0.80 (95% CI 0.59 to 1.08); log RoM: ‐0.22 (‐0.53 to 0.08); I² not applicable

71
(1 study)

⊕⊕⊝⊝
Lowa,b

RoM (95% CrI): 0.80
(0.55 to 1.17),
SUCRA = 0.853, mean Pr(best) = 0.608, mean rank = 1.88

Cholinesterase inhibitor
vs placebo

Median length of ICU
stay: 8 days for placebo

17.53 days of ICU stay (95% CrI 11.76 to
26.14) corresponding to 8 days in the placebo group

RoM: exp(0.78) = 2.18 (95% CI 1.58 to 3.03); log RoM: 0.78 (0.46 to 1.11); I² not applicable

104
(1 study)

⊕⊕⊕⊝
Moderatea

RoM (95% CrI): 2.19
(1.47 to 3.27),
SUCRA = 0.002, mean Pr(best) = 0, mean rank = 6.99

Opioid
vs placebo

No study reported this comparison

7.40 days of ICU stay (95% CrI 4.95 to
11.24) corresponding to 8 days in the opioid group

Pairwise meta‐analysis not performed

0

(0 studies)

⊕⊕⊝⊝
Very lowa,c

RoM (95% CrI): 0.92
(0.62 to 1.40),
SUCRA = 0.639, mean Pr(best) = 0.238, mean rank = 3.17

*The basis for the assumed risk (e.g. the median control group risk across studies). The corresponding risk (and its 95% CrI) is calculated as the assumed risk multiples the ratio of means (and its 95% CrI) based on NMA.
CI: confidence interval; CrI: credible interval; HMG‐CoA: 5‐hydroxy‐3‐methylglutaryl‐coenzyme A reductase inhibitor; NMA: network meta‐analysis; Pr(best): probability(best); RoM: ratio of means; SUCRA: surface under the cumulative ranking curve.

GRADE Working Group grades of evidence.
High quality: further research is very unlikely to change our confidence in the estimate of effect.
Moderate quality: further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate.
Low quality: further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate.
Very low quality: we are very uncertain about the estimate.

aDowngraded one level for imprecision (wide credible interval).
bDowngraded one level for single trial with risk of bias and indirectness.

cDowngraded two levels for only indirect evidence available and risk of bias of a single trial informing opioid vs typical antipsychotic.

Figures and Tables -
Summary of findings 4. Length of ICU stay
Summary of findings 5. Length of hospital stay

Outcome: length of hospital stay

Patient or population: critically ill adult with confirmed or at high risk of delirium
Settings: intensive care units in Australia and New Zealand, Canada, Egypt, Netherlands, Turkey, USA, UK
Intervention: any pharmacological intervention

Control: placebo or active comparator

Outcomes

Illustrative comparative risks* (95% CI)

Ratio of means (RoM) based on log RoM estimates from meta‐analysis

(IV, random, 95% CI)

Number of participants
(studies)

Quality of the evidence
(GRADE) based on NMA

NMA results (assuming
consistency equations)

Assumed risk

Corresponding risk based on NMA estimates

Placebo/Comparator

Intervention drug

Typical AP
vs placebo

Median length of hospital stay: 13 to 26 days for placebo

16.48 days of hospital stay (95% CrI 11.74 to 21.29) corresponding to 18 days in the placebo group

RoM: exp(‐0.12) = 0.89 (95% CI 0.68 to 1.15); log RoM: ‐0.12 (‐0.38 to 0.14); I² = 72%

479
(2 studies)

⊕⊕⊝⊝
Lowa,b

RoM (95% CrI): 0.92
(0.65 to 1.18),
SUCRA = 0.722, mean Pr(best) = 0.235, mean rank = 2.67

Atypical AP
vs placebo

Median length of hospital stay: 6 to 26 days for placebo

16.69 days of hospital stay (95% CrI 12.47 to 20.79) corresponding to 18 days in the placebo group

RoM: exp(‐0.04) = 0.96 (95% CI 0.88 to 1.05); log RoM: ‐0.04 (‐0.13 to 0.05); I² = 0%

511
(3 studies)

⊕⊕⊕⊝

Moderateb

RoM (95% CrI): 0.93
(0.69 to 1.16),
SUCRA = 0.693, mean Pr(best) = 0.218, mean rank = 2.84

Statin (HMG‐CoA)
vs placebo

Mean length of hospital stay: 22 to 23.1 days for placebo

17.55 days of hospital stay (95% CrI 12.45 to 23.47) corresponding to 18 days in the placebo group

RoM: exp(‐0.01) = 0.99 (95% CI 0.88 to 1.13); log RoM: ‐0.01 (‐0.13 to 0.12); I² = 0%

369

(2 studies)

⊕⊕⊕⊝

Moderateb

RoM (95% CrI): 0.98
(0.69 to 1.30),
SUCRA = 0.537, mean Pr(best) = 0.147, mean rank = 3.78

Alpha2 agonist
vs placebo

Median length of hospital stay: 12.5 days for placebo

19.80 days of hospital stay (95% CrI 12.37 to 31.52) corresponding to 18 days in the placebo group

RoM: exp(0.09) = 1.09 (95% CI 0.84 to 1.42); log RoM: 0.09 (‐0.17 to 0.35); I² not applicable

71

(1 study)

⊕⊕⊕⊝

Moderatec

RoM (95% CrI): 1.10
(0.69 to 1.75),
SUCRA = 0.301, mean Pr(best) = 0.090, mean rank = 5.19

Cholinesterase Inhibitor
vs placebo

Median length of hospital stay: 25 days for placebo

20.00 days of hospital stay (95% CrI 12.64 to 31.93) corresponding to 18 days in the placebo group

RoM: exp(0.11) = 1.12 (95% CI 0.86 to 1.43); log RoM: 0.11 (‐0.15 to 0.36); I² not applicable

104

(1 study)

⊕⊕⊕⊝

Moderatec

RoM (95% CrI): 1.11
(0.70 to 1.77),
SUCRA = 0.280, mean Pr(best) = 0.078, mean rank = 5.32

Opioid
vs placebo

No study reported this comparison

17.51 days of hospital stay (95% CrI 9.89 to 28.78) corresponding to 18 days in the opioid group

Pairwise meta‐analysis not performed

0

(0 studies)

⊕⊕⊝⊝
Very lowb,d

RoM (95% CrI): 0.97
(0.55 to 1.60),
SUCRA = 0.532, mean Pr(best) = 0.225, mean rank = 3.81

*The basis for the assumed risk (e.g. the median control group risk across studies). The corresponding risk (and its 95% CrI) is calculated as the assumed risk multiples the ratio of means (and its 95% CrI) based on NMA.
CI: confidence interval; CrI: credible interval; HMG‐CoA: 5‐hydroxy‐3‐methylglutaryl‐coenzyme A reductase inhibitor; NMA: network meta‐analysis; Pr(best): probability(best); RoM: ratio of means; SUCRA: surface under the cumulative ranking curve.

GRADE Working Group grades of evidence.
High quality: further research is very unlikely to change our confidence in the estimate of effect.
Moderate quality: further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate.
Low quality: further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate.
Very low quality: we are very uncertain about the estimate.

aDowngraded one level for heterogeneity (I² of 50% to 75%, > 75% considered as medium and large heterogeneity).
bDowngraded one level for imprecision (wide credible interval).
cDowngraded one level for single small trial with risk of bias and indirectness.
dDowngraded two levels for only indirect evidence available and risk of bias of a single trial informing opioid vs typical antipsychotic.

Figures and Tables -
Summary of findings 5. Length of hospital stay
Summary of findings 6. QTc prolongation

Outcome: QTc prolongation

Patient or population: critically ill adult with confirmed or at high risk of delirium
Settings: intensive care units in Australia and New Zealand, Canada, Egypt, Netherlands, Turkey, USA, UK
Intervention: any pharmacological intervention

Control: placebo or active comparator

Comparisons

Illustrative comparative risks* (95% CI)

Relative effect

OR (95% CI)

Absolute effect

(auto calculation using
GRADEpro GDT)

Number of
participants
(studies)

Quality of the evidence
(GRADE)

Assumed risk

Corresponding risk

Placebo/Comparator

Intervention drug

Typical antipsychotic vs placebo

62 per 1000

78 per 1000

1.26 (0.68 to 2.34)
I² = 0%

15 more per 1000
(from 19 fewer to
72 more)

656
(4 studies)

⊕⊕⊕⊕
High

Atypical antipsychotic vs placebo

90 per 1000

118 per 1000

1.28 (0.45 to 3.66)
I² = 56%

22 more per 1000
(from 48 fewer to
176 more)

577
(4 studies)

⊕⊕⊕⊝
Moderatea

Typical antipsychotic
vs atypical antipsychotic

114 per 1000

66 per 1000

0.55 (0.28 to 1.08)
I² = 0%

48 fewer per 1000
(from 79 fewer to
8 more)

447
(2 studies)

⊕⊕⊕⊕
High

Alpha2 agonist vs
typical antipsychotic

400 per 1000

400 per 1000

1.00 (0.17 to 5.98)
I² not applicable

0 fewer per 1000
(from 298 fewer to
399 more)

20
(1 study)

⊕⊕⊝⊝
Lowb

Alpha2 agonist
vs 5HT3 inhibitor

0 per 1000

0 per 1000

OR not estimable
I² not applicable

Not estimable

64
(1 study)

⊕⊕⊝⊝
Lowb

*The basis for the assumed risk (e.g. the median control group risk across studies). The corresponding risk (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI).

CI: confidence interval; OR: odds ratio.

GRADE Working Group grades of evidence.
High quality: further research is very unlikely to change our confidence in the estimate of effect.
Moderate quality: further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate.
Low quality: further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate.
Very low quality: we are very uncertain about the estimate.

aDowngraded one level for heterogeneity (I² of 50% to 75%, > 75% considered as medium and large heterogeneity).

bDowngraded two levels for imprecision (wide confidence interval, single small trial with risk of bias).

Figures and Tables -
Summary of findings 6. QTc prolongation
Table 1. Duration of delirium: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)

Alpha2
agonist

0.807

0.819

0.907

0.922

0.977

0.938

0.72
(0.33, 1.87)

Atypical

antipsychotic

0.610

0.858

0.861

0.975

0.927

0.66
(0.21, 2.13)

0.92
(0.36, 2.11)

Opioid

0.627

0.685

0.924

0.660

0.60
(0.26, 1.48)

0.84
(0.50, 1.25)

0.92
(0.42, 1.97)

Typical

antipsychotic

0.646

0.945

0.607

0.55
(0.22, 1.43)

0.77
(0.37, 1.38)

0.84
(0.30, 2.25)

0.92
(0.47, 1.73)

Statin

0.909

0.396

0.31
(0.10, 0.97)

0.44
(0.17, 1.00)

0.48
(0.14, 1.51)

0.52
(0.21, 1.25)

0.57
(0.22, 1.49)

CHE

Inhibitor

0.054

0.58
(0.26, 1.27)

0.80
(0.50, 1.11)

0.88
(0.37, 2.01)

0.96
(0.64, 1.36)

1.05
(0.61, 1.77)

1.84
(0.82, 4.10)

Placebo

CHE: cholinesterase.

Crl: credible interval.

RoM: ratio of means.

A complete summary of estimates for efficacy from the random‐effects (RE) consistency model assuming vague priors is displayed.

Treatments other than placebo are in the order of decreasing surface under the cumulative ranking curve (SUCRA) value from upper left to lower right. For each comparison, the lower/right‐most treatment is the reference treatment. For example, the RoM estimate of 0.58 (95% CrI 0.26 to 1.27) in the lower triangle suggests a 42% reduction in the mean duration of delirium with alpha2 agonists compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 93.8% that alpha2 agonists are better than placebo in terms of duration of delirium. Estimates which ruled out the possibility of no difference based on pairwise RoM estimates are shown in bold font.

Figures and Tables -
Table 1. Duration of delirium: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)
Table 2. Duration of delirium: mean SUCRA value, mean probability to be the best, and mean rank for each treatment

RE consistency model

Mean SUCRA

Mean Pr(best)

Mean ranka

Alpha2 agonist

0.895

0.717

1.63 (1 to 6)

Atypical antipsychotic

0.738

0.114

2.57 (1 to 5)

Opioid

0.578

0.129

3.53 (1 to 7)

Typical antipsychotic

0.468

0.010

4.19 (2 to 6)

Placebo

0.403

0.001

4.58 (3 to 6)

Statin

0.365

0.023

4.81 (2 to 7)

CHE inhibitor

0.054

0.006

6.68 (3 to 7)

aMean rank with 2.5% and 97.5% quantiles in parentheses.

CHE: cholinesterase.

Pr: probability.

RE: random‐effects.

SUCRA: surface under the cumulative ranking curve.

Figures and Tables -
Table 2. Duration of delirium: mean SUCRA value, mean probability to be the best, and mean rank for each treatment
Table 3. Delirium‐ and coma‐free days: league table of posterior median pairwise RoM and 95% (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)

Atypical
antipsychotic

0.784

0.854

0.898

1.15
(0.58, 2.40)

Typical

antipsychotic

0.768

0.781

1.46
(0.45, 5.41)

1.27
(0.40, 4.30)

Statin

0.340

1.31
(0.69, 2.83)

1.14
(0.64, 2.16)

0.90
(0.32, 2.52)

Placebo

RoM: ratio of means.

A complete summary of estimates for efficacy from the random‐effects (RE) consistency model assuming vague priors is displayed.

Treatments other than placebo are in the order of decreasing surface under the cumulative ranking curve (SUCRA) value from upper left to lower right. For each comparison, the lower/right‐most treatment is the reference treatment. For example, the RoM estimate of 1.31 (95% credible interval (CrI) 0.69 to 2.83) in the lower triangle suggests a 31% increase in mean delirium‐ and coma‐free days with atypical antipsychotics compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 89.8% that atypical antipsychotics are better than placebo in terms of delirium‐ and coma‐free days.

Figures and Tables -
Table 3. Delirium‐ and coma‐free days: league table of posterior median pairwise RoM and 95% (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)
Table 4. Delirium‐ and coma‐free days: mean SUCRA value, mean probability to be the best, and mean rank for each treatment

RE consistency model

Mean SUCRA

Mean Pr(best)

Mean ranka

Atypical antipsychotic

0.845

0.690

1.46 (1 to 4)

Typical antipsychotic

0.589

0.160

2.23 (1 to 4)

Placebo

0.327

0.033

3.02 (1 to 4)

Statin

0.239

0.116

3.28 (1 to 4)

aMean rank with 2.5% and 97.5% quantiles in parentheses.

Pr: probability.

RE: random‐effects.

SUCRA: surface under the cumulative ranking curve.

Figures and Tables -
Table 4. Delirium‐ and coma‐free days: mean SUCRA value, mean probability to be the best, and mean rank for each treatment
Table 5. Days with coma: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)

Typical

antipsychotic

0.740

0.815

0.905

0.83
(0.34, 1.87)

Statin

0.532

0.651

0.81
(0.44, 1.56)

0.98
(0.41, 2.58)

Atypical

antipsychotic

0.612

0.77
(0.43, 1.29)

0.92
(0.49, 1.80)

0.94
(0.48, 1.72)

Placebo

Crl: credible interval.

RoM: ratio of means.

A complete summary of estimates for efficacy from the random‐effects (RE) consistency model assuming vague priors is displayed.

Treatments other than placebo are in the order of decreasing surface under the cumulative ranking curve (SUCRA) value from upper left to lower right. For each comparison, the lower/right‐most treatment is the reference treatment. For example, the RoM estimate of 0.77 (95% CrI 0.43 to 1.29) in the lower triangle suggests a 23% reduction in mean coma days with typical antipsychotics compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 90.5% that typical antipsychotics are better than placebo in terms of days with coma.

Figures and Tables -
Table 5. Days with coma: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)
Table 6. Days with coma: mean SUCRA values, mean probability to be the best, and mean rank for each treatment

RE consistency model

Mean SUCRA

Mean Pr(best)

Mean ranka

Typical antipsychotic

0.820

0.620

1.54 (1 to 4)

Statin

0.481

0.222

2.56 (1 to 4)

Atypical antipsychotic

0.422

0.132

2.73 (1 to 4)

Placebo

0.278

0.026

3.17 (1 to 4)

aMean rank with 2.5% and 97.5% quantiles in parentheses.

Pr: probability.

RE: random‐effects.

SUCRA: surface under the cumulative ranking curve

Figures and Tables -
Table 6. Days with coma: mean SUCRA values, mean probability to be the best, and mean rank for each treatment
Table 7. Duration of mechanical ventilation: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)

Alpha2
agonist

0.973

0.973

0.958

0.978

0.986

0.59
(0.34, 1.01)

Typical

antipsychotic

0.665

0.628

0.805

0.754

0.57
(0.33, 1.02)

0.95
(0.72, 1.35)

Atypical

antipsychotic

0.527

0.729

0.582

0.56
(0.26, 1.14)

0.94
(0.58, 1.52)

0.98
(0.53, 1.70)

Opioid

0.645

0.517

0.50
(0.26, 0.97)

0.85
(0.52, 1.45)

0.89
(0.52, 1.47)

0.90
(0.46, 1.87)

Statin

0.274

0.55
(0.34, 0.89)

0.93
(0.72, 1.24)

0.98
(0.71, 1.28)

0.99
(0.58, 1.76)

1.10
(0.71, 1.69)

Placebo

Crl: credible interval.

RoM: ratio of means.

A complete summary of estimates for efficacy from the random‐effects (RE) consistency model assuming vague priors is displayed.

Treatments other than placebo are in the order of decreasing surface under the cumulative ranking curve (SUCRA) value from upper left to lower right. For each comparison, the lower/right‐most treatment is the reference treatment. For example, the RoM estimate of 0.55 (95% CrI 0.34 to 0.89) in the lower triangle suggests a 45% reduction in the mean duration of mechanical ventilation with alpha2 agonists compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 98.6% that alpha2 agonists are better than placebo for the duration of mechanical ventilation. Estimates which ruled out the possibility of no difference based on pairwise RoM estimates are shown in bold font.

Figures and Tables -
Table 7. Duration of mechanical ventilation: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)
Table 8. Duration of mechanical ventilation: mean SUCRA values, mean probability to be the best, and mean rank for each treatment

RE consistency model

Mean SUCRA

Mean Pr(best)

Mean ranka

Alpha2 agonists

0.974

0.931

1.13 (1 to 3)

Typical antipsychotic

0.576

0.009

3.12 (2 to 6)

Atypical antipsychotic

0.440

0.012

3.80 (2 to 6)

Opioid

0.410

0.033

3.95 (1 to 6)

Placebo

0.377

0.001

4.11 (2 to 6)

Statin

0.223

0.014

4.88 (2 to 6)

aMean rank with 2.5% and 97.5% quantiles in parentheses.

Pr: probability.

RE: random‐effects.

SUCRA: surface under the cumulative ranking curve.

Figures and Tables -
Table 8. Duration of mechanical ventilation: mean SUCRA values, mean probability to be the best, and mean rank for each treatment
Table 9. Length of ICU stay: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)

Alpha2 agonist

0.766

0.705

0.858

0.902

0.999

0.886

0.87
(0.58, 1.29)

Atypical

antipsychotic

0.499

0.792

0.856

0.999

0.874

0.87
(0.50, 1.49)

1.00
(0.65, 1.51)

Opioid

0.647

0.736

0.996

0.658

0.81
(0.54, 1.21)

0.93
(0.77, 1.12)

0.93
(0.64, 1.36)

Typical

antipsychotic

0.714

0.998

0.559

0.75
(0.47, 1.20)

0.87
(0.64, 1.20)

0.87
(0.53, 1.44)

0.93
(0.68, 1.29)

Statin

0.995

0.278

0.37
(0.21, 0.63)

0.42
(0.28, 0.65)

0.42
(0.24, 0.75)

0.45
(0.30, 0.70)

0.49
(0.30, 0.79)

CHE inhibitor

0.001

0.80
(0.55, 1.17)

0.92
(0.80, 1.08)

0.92
(0.62, 1.40)

0.99
(0.85, 1.17)

1.07
(0.81, 1.41)

2.19
(1.47, 3.27)

Placebo

CHE: cholinesterase.

Crl: credible interval.

ICU: intensive care unit.

RoM: ratio of means.

A complete summary of estimates for efficacy from the random‐effects (RE) consistency model assuming vague priors is displayed.

Treatments other than placebo are in the order of decreasing surface under the cumulative ranking curve (SUCRA) value from upper left to lower right. For each comparison, the lower/right‐most treatment is the reference treatment. For example, the RoM estimate of 0.80 (95% CrI 0.55 to 1.17) in the lower triangle suggests a 20% reduction in mean length of ICU stay with alpha2 agonists compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 88.6% that alpha2 agonists are better than placebo for the length of ICU stay. Estimates which ruled out the possibility of no difference based on pairwise RoM estimates are shown in bold font.

Figures and Tables -
Table 9. Length of ICU stay: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)
Table 10. Length of ICU stay: mean SUCRA value, mean probability to be the best, and mean rank for each treatment

RE consistency model

Mean SUCRA

Mean Pr(best)

Mean ranka

Alpha2 agonists

0.853

0.608

1.88 (1 to 6)

Atypical antipsychotic

0.709

0.106

2.75 (1 to 5)

Opioid

0.639

0.238

3.17 (1 to 6)

Typical antipsychotic

0.496

0.014

4.02 (2 to 6)

Placebo

0.457

0.004

4.26 (2 to 6)

Statin

0.344

0.030

4.93 (1 to 6)

CHE inhibitor

0.002

0.000

6.99 (7 to 7)

aMean rank with 2.5% and 97.5% quantiles in parentheses.

CHE: cholinesterase.

ICU: intensive care unit.

Pr: probability.

RE: random‐effects.

SUCRA: surface under the cumulative ranking curve.

Figures and Tables -
Table 10. Length of ICU stay: mean SUCRA value, mean probability to be the best, and mean rank for each treatment
Table 11. Length of hospital stay: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)

Typical
antipsychotic

0.546

0.673

0.654

0.810

0.828

0.820

0.99
(0.70, 1.38)

Atypical

antipsychotic

0.651

0.604

0.801

0.818

0.827

0.94
(0.60, 1.43)

0.95
(0.63, 1.42)

Statin

0.498

0.716

0.736

0.596

0.94
(0.60, 1.47)

0.95
(0.55, 1.67)

1.00
(0.55, 1.88)

Opioid

0.685

0.700

0.561

0.83
(0.46, 1.39)

0.84
(0.48, 1.38)

0.88
(0.49, 1.52)

0.88
(0.42, 1.73)

Alpha2 agonist

0.515

0.305

0.82
(0.46, 1.37)

0.83
(0.47, 1.36)

0.88
(0.49, 1.49)

0.87
(0.41, 1.70)

0.99
(0.51, 1.90)

CHE inhibitor

0.278

0.92
(0.65, 1.18)

0.93
(0.69, 1.16)

0.98
(0.69, 1.30)

0.97
(0.55, 1.60)

1.10
(0.69, 1.75)

1.11
(0.70, 1.77)

Placebo

CHE: cholinesterase.

Crl: credible interval.

RoM: ratio of means.

Treatments other than placebo are in the order of decreasing surface under the cumulative ranking curve (SUCRA) value from upper left to lower right. For each comparison, the lower/right‐most treatment is the reference treatment. For example, the RoM estimate of 0.92 (95% CrI 0.65 to 1.18) in the lower triangle suggests an 8% reduction in mean length of hospital stay with typical antipsychotics compared to placebo. The corresponding probability estimate in the upper triangle suggests a probability of 82% that typical antipsychotics are better than placebo for length of hospital stay.

Figures and Tables -
Table 11. Length of hospital stay: league table of posterior median pairwise RoM and 95% CrI (lower triangle), and pairwise probabilities that a treatment is better than another (upper triangle)
Table 12. Length of hospital stay: mean SUCRA value, mean probability to be the best, and mean rank for each treatment

RE consistency model

Mean SUCRA

Mean Pr(best)

Mean ranka

Typical antipsychotic

0.722

0.235

2.67 (1 to 6)

Atypical antipsychotic

0.693

0.218

2.84 (1 to 6)

Statin

0.537

0.147

3.78 (1 to 7)

Opioid

0.532

0.225

3.81 (1 to 7)

Placebo

0.435

0.008

4.39 (2 to 6)

Alpha2 agonists

0.301

0.090

5.19 (1 to 7)

CHE inhibitor

0.280

0.078

5.32 (1 to 7)

aMean rank with 2.5% and 97.5% quantiles in parentheses.

CHE: cholinesterase.

Pr: probability.

RE: random‐effects.

SUCRA: surface under the cumulative ranking curve.

Figures and Tables -
Table 12. Length of hospital stay: mean SUCRA value, mean probability to be the best, and mean rank for each treatment
Comparison 1. Duration of delirium (log ratio of means scale)

Outcome or subgroup title

No. of studies

No. of participants

Statistical method

Effect size

1 Duration of delirium (log units) Show forest plot

11

Mean Difference (IV, Random, 95% CI)

Subtotals only

1.1 Typical AP vs placebo

4

608

Mean Difference (IV, Random, 95% CI)

0.02 [‐0.09, 0.13]

1.2 Atypical AP vs placebo

4

500

Mean Difference (IV, Random, 95% CI)

‐0.31 [‐0.71, 0.10]

1.3 Statin vs placebo

2

414

Mean Difference (IV, Random, 95% CI)

0.07 [‐0.09, 0.22]

1.4 Alpha‐2 agonist vs placebo

1

71

Mean Difference (IV, Random, 95% CI)

‐0.55 [‐0.85, ‐0.24]

1.5 Cholinesterase inhibitor vs placebo

1

104

Mean Difference (IV, Random, 95% CI)

0.61 [0.22, 0.99]

1.6 Typical AP vs atypical AP

2

447

Mean Difference (IV, Random, 95% CI)

0.05 [‐0.07, 0.17]

1.7 Typical AP vs opioid

1

53

Mean Difference (IV, Random, 95% CI)

0.09 [‐0.17, 0.34]

Figures and Tables -
Comparison 1. Duration of delirium (log ratio of means scale)
Comparison 2. Delirium‐free and coma‐free days (log ratio of means scale)

Outcome or subgroup title

No. of studies

No. of participants

Statistical method

Effect size

1 Delirium‐free and coma‐free days Show forest plot

4

Mean Difference (IV, Random, 95% CI)

Subtotals only

1.1 Typical AP vs placebo

3

588

Mean Difference (IV, Random, 95% CI)

0.13 [‐0.08, 0.34]

1.2 Atypical AP vs placebo

2

440

Mean Difference (IV, Random, 95% CI)

0.36 [‐0.15, 0.87]

1.3 Statin vs placebo

1

142

Mean Difference (IV, Random, 95% CI)

‐0.11 [‐0.33, 0.11]

1.4 Typical AP vs atypical AP

2

447

Mean Difference (IV, Random, 95% CI)

‐0.08 [‐0.21, 0.05]

Figures and Tables -
Comparison 2. Delirium‐free and coma‐free days (log ratio of means scale)
Comparison 3. Days with coma (log ratio of means scale)

Outcome or subgroup title

No. of studies

No. of participants

Statistical method

Effect size

1 Days with coma (log units) Show forest plot

5

Mean Difference (IV, Random, 95% CI)

Subtotals only

1.1 Typical AP vs placebo

3

588

Mean Difference (IV, Random, 95% CI)

‐0.29 [‐0.71, 0.12]

1.2 Atypical AP vs placebo

2

440

Mean Difference (IV, Random, 95% CI)

0.06 [‐0.13, 0.26]

1.3 Statin vs placebo

2

414

Mean Difference (IV, Random, 95% CI)

‐0.10 [‐0.32, 0.11]

1.4 Typical AP vs atypical AP

2

447

Mean Difference (IV, Random, 95% CI)

‐0.15 [‐0.34, 0.04]

Figures and Tables -
Comparison 3. Days with coma (log ratio of means scale)
Comparison 4. Duration of mechanical ventilation (log ratio of means scale)

Outcome or subgroup title

No. of studies

No. of participants

Statistical method

Effect size

1 Duration of mechanical ventilation (log units) Show forest plot

7

Mean Difference (IV, Random, 95% CI)

Subtotals only

1.1 Typical AP vs placebo

3

515

Mean Difference (IV, Random, 95% CI)

‐0.08 [‐0.23, 0.06]

1.2 Atypical AP vs placebo

3

476

Mean Difference (IV, Random, 95% CI)

‐0.02 [‐0.17, 0.14]

1.3 Statin vs placebo

1

272

Mean Difference (IV, Random, 95% CI)

0.09 [‐0.11, 0.29]

1.4 Alpha‐2 agonist vs placebo

1

71

Mean Difference (IV, Random, 95% CI)

‐0.59 [‐0.89, ‐0.29]

1.5 Typical AP vs atypical AP

2

447

Mean Difference (IV, Random, 95% CI)

‐0.17 [‐0.67, 0.33]

1.6 Typical AP vs opioid

1

53

Mean Difference (IV, Random, 95% CI)

‐0.06 [‐0.37, 0.24]

Figures and Tables -
Comparison 4. Duration of mechanical ventilation (log ratio of means scale)
Comparison 5. Length of ICU stay (log ratio of means scale)

Outcome or subgroup title

No. of studies

No. of participants

Statistical method

Effect size

1 Length of ICU stay (log units) Show forest plot

10

Mean Difference (IV, Random, 95% CI)

Subtotals only

1.1 Typical AP vs placebo

4

618

Mean Difference (IV, Random, 95% CI)

0.01 [‐0.11, 0.13]

1.2 Atypical AP vs placebo

4

577

Mean Difference (IV, Random, 95% CI)

‐0.09 [‐0.18, ‐0.00]

1.3 Statin vs placebo

1

272

Mean Difference (IV, Random, 95% CI)

0.06 [‐0.09, 0.21]

1.4 Alpha‐2 agonist vs placebo

1

71

Mean Difference (IV, Random, 95% CI)

‐0.22 [‐0.53, 0.08]

1.5 Cholinesterase inhibitor vs placebo

1

104

Mean Difference (IV, Random, 95% CI)

0.78 [0.46, 1.11]

1.6 Typical AP vs atypical AP

2

447

Mean Difference (IV, Random, 95% CI)

0.01 [‐0.13, 0.16]

1.7 Typical AP vs opioid

1

53

Mean Difference (IV, Random, 95% CI)

0.07 [‐0.24, 0.37]

Figures and Tables -
Comparison 5. Length of ICU stay (log ratio of means scale)
Comparison 6. Length of hospital stay (log ratio of means scale)

Outcome or subgroup title

No. of studies

No. of participants

Statistical method

Effect size

1 Length of hospital stay (log units) Show forest plot

9

Mean Difference (IV, Random, 95% CI)

Subtotals only

1.1 Typical AP vs placebo

2

479

Mean Difference (IV, Random, 95% CI)

‐0.12 [‐0.38, 0.14]

1.2 Atypical AP vs placebo

3

511

Mean Difference (IV, Random, 95% CI)

‐0.04 [‐0.13, 0.05]

1.3 Statin vs placebo

2

369

Mean Difference (IV, Random, 95% CI)

‐0.01 [‐0.13, 0.12]

1.4 Alpha‐2 agonist vs placebo

1

71

Mean Difference (IV, Random, 95% CI)

0.09 [‐0.17, 0.35]

1.5 Cholinesterase Inhibitor vs placebo

1

104

Mean Difference (IV, Random, 95% CI)

0.11 [‐0.15, 0.36]

1.6 Typical AP vs atypical AP

1

382

Mean Difference (IV, Random, 95% CI)

0.04 [‐0.09, 0.17]

1.7 Typical AP vs opioid

1

53

Mean Difference (IV, Random, 95% CI)

‐0.06 [‐0.26, 0.13]

Figures and Tables -
Comparison 6. Length of hospital stay (log ratio of means scale)
Comparison 7. Mortality

Outcome or subgroup title

No. of studies

No. of participants

Statistical method

Effect size

1 Mortality Show forest plot

10

Odds Ratio (M‐H, Random, 95% CI)

Subtotals only

1.1 Typical AP vs placebo

4

656

Odds Ratio (M‐H, Random, 95% CI)

0.98 [0.69, 1.40]

1.2 Atypical AP vs placebo

4

577

Odds Ratio (M‐H, Random, 95% CI)

1.00 [0.66, 1.52]

1.3 Cholinesterase inhibitor vs placebo

1

104

Odds Ratio (M‐H, Random, 95% CI)

1.77 [0.74, 4.26]

1.4 Statin vs placebo

2

414

Odds Ratio (M‐H, Random, 95% CI)

1.42 [0.84, 2.39]

1.5 Typical AP vs opioid

1

53

Odds Ratio (M‐H, Random, 95% CI)

2.17 [0.18, 25.46]

1.6 Typical AP vs atypical AP

2

447

Odds Ratio (M‐H, Random, 95% CI)

0.90 [0.59, 1.39]

Figures and Tables -
Comparison 7. Mortality
Comparison 8. Akathisia

Outcome or subgroup title

No. of studies

No. of participants

Statistical method

Effect size

1 Akathisia Show forest plot

2

Odds Ratio (M‐H, Random, 95% CI)

Subtotals only

1.1 Typical AP vs placebo

2

212

Odds Ratio (M‐H, Random, 95% CI)

1.34 [0.49, 3.67]

1.2 Atypical AP vs placebo

1

66

Odds Ratio (M‐H, Random, 95% CI)

1.04 [0.31, 3.50]

1.3 Typical AP vs atypical AP

1

65

Odds Ratio (M‐H, Random, 95% CI)

1.6 [0.50, 5.09]

Figures and Tables -
Comparison 8. Akathisia
Comparison 9. Arrhythmias

Outcome or subgroup title

No. of studies

No. of participants

Statistical method

Effect size

1 Arrhythmias Show forest plot

4

Odds Ratio (M‐H, Random, 95% CI)

Subtotals only

1.1 Typical AP vs placebo

3

588

Odds Ratio (M‐H, Random, 95% CI)

3.09 [1.11, 8.62]

1.2 Atypical AP vs placebo

2

440

Odds Ratio (M‐H, Random, 95% CI)

0.0 [0.0, 0.0]

1.3 Alpha‐2 agonist vs typical AP

1

20

Odds Ratio (M‐H, Random, 95% CI)

1.0 [0.11, 8.95]

1.4 Typical AP vs atypical AP

2

447

Odds Ratio (M‐H, Random, 95% CI)

5.0 [0.24, 104.84]

Figures and Tables -
Comparison 9. Arrhythmias
Comparison 10. QTc prolongation

Outcome or subgroup title

No. of studies

No. of participants

Statistical method

Effect size

1 QTc prolongation Show forest plot

8

Odds Ratio (M‐H, Random, 95% CI)

Subtotals only

1.1 Typical AP vs placebo

4

656

Odds Ratio (M‐H, Random, 95% CI)

1.26 [0.68, 2.34]

1.2 Atypical AP vs placebo

4

577

Odds Ratio (M‐H, Random, 95% CI)

1.28 [0.45, 3.66]

1.3 Typical AP vs atypical AP

2

447

Odds Ratio (M‐H, Random, 95% CI)

0.55 [0.28, 1.08]

1.4 Alpha‐2 agonist vs typical AP

1

20

Odds Ratio (M‐H, Random, 95% CI)

1.0 [0.17, 5.98]

1.5 Alpha‐2 agonist vs 5HT3 inhibitor

1

64

Odds Ratio (M‐H, Random, 95% CI)

0.0 [0.0, 0.0]

Figures and Tables -
Comparison 10. QTc prolongation
Comparison 11. Extrapyramidal symptoms

Outcome or subgroup title

No. of studies

No. of participants

Statistical method

Effect size

1 Extrapyramidal symptoms Show forest plot

6

Odds Ratio (M‐H, Random, 95% CI)

Subtotals only

1.1 Typical AP vs placebo

4

656

Odds Ratio (M‐H, Random, 95% CI)

0.75 [0.26, 2.21]

1.2 Atypical AP vs placebo

3

476

Odds Ratio (M‐H, Random, 95% CI)

0.47 [0.11, 1.97]

1.3 Typical AP vs atypical AP

3

520

Odds Ratio (M‐H, Random, 95% CI)

2.22 [0.59, 8.38]

Figures and Tables -
Comparison 11. Extrapyramidal symptoms