Elsevier

Resuscitation

Volume 156, November 2020, Pages 196-201
Resuscitation

Simulation and education
The use of drones and a machine-learning model for recognition of simulated drowning victims—A feasibility study

https://doi.org/10.1016/j.resuscitation.2020.09.022Get rights and content

Abstract

Background

Submersion time is a strong predictor for death in drowning, already 10 min after submersion, survival is poor. Traditional search efforts are time-consuming and demand a large number of rescuers and resources. We aim to investigate the feasibility and effectiveness of using drones combined with an online machine learning (ML) model for automated recognition of simulated drowning victims.

Methods

This feasibility study used photos taken by a drone hovering at 40 m altitude over an estimated 3000 m2 surf area with individuals simulating drowning. Photos from 2 ocean beaches in the south of Sweden were used to (a) train an online ML model (b) test the model for recognition of a drowning victim.

Results

The model was tested for recognition on n = 100 photos with one victim and n = 100 photos with no victims. In drone photos containing one victim (n = 100) the ML model sensitivity for drowning victim recognition was 91% (95%CI 84.9%–96.2%) with a median probability score that the finding was human of 66% (IQR 52−71). In photos with no victim (n = 100) the ML model specificity was 90% (95%CI: 83.9%–95.6%). False positives were present in 17.5% of all n = 200 photos but could all be ruled out manually as false objects.

Conclusions

The use of a drone and a ML model was feasible and showed satisfying effectiveness in identifying a submerged static human simulating drowning in open water and favorable environmental conditions. The ML algorithm and methodology should be further optimized, again tested and validated in a real-life clinical study.

Section snippets

Background

Approximately 400,000 individuals suffer from fatal accidental drowning annually, with a majority of cases in low and middle income countries (LMIC).1 The number of fatal drownings are most likely underestimated and pose a large Global health problem affecting mainly children and adolescents.2 Drowning is defined as the “process of experiencing respiratory impairment from submersion/immersion in liquid”, and thus also includes incidents without fatal outcome.3, 4 Taking into account non-fatal

Methods

This is a simulation study using a dataset of drone-photos of voluntarily submerged surf lifeguards and an online machine learning model in order to test the potential of recognizing simulated drowning victims. Photos were used to train a ML-model and tested on another dataset to test sensitivity and specificity in recognizing victims. False positives were ruled out manually by a single researcher who was blinded from data collection and training of the ML-model.

Sample

A total number of 1 386 photos were collected over 8 summer months (May–August) during 2018 and 2019 with no adverse events reported during data collection. After exclusion of doublets and photos containing two individuals 593 photos from 35 series of 38 individuals (23 male lifeguards-61%) were eligible for final training, testing and analysis using an online ML model-tool for detection of a drowning victim.

Feasibility, effectivity

In 100 drone images when the search area contained one victim the ML model sensitivity

Discussion

The main finding of this study was that an off the shelf drone and a low-cost ML model was feasible in this setting for automated recognition of simulated drowning victims with acceptable sensitivity and specificity, both reaching values of >90%. Adding simultaneous manual interpretation by the researcher, all false objects could be ruled out. A methodology of combining automated recognition using ML alongside the manual interpretation of one pilot and one assistant screening the tablet display

Limitations

This is a simulation study, there are differences in simulated settings vs real life settings. There is limited information of the online ML models prerequisites and thereby potential to actually detect human victims in this setting. A different ML model may display results in another way. Drone photos were relatively homogenous as they were taken summertime in Sweden from only n = 35 different photo series, more variation in beach settings and more photo series could perhaps have changed the

Conclusion

The use of a drone and a ML model was feasible and showed satisfying effectiveness in identifying a submerged static human simulating drowning in open water and favorable environmental conditions. The ML algorithm and methodology should be further optimized, again tested and validated in a real-life clinical study.

Funding

The study was partly funded by the Laerdal Foundation.

Conflict of interest

None of the authors have any financial or personal relationships with other people or organisations that could inappropriately influence (bias) their work. None of the authors have any conflict of interest.

CRediT authorship contribution statement

A. Claesson: Conceptualization, Data curation, Formal analysis, Funding acquisition, Project administration, Visualization, Writing - original draft, Writing - review & editing. S. Schierbeck: Writing - original draft, Writing - review & editing. J. Hollenberg: Conceptualization, Funding acquisition, Resources, Writing - review & editing. S. Forsberg: Conceptualization, Methodology, Writing - review & editing. P. Nordberg: Conceptualization, Methodology, Writing - review & editing. M. Ringh:

Acknowledgements

We wish to acknowledge the Swedish lifesaving society and the Göteborg and Tylösand SLSC for their contribution in collecting data.

Cited by (0)

View full text
© 2020 Elsevier B.V. All rights reserved.