Elsevier

Gait & Posture

Volume 26, Issue 2, July 2007, Pages 194-199
Gait & Posture

Evaluation of a threshold-based tri-axial accelerometer fall detection algorithm

We dedicate this paper to the memory of our late colleague and co-author, Jacinta O’Brien, who died suddenly on February 3, 2004: “Ar dheis Dé go raibh a hanam dílis”.
https://doi.org/10.1016/j.gaitpost.2006.09.012Get rights and content

Abstract

Using simulated falls performed under supervised conditions and activities of daily living (ADL) performed by elderly subjects, the ability to discriminate between falls and ADL was investigated using tri-axial accelerometer sensors, mounted on the trunk and thigh. Data analysis was performed using MATLAB to determine the peak accelerations recorded during eight different types of falls. These included; forward falls, backward falls and lateral falls left and right, performed with legs straight and flexed. Falls detection algorithms were devised using thresholding techniques. Falls could be distinguished from ADL for a total data set from 480 movements. This was accomplished using a single threshold determined by the fall-event data-set, applied to the resultant-magnitude acceleration signal from a tri-axial accelerometer located at the trunk.

Introduction

Falls affect over one in every three elderly people [1], [2]. They are the leading cause of injury deaths [3] and of injury-related hospitalisation [4] among the elderly population. Injuries sustained from falls can include broken or fractured bones, superficial cuts and abrasions as well as soft tissue damage [2], [5]. A serious consequence of sustaining a fall is also the ‘long-lie’, which is identified as involuntarily remaining on the ground for an hour or more following a fall [6]. The ‘long-lie’ occurs in more than 20% of elderly people admitted to hospital as a result of a fall [7]. Half of elderly people who experience a ‘long-lie’ die within 6 months, even if no direct injury from the fall has occurred [8].

Detection of a fall, either through automatic fall detection or through a personal emergency response system (PERS) might reduce the occurrence of the ‘long-lie’, by minimizing the time between the fall and the arrival of medical attention [9]. The most common existing PERS, the push-button pendant, is not always satisfactory because during a loss of consciousness or a faint the pendant might not be activated [10]. Moreover, some elderly people do not activate their PERS, even when they have the ability to do so [11].

A number of different approaches for the automatic detection of falls have appeared in recent years [12]. Some detect the impact of the body with the ground or the near horizontal orientation of the faller following a fall [12]. Most fall-detection systems detect the shock received by the body upon impact using accelerometers [12], [13], [14], [15]. For example, Diaz et al. [13] developed a primary fall-detection system which consisted of a small adhesive sensor patch that could be attached to the sacrum. Its fall detection accuracy was 100% (100% sensitivity) and only 7.5% of activities of daily living (ADL) were misdetected as falls. Hwang et al. [14] used a tri-axial accelerometer and gyroscope, both placed at the chest, to successfully distinguish between falls and ADL. This system had a sensitivity of 95.5% and specificity of 100%. However, ADL testing of the system was only for three young adults who performed sitting and a daily life activity.

To date fall-detection systems have used young subjects to test the extent of misdetection of ADL as falls. Elderly people often move differently than younger people as they typically have less control over the speed of their body movements due to reduced muscle strength with old age. As a result elderly people may “fall” into a chair when sitting down instead of sitting in a controlled manner and thus would be expected to produce higher peak accelerations when performing certain ADL. Thus, it was considered appropriate by the authors of this paper that the ADL-based measurements be performed using elderly subjects to increase the robustness of the test methodology.

This paper describes the development and testing of a threshold-based algorithm capable of automatically discriminating between a fall-event and an ADL, using tri-axial accelerometers. The accelerometer signals were acquired from simulated falls performed by healthy young subjects and from activities of daily living performed by elderly adults in their own homes. When a person falls and contacts the ground the forces to the body exceed those experienced during normal daily activities. We predicted that trunk and thigh tri-axial accelerometer signals would have peak values during a fall, which would be distinct from the signals produced during the performance of normal ADL.

Section snippets

Materials and method

Trunk and thigh longitudinal, anterior/posterior and medial–lateral accelerometer readings were recorded during simulated falls and ADL tasks. As it was not appropriate to subject elderly people to simulated falls, the first study involved young subjects performing simulated falls, in a safe controlled environment, under the supervision of a physical education professional. The second study involved elderly subjects performing ADL tasks in their own homes. Both studies were completed with

Results

The UFT and LFT for the trunk and thigh obtained by analysing the accelerometer signals from the 240 falls are summarised in Table 1. The trunk and thigh thresholds were also applied to determine the number of ADL tasks correctly identified as non-falls. The results show that 67–100% of ADL tasks were correctly classified Table 1. The UFT for each signal gave higher specificity than the LFT value. The UFTTHIGH provided a specificity of 83.3%, which was better than the LFTTHIGH (67.1%). For the

Discussion and conclusion

We have investigated signals from tri-axial accelerometers placed at the trunk and thigh, to determine if their peak values could be used to discriminate between ADL and falls. Using the trunk upper fall threshold (UFTTRUNK), all ADL tasks were correctly detected as non-falls. The closest this signal threshold came to being exceeded by one of the 240 ADL tasks was for, “Sitting on a kitchen chair”. Even for this task a reasonable error margin of 0.36 g (14.3%) was available.

Based on our results,

Acknowledgments

The authors wish to acknowledge the assistance of Mr. David Mahedy, Director of Sport at the University of Limerick for providing access to the crash mat facilities during the course of this research project. The authors wish to thank John Kiely for his assistance in supervising a number of the fall events and Dr. Richard Conway for his invaluable contribution and direction.

References (24)

  • R.E. Roush et al.

    Impact of a personal emergency response system on hospital utilization by community-residing elders

    South Med J

    (1995)
  • R.J. Gurley et al.

    Persons found in their homes helpless or dead

    N Engl J Med

    (1996)
  • Cited by (708)

    View all citing articles on Scopus
    View full text