Development of an automated fall detection device specific to wheelchair users

Abstract

Purpose: Falls are a major health concern among the 65 millions individuals who use a wheelchair worldwide.¹ Older adults are four times more likely to use wheelchairs.² More than 60% of wheelchair users are affected by falls.³ After a fall, wheelchair users spend an average of 9 minutes (range 1-45 min) on the floor and 80% require assistance to recover.^4,5 Remaining on the ground for an extended period of time after a fall is associated with a risk of future injurious falls, long term care admissions, and death.⁶ Automated fall detection devices can be useful to provide timely assistance and minimize the consequences of a long lie.⁷ Although used widely in ambulatory populations, the ability to detect falls among wheelchair users is limited. Abou, et al⁸ examined the efficacy of a native fall detection app built into an Apple watch. The device was only able to detected 4.7% of falls from a wheelchair. The development of a fall detection device specific to wheelchair users is needed to facilitate a quick and accurate response after a fall. This presentation will describe the initial development of a fall detection algorithm and device preferences of older adult wheelchair users (target population).

Methods: Ethical approval obtained. The algorithm was developed based on repeated falls from a wheelchair with 30 participants.  Research-grade accelerometers were attached to the participants’ wrist, chest and head. Participants were first asked to performed a series of activities of daily living (e.g. transfers) to train the algorithm on the common, non-fall activities. Next, participants were asked to fall from a manual wheelchair in four directions: forward, left sideways, right sideways and backwards onto a pad.  Three fall trials were performed in each direction (12 falls total). For each trial, data were synchronized on the start impulse and divided into 4s sample windows. A 4s data window centered on the fall event or the non-fall data was created in each trial. 38 features were generated for each 4s trial window based on standard protocols.⁹ Once the features were generated and selected, a feature vector was formed. Features were then normalized in the range between fall and non-fall classes. During classification, feature vectors extracted from the data were fed into the model, which recognizes the activity (non-fall or fall) of the user. An SVM technique with a K-Fold Cross Validation training classifier was used to avoid overfitting.⁹ Feedback on the physical specifications of a fall detection device was obtain from interviews with the target population.

Results and Discussion: Accelerometers mounted at the chest and wrist presented with the best prediction outcomes to detect falls from a wheelchair, accuracy = 99.5% and 99.6%, respectively. Fifteen older adults who use wheelchairs reported they preferred a fall detection device to be in watch form.  Participants wanted the ability to modify who is contacted in the event of a fall and to cancel a notification of fall in the event of a false alarm. Further testing will be performed to examine accuracy in a community setting.