Measuring fall risk of the elderly with IMU sensors by developing a Convolutional Neural Network

Master thesis (2022)

Authors

C.M. Claassen Mechanical Engineering

Contributors

E. van der Kruk Biomechatronics & Human-Machine Control - Mechanical, Maritime and Materials Engineering (supervisor 1)
F.C.T. van der Helm Biomechatronics & Human-Machine Control - Mechanical, Maritime and Materials Engineering (supervisor 1)
F Lefeber (supervisor 1)
J. Harlaar Biomechatronics & Human-Machine Control - Mechanical, Maritime and Materials Engineering (supervisor 2)
Faculty
Mechanical Engineering
Copyright: © 2022 Carlijn Claassen
More Info
expand_more

Published Date

25-08-2022

Language

English

Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Faculty

Mechanical Engineering

Abstract

The combination of the high number and the consequences of falls in older adults led to the development of fall risk assessments; non-sensor-based and sensor-based. Multiple studies used ML for older adults' fall risk prediction using raw IMU data. This study's objective was to develop a DL algorithm that predicts the fall risk of people living in a geriatric rehabilitation department using raw data collected from IMUs positioned at the ankles during the 10-m walk test.
Raw IMU data of 97 participants were used. The participants were classified as low, increased or high fall risk based on the Performance Oriented Mobility Assessment (POMA). Accelerometer and gyroscope's resultant time-series sequences (n=1037) were used as input for the Convolutional Neural Network (CNN) that was optimised and trained with 80% and tested with 20% of the participants. The results were compared with the performance of an existing portable sensor-based fall risk assessment called the Smart Floor (SF). The macro F1 of the unweighted (40%) and weighted (41%) multiclass classification CNNs was lower than the macro F1 of the SF (49%). The binary classification CNN's macro F1 (56%) was slightly lower than the SF's performance. All CNNs were better at predicting high-risk sequences. All models had poor performance when all three POMA fall risk categories should have been predicted. Adjustments to the data collection and CNN optimisation methods should be performed to study the possibility of predicting fall risk using raw IMU data in geriatric rehabilitation centres.