Authors:
Aneta Lisowska
1
;
Alison O'Neil
2
and
Ian Poole
2
Affiliations:
1
Toshiba Medical Visualization Systems Europe Ltd. and Heriot-Watt University, United Kingdom
;
2
Toshiba Medical Visualization Systems Europe Ltd., United Kingdom
Keyword(s):
Fall Detection, Accelerometer Data, Machine Learning, Wearable Devices.
Related
Ontology
Subjects/Areas/Topics:
Biomedical Engineering
;
Biomedical Signal Processing
;
Devices
;
Distributed and Mobile Software Systems
;
Health Engineering and Technology Applications
;
Health Information Systems
;
Human-Computer Interaction
;
ICT, Ageing and Disability
;
Mobile Technologies
;
Mobile Technologies for Healthcare Applications
;
Neural Rehabilitation
;
Neurotechnology, Electronics and Informatics
;
Pattern Recognition and Machine Learning
;
Physiological Computing Systems
;
Software Engineering
;
Wearable Sensors and Systems
Abstract:
Falls in seniors can lead to serious physical and psychological consequences. A fall detector can allow a fallen
person to receive medical intervention promptly after the incident. The accelerometer data from smartphones
or wearable devices can be used to detect falls without serious privacy intrusion. Common machine learning
approaches to fall detection include supervised and novelty based methods. Previous studies have found
that supervised methods have superior performance when tested on participants from the population cohort
resembling the one they were trained on. In this study, we investigate if the performance remains superior
when they are tested on a distinctly different population cohort. We train the supervised algorithms on data
gathered using a wearable Silmee device (Cohort 1) and test on smartphone data from a publicly available
data set (Cohort 2). We show that the performance of the supervised methods decreases when they are tested
on distinctly different data, but
that the decrease is not substantial. Novelty based fall detectors have better
performance, suggesting that novelty based detectors might be better suited for real life applications.
(More)