Detecting falls as novelties in acceleration patterns acquired with smartphones

Area of Science:

• Biomedical Engineering
• Gerontology
• Machine Learning

Background:

• Falls in the elderly represent a significant public health issue, with current detection methods lacking efficiency.
• Existing fall detection systems often rely on acceleration data from smartphones and supervised machine learning, but use simulated falls from younger individuals, potentially limiting real-world applicability.
• A key challenge is the discrepancy between simulated falls and actual fall scenarios experienced by older adults.

Purpose of the Study:

• To explore the efficacy of novelty detection methods for fall detection in the elderly, using only data from activities of daily living (ADL).
• To compare a selected novelty detection method against a state-of-the-art supervised machine learning algorithm.
• To assess the adaptability of novelty detection systems to new ADL through on-the-fly retraining.

Main Methods:

• Utilized smartphone-recorded data, collecting ADL from ten volunteers in real-life scenarios and simulated falls.
• Investigated several novelty detection techniques, selecting the nearest neighbour-based (NN) method as the most suitable.
• Compared the performance of the adapted NN method against a generic Support Vector Machine (SVM) under various conditions.

Main Results:

• The study found that a generic Support Vector Machine (SVM) generally outperformed the adapted nearest neighbour (NN) novelty detection method across most tested conditions.
• Novelty detection offers a potential advantage in adaptability by learning from true ADL and identifying falls as deviations from normal movement.
• Publicly available datasets were created to enhance the reproducibility of fall detection research.

Conclusions:

• While a generic SVM showed superior performance in this study, novelty detection methods present a promising avenue for developing adaptive fall detection systems.
• The ability to continuously learn and adapt to new ADL could lead to more personalized and effective fall detection for the elderly.
• Further research is warranted to optimize novelty detection algorithms for real-world elderly fall detection applications.