A Study of One-Class Classification Algorithms for Wearable Fall Sensors.

Area of Science:

• Wearable technology
• Biomedical engineering
• Machine learning

Background:

• Automatic fall detection is crucial for monitoring individuals, especially the elderly.
• Wearable inertial sensors and machine learning algorithms are increasingly used for this purpose.
• Supervised learning methods require fall data, which is difficult to collect.

Purpose of the Study:

• To systematically evaluate the performance of One-Class Classifiers (OCCs) for fall detection.
• To compare OCCs with traditional supervised methods.
• To identify optimal features and hyperparameters for OCCs in fall detection.

Main Methods:

• Utilized nine public datasets containing falls and Activities of Daily Living (ADLs).
• Applied and analyzed various typical OCCs with diverse input features and hyperparameters.
• Evaluated performance using standard metrics like specificity and sensitivity.

Main Results:

• OCCs achieved performance metrics comparable to supervised algorithms, with specificity and sensitivity exceeding 95%.
• The diversity of ADLs used for training significantly impacts OCC performance.
• High-mobility ADLs, if not adequately represented in training data, can lead to increased false alarms.

Conclusions:

• OCCs offer a viable alternative to supervised methods for fall detection, especially when fall data is scarce.
• Careful selection and diversity of ADLs during OCC training are essential for reliable fall detection systems.
• Further research should focus on optimizing OCCs with comprehensive ADL datasets to enhance robustness and reduce false positives.