A Machine Learning Multi-Class Approach for Fall Detection Systems Based on Wearable Sensors with a Study on Sampling...

Area of Science:

• Biomedical Engineering
• Gerontology
• Machine Learning

Background:

• Falls pose significant risks to the elderly, leading to severe injuries, especially with delayed assistance.
• Existing fall detection systems (FDS) often lack nuanced classification of fall events.
• Previous research established the utility of waist-worn inertial measurement units for fall detection.

Purpose of the Study:

• To develop and evaluate an improved Fall Detection System (FDS) using machine learning.
• To introduce a multi-class classification approach for fall detection, categorizing falls into pre-fall, impact, and post-fall phases.
• To investigate the impact of sensor sampling rates on FDS performance.

Main Methods:

• Utilized the publicly available SisFall dataset, including activities of daily living and falls.
• Applied preprocessing and feature extraction techniques to the inertial sensor data.
• Compared five machine learning algorithms, focusing on ensemble methods like Random Forest and Gradient Boosting.
• Experimented with sensor sampling rates ranging from 1 Hz to 200 Hz.

Main Results:

• Ensemble learning algorithms, specifically Random Forest and Gradient Boosting, demonstrated superior performance.
• Achieved high accuracy with Sensitivity and Specificity both approaching 99%.
• A sampling rate of 50 Hz was found to be generally sufficient for accurate fall detection, though higher rates showed marginal improvements.

Conclusions:

• The proposed multi-class classification approach significantly enhances fall detection accuracy.
• Machine learning, particularly ensemble methods, is highly effective for real-time fall event analysis.
• Optimizing sensor sampling rates is crucial for balancing FDS performance and data processing efficiency, with 50 Hz being a practical threshold.