An Event-Triggered Machine Learning Approach for Accelerometer-Based Fall Detection.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Traditional fixed-size sliding window methods for fall detection using accelerometers lose crucial information by not segmenting fall stages.
• This limitation can reduce the accuracy of machine learning classifiers in detecting falls.

Purpose of the Study:

• To propose an event-triggered machine learning (EvenT-ML) approach for fall detection that aligns data segments with distinct fall stages (pre-impact, impact, post-impact).
• To evaluate the effectiveness of EvenT-ML in improving fall detection rates and reducing computational costs.

Main Methods:

• Developed the EvenT-ML approach to segment data based on fall event triggers, aligning segments with specific fall stages.
• Utilized Classification and Regression Trees (CART), k-nearest neighbor (k-NN), logistic regression (LR), and support vector machine (SVM) classifiers.
• Evaluated the approach on two publicly available datasets using chest-worn and waist-worn sensors.

Main Results:

• EvenT-ML achieved high F-scores: 98% for chest-worn sensors and 92% for waist-worn sensors.
• Demonstrated significant reductions in computational cost, up to 8-fold for chest sensors and 78-fold for waist sensors, compared to fixed-size sliding window methods.
• Outperformed existing fall detection approaches in terms of F-score.

Conclusions:

• Aligning feature segments with fall stages using EvenT-ML significantly enhances fall detection accuracy.
• The EvenT-ML approach substantially reduces computational overhead, making it more efficient for real-time fall detection systems.