Enhancing Real-World Fall Detection Using Commodity Devices: A Systematic Study.

Area of Science:

• Wearable technology
• Biomedical engineering
• Gerontology

Background:

• Smartphones and smartwatches offer non-intrusive fall detection using built-in sensors.
• Existing fall detection systems struggle with soft falls and real-world data discrepancies.
• Offline-trained models show performance drops when deployed in real-time settings.

Purpose of the Study:

• To investigate if combining gyroscope and accelerometer data from wrist and hip locations improves fall detection accuracy.
• To address the performance gap between offline and real-time fall detection models.
• To reduce false positives in real-world fall detection scenarios.

Main Methods:

• Systematic experimentation with accelerometer and gyroscope data from wrist and hip sensors.
• Utilized a Transformer-based neural network for fall detection model development.
• Retrained the model using real-time feedback data, including false and true positives.

Main Results:

• Combining hip and wrist accelerometer data achieved an 88% F1-score offline.
• Real-time deployment of the initial model resulted in numerous false positives.
• Retraining with real-time feedback improved the F1-score to 92%, reducing false positives and maintaining true fall detection accuracy.

Conclusions:

• Multi-sensor fusion (wrist and hip accelerometers) enhances fall detection performance.
• Real-time feedback is crucial for refining fall detection models in uncontrolled environments.
• The improved model demonstrates good generalization for older adults' movement patterns with minimal false positives.