A Class-Imbalanced Deep Learning Fall Detection Algorithm Using Wearable Sensors.

Area of Science:

• Gerontology
• Biomedical Engineering
• Signal Processing

Background:

• Falls pose a significant health risk to the elderly, potentially causing irreversible injuries.
• Timely treatment after a fall is crucial, highlighting the need for accurate fall detection.
• Existing fall detection algorithms often perform poorly on real-world data due to class imbalance.

Purpose of the Study:

• To develop a fall detection algorithm capable of effectively distinguishing falls from Activities of Daily Life (ADL) signals.
• To address the challenge of class imbalance in wearable sensor data for fall detection.
• To improve the performance of fall detection systems in real-life applications.

Main Methods:

• Proposed a novel algorithm for fall detection.
• Evaluated the algorithm on class-imbalanced datasets representative of real-world scenarios.
• Compared the proposed method against state-of-the-art fall detection algorithms.

Main Results:

• The proposed algorithm demonstrated high performance on multiple evaluation metrics.
• Achieved a sensitivity of 99.33%, specificity of 91.86%, F-Score of 98.44%, and AUC of 98.35%.
• Outperformed existing algorithms, particularly on class-imbalanced data.

Conclusions:

• The developed algorithm is effective in distinguishing falls from ADL signals, even with imbalanced data.
• The proposed method is more suitable for real-life fall detection applications compared to previous approaches.
• This advancement can enhance the safety and timely treatment for elderly individuals at risk of falling.