Transfer Learning on Small Datasets for Improved Fall Detection.

Area of Science:

• Gerontology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Falls in the elderly lead to severe health issues and mortality.
• Existing AI-powered fall detection systems lack real-time smartwatch application for seniors.
• Smartwatch systems face challenges with insufficient personalized data and model drift across devices.

Purpose of the Study:

• To develop and evaluate a transfer learning strategy for smartwatch-based fall detection in the elderly.
• To address limitations of small datasets and model generalization across heterogeneous devices.
• To improve the accuracy and reduce false positives of fall detection systems.

Main Methods:

• Collected accelerometer datasets from diverse devices (Microsoft Band, Huawei Watch, meta-sensor).
• Applied a transfer learning strategy to train fall detection models.
• Evaluated model performance across different datasets and devices.

Main Results:

• Transfer learning improved fall detection performance, with an average F1 score increase of over 10% and AUC increase of over 0.15.
• The transfer learning approach significantly reduced false positive prediction rates compared to non-transfer learning methods.
• Demonstrated the effectiveness of transfer learning in generalizing models across heterogeneous smartwatch devices.

Conclusions:

• Transfer learning is a viable strategy to overcome data scarcity and model drift in smartwatch fall detection.
• The proposed method enhances the reliability and scalability of real-time fall detection systems for the elderly.
• This approach holds promise for developing more robust and widely applicable fall detection technologies.