Enhancing Wearable Fall Detection System via Synthetic Data.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Data Science

Background:

• Deep learning models require substantial training data, which is often limited in clinical applications, particularly for fall detection.
• Acquiring diverse and comprehensive real-world fall datasets is a significant challenge, hindering the development of robust detection systems.

Purpose of the Study:

• To investigate methods for generating realistic synthetic multivariate fall data to augment limited real-world datasets.
• To evaluate the effectiveness of Diffusion-based generative AI and video-based pose estimation for synthetic fall data generation.
• To assess the impact of synthetic data on the performance of deep learning models for fall detection.

Main Methods:

• Applied conventional time-series augmentation, Diffusion-based generative AI, and video-based pose estimation to extract fall segments from public footage.
• Generated synthetic multivariate fall data using Diffusion models and video-based pose estimation, tailored for specific sensor placements.
• Evaluated synthetic data quality using quantitative metrics (FID, Discriminative Score, Predictive Score, JSD, KS test) and visual inspection.

Main Results:

• Diffusion-based synthesis generated the most realistic and distributionally aligned synthetic fall data.
• Incorporating Diffusion-based synthetic data improved the offline F1-score of a long short-term memory (LSTM) model by 7-10%.
• Real-time fall detection performance, tested with the SmartFall App, was boosted by 24% with the inclusion of synthetic data.

Conclusions:

• Synthetic data generation, particularly using Diffusion models, is a viable strategy to overcome data limitations in clinical fall detection.
• The developed methods create realistic and diverse synthetic fall data, enhancing the robustness and real-world applicability of fall detection models.
• This research highlights promising, underexplored directions in leveraging generative AI and video analysis for improved fall detection systems.