AI-Generated Fall Data: Assessing LLMs and Diffusion Model for Wearable Fall Detection.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Data Science

Background:

• Training fall detection systems is hindered by a lack of real-world fall data, especially from elderly populations.
• Synthetic data generation offers a potential solution to augment limited real-world datasets.
• Evaluating advanced AI models for realistic data simulation is crucial for improving healthcare technologies.

Purpose of the Study:

• To investigate the efficacy of Large Language Models (LLMs) in generating synthetic fall data for fall detection systems.
• To compare the performance of text-to-motion and text-to-text LLMs against a diffusion-based method for synthetic data generation.
• To assess the impact of LLM-generated synthetic data on the performance of a Long Short-Term Memory (LSTM) fall detection model.

Main Methods:

• Generated synthetic fall data using various LLMs (GPT4o, GPT4, Gemini) and text-to-motion models (SATO, ParCo).
• Integrated synthetic datasets with real-world baseline datasets for training and evaluating an LSTM fall detection model.
• Compared LLM-generated data and diffusion-based synthetic data against real accelerometer data distributions.

Main Results:

• LLM-generated synthetic data improved fall detection performance, particularly in low-frequency settings (20 Hz), but showed instability in high-frequency datasets (200 Hz).
• Text-to-motion models yielded more biomechanically realistic data than text-to-text models, though their impact on detection varied.
• Diffusion-based synthetic data closely matched real data distributions but did not consistently enhance model performance.

Conclusions:

• The effectiveness of synthetic data for fall detection is contingent on dataset characteristics, sensor placement, and fall representation.
• LLMs offer a viable approach for generating synthetic fall data, with text-to-motion models showing potential for realistic biomechanical simulation.
• Further research is needed to optimize synthetic data generation strategies for robust and reliable fall detection systems across diverse conditions.