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

TL;DR

The study tackles the scarcity of elderly fall data by evaluating synthetic accelerometer data generated with LLMs (text-to-motion and text-to-text) and diffusion methods for wearable fall detection. It augments four real baselines (SMM, KFall, UMAFall, SisFall) with synthetic data and trains an $LSTM$ classifier on sliding windows of size $W=128$ with stride $S=10$, comparing performance and distributional alignment across methods. Diffusion-TS consistently yields the closest match to real data (lowest $JSD$, highest Coverage), while LLM-based data show variable alignment and dataset-dependent impact on F1-scores, performing best at low frequencies ($20$ Hz) and degrading at high frequencies ($200$ Hz). Text-to-motion models better capture joint-specific dynamics; however, demographic conditioning (gender/age) is limited, and the practical augmentation gains are modest, suggesting a practical path through fine-tuning text-to-motion generators and automating prompt strategies for robust synthetic fall data generation.

Abstract

Training fall detection systems is challenging due to the scarcity of real-world fall data, particularly from elderly individuals. To address this, we explore the potential of Large Language Models (LLMs) for generating synthetic fall data. This study evaluates text-to-motion (T2M, SATO, ParCo) and text-to-text models (GPT4o, GPT4, Gemini) in simulating realistic fall scenarios. We generate synthetic datasets and integrate them with four real-world baseline datasets to assess their impact on fall detection performance using a Long Short-Term Memory (LSTM) model. Additionally, we compare LLM-generated synthetic data with a diffusion-based method to evaluate their alignment with real accelerometer distributions. Results indicate that dataset characteristics significantly influence the effectiveness of synthetic data, with LLM-generated data performing best in low-frequency settings (e.g., 20Hz) while showing instability in high-frequency datasets (e.g., 200Hz). While text-to-motion models produce more realistic biomechanical data than text-to-text models, their impact on fall detection varies. Diffusion-based synthetic data demonstrates the closest alignment to real data but does not consistently enhance model performance. An ablation study further confirms that the effectiveness of synthetic data depends on sensor placement and fall representation. These findings provide insights into optimizing synthetic data generation for fall detection models.

Figures (6)

• Figure 1: Illustration of smartwatch-based fall detection: (a) A person wearing a smartwatch equipped with inertial sensors, which continuously monitor movement. (b) Representation of the internal accelerometer of the smartwatch, capturing movement along three axes: $x$ (red), $y$ (green), and $z$ (blue). (c) The accelerometer data, corresponding to the $x$, $y$, and $z$-axes, are visualized as time-series signals. (d) This data is then processed by machine/deep learning algorithms to detect patterns indicating a fall.
• Figure 2: Overview of the data generation process using two categories of pre-trained large language models (LLMs). (a) Text-to-Motion: Motion data is generated using models, T2M-GPT, SATO, and ParCo, and then, joint-specific data is extracted to create synthetic accelerometer data. (b) Text-to-Text: Three LLMs, GPT4o (ChatGPT), GPT4 (Copilot), and Gemini-1.5-Flash-8B (Google AI Studio), take prompts to simulate fall scenarios and generate accelerometer data directly for the left wrist.
• Figure 3: Process for generating synthetic accelerometer data with different characteristics using Text-to-Motion and Text-to-Text LLMs. (a) Text-to-Motion: Data selection from specific joints based on the SMPL human model, highlighting selected indices for neutral, left wrist, right wrist, and waist/pelvic data. (b) Text-to-Text: Prompt-based fall data generation using GPT4o, GPT4, and Gemini (1.5-Flash-8B).
• Figure 4: P-values from Kolmogorov-Smirnov tests comparing neutral and specific datasets for Text-to-Motion (T2M, ParCo, and SATO) and Text-to-Text LLMs (Gemini, GPT4, and GPT40).
• Figure 5: Comparison of normalized value distributions of real falls of four datasets, SmartFallMM, KFall, UMAFall, and SisFall, versus synthetic fall data generated from diffusion-based, text-to-motion, and text-to-text models.