Predicting Fall Risk in Community-Dwelling Older Adults Using a Fine-Tuned Quantized Large Language Model.

Computerized posturography has been employed to quantify an individual's intrinsic balance control under varying stances, thereby presenting the potential to support autonomous and ambulatory fall risk assessment when integrated with machine learning (ML) techniques. However, the superiority of posturography-based approaches over conventional methods such as questionnaires or physical performance tests remain insufficiently documented. In this study, we compared the predictive performance of various combinations of input data and introduced a novel ML approach that incorporates a Large Language Model (LLM) to enhance prediction while enabling feature-based, summarized explanations to improve the transparency of the predictions. We followed 206 community-dwelling older adults over a 6-month period to monitor fall events. At baseline, all participants completed a survey capturing demographic information, self-reported questionnaires, various physical performance tests, and four standing tasks assessed via tracker-based posturography. The predictive validity of these data in distinguishing fallers from non-fallers was evaluated using traditional ML models, and an LLM enhanced with Quantized Low-Rank Adaptation (QLoRA). The 6-month fall incidence was 16.9%. Traditional ML models achieved an area under the curve (AUC) ranging from 0.54 to 0.71 using different combinations of questionnaire responses, physical performance data, and posturographic parameters. Notably, a higher AUC (0.88) and accuracy (0.86) were achieved by applying the LLM with QLoRA to posturographic parameters alone. In conclusion, this study contributes to a deeper understanding of the relationship between postural control and fall risk, and demonstrates the potential of LLMs to improve predictive accuracy while minimizing the need for labor-intensive expert annotation.