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

Area of Science:

• Gerontology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Computerized posturography quantifies balance control for fall risk assessment.
• Machine learning (ML) integration shows promise, but posturography's superiority over conventional methods needs documentation.
• Existing ML models lack transparency in fall risk prediction.

Purpose of the Study:

• Compare predictive performance of various data combinations for fall risk.
• Introduce a novel ML approach using a Large Language Model (LLM) for enhanced prediction and transparency.
• Evaluate LLM's ability to provide feature-based explanations for predictions.

Main Methods:

• Followed 206 community-dwelling older adults for 6 months to track fall events.
• Collected baseline data: demographics, questionnaires, physical tests, and posturography.
• Evaluated traditional ML models and an LLM with Quantized Low-Rank Adaptation (QLoRA) for predictive validity.

Main Results:

• 6-month fall incidence was 16.9%.
• Traditional ML models achieved an Area Under the Curve (AUC) of 0.54–0.71.
• LLM with QLoRA on posturography alone achieved a higher AUC (0.88) and accuracy (0.86).

Conclusions:

• Postural control is strongly related to fall risk in older adults.
• LLMs, particularly with QLoRA, significantly enhance fall risk prediction accuracy using posturography.
• LLMs offer improved transparency and reduce the need for expert annotation in fall risk assessment.