Large Language Model-Driven Throat-Wearable Sensing System for Real-Time Recognition and Evaluation of Swallowing Disorders.

Abstract

Swallowing disorders are a common complication after stroke, yet current assessment methods rely largely on clinical observation and subjective screening, limiting continuous and objective evaluation. Here, we report a large language model (LLM)-driven throat-wearable sensing system (TWSS) for real-time monitoring of laryngeal activity and quantitative assessment of swallowing function. TWSS consists of a flexible sensing patch and a structured signal sequence-based LLM framework (S3-LLM). The flexible patch integrates a stretchable sensor with a wireless circuit module, enabling conformal attachment to the throat and real-time acquisition of physiological signals. Owing to its dual sensitivity to pressure and strain, the sensor can capture subtle and complex laryngeal movements associated with different physiological activities. S3-LLM converts them into structured signal sequences and combines temporal encoding with parameter-efficient fine-tuning, thereby exploiting the representation and generalization capabilities of LLMs under few-shot conditions. In a clinical validation involving 20 participants, TWSS achieved an accuracy of 92.4% for recognizing normal laryngeal activities and 87.6% for evaluating swallowing function, approximately 20% improvement over conventional models. These results demonstrate that TWSS provides a promising wearable platform for continuous, objective, and quantitative assessment of swallowing disorders and highlights the potential for personalized healthcare.