LLM-Powered Dysphagia Screening With Multimodal Physiological Signal Analysis and Medically Informed Prompts.

The proposed LLM-based dysphagia screening framework achieved a classification accuracy of 96.3%, significantly outperforming baseline models.

• The framework integrates multimodal physiological signals including laryngeal vibration, nasal airflow, and swallowing sound.
• A medically-informed prompt template was designed to incorporate individual attributes, key biosignal features, and task instructions.
• Performance was described as significantly outperforming baseline models.
• The model leverages the reasoning capabilities of large language models (LLMs) for signal analysis.

The study recruited 217 participants generating 2,664 total samples for dysphagia screening evaluation.

• A total of 217 participants were recruited, including 109 post-stroke patients with dysphagia and 108 healthy individuals.
• The dataset comprised 1,391 dysphagic samples and 1,273 healthy control samples.
• Participants included both post-stroke patients and healthy individuals as controls.
• The dataset was designed to address challenges of limited data and complex multimodal signals.

The model demonstrated robust performance in few-shot learning settings, indicating strong generalization capabilities.

• The few-shot learning evaluation was used to assess model generalization with limited training data.
• Robust performance in few-shot settings was explicitly noted as a key strength of the framework.
• This characteristic is described as indicating 'strong generalization capabilities.'
• The few-shot robustness addresses the challenge of limited clinical data availability.

Traditional diagnostic techniques for dysphagia are limited by accessibility, reliability, and invasiveness, motivating the development of automated screening methods.

• Traditional techniques identified as limited include bedside screening and videofluoroscopic swallowing studies (VFSS).
• Dysphagia is described as a common complication among stroke patients.
• Dysphagia significantly increases the risk of aspiration pneumonia, malnutrition, and mortality.
• The framework was proposed specifically to address challenges of limited data and complex multimodal signals in clinical dysphagia detection.

The medically-informed prompt template was specifically designed to guide the LLM to focus on dysphagia-related patterns by incorporating individual attributes, key biosignal features, and task instructions.

• The prompt template incorporated three elements: individual attributes, key biosignal features, and task instructions.
• The design goal was to effectively guide the LLM to focus on dysphagia-related patterns.
• This prompt-driven reasoning approach was identified as a key component of the framework's effectiveness.
• The framework is described as offering 'extensive applicability in clinical practice.'