A multimodal multi-agent LLM framework for identifying key drivers of sleep disorders.

Introduction: Sleep quality and sleep disorders are influenced by interacting lifestyle, behavioral, physiological, and occupational determinants, but most existing studies examine these factors in isolation. Traditional statistical methods may be limited in modeling complex interactions, while many machine-learning approaches remain insufficiently interpretable for clinically meaningful sleep research. Methods: We developed an interpretable large language model (LLM)-based multi-agent multimodal framework for sleep disorder analysis. The framework includes three specialized agents: a Data Analyst Agent for identifying correlations, feature relevance, and interaction effects; a Physiology and Health Analyst Agent for contextual interpretation; and a Validation Analyst Agent for evaluating evidential grounding and consistency. The framework was applied to public and synthetic sleep-health datasets. Results: Pairwise bootstrap analyses showed weak or uncertain associations between bedtime consistency, light exposure, caffeine intake, stress, heart rate, and continuous sleep-duration or sleep-quality outcomes, whereas high caffeine intake was associated with elevated sleep-disorder risk. Physical activity effects differed by activity type: agility drills, jump tests, and lateral moves were more frequently linked to insomnia, whereas endurance running showed a stronger association with sleep apnea. Occupational context, psychological stress, stimulant use, and physiological indicators jointly influenced sleep disorder profiles, although several pairwise physiological associations remained weak and should be interpreted cautiously. Discussion: The proposed framework enhances interpretability, supports evidence-grounded reasoning, and reduces unsupported claims in multimodal sleep analysis. Because the second dataset was synthetic, the cross-dataset analysis should be interpreted as a controlled distributional robustness check rather than external clinical validation or evidence of broad clinical generalizability.