0328 LLM-Driven Multimodal Sequence Interpretation for Home-Based Sleep and Respiratory Assessment Using Fingertip Wearable Signals

Abstract Introduction Fingertip wearables enable continuous home monitoring of nocturnal physiology, including sleep, peripheral oxygen saturation (SpO₂), autonomic nervous system (ANS) activity, and apnea events. While each signal is informative, their clinical meaning emerges only when integrated, as sleep transitions, respiratory disturbances, and autonomic responses interact over short timescales. Conventional pipelines struggle to combine these heterogeneous signals. Large language model–driven agents can reason over complex multimodal patterns, synthesizing raw data into clinically meaningful assessments. This study evaluates whether multimodal sequence analysis combined with an LLM-based diagnostic agent can produce robust, human-interpretable insights into sleep and respiratory function using data derived from the TipTraQ device. Methods Full-night time series were temporally aligned to preserve physiological synchrony. Sequence analysis characterized intra-channel structure and cross-channel coupling, revealing apnea-driven disruptions, delayed desaturation, ANS activation during sleep transitions, and fragmentation during hypoxemia. These sequences were then provided to a diagnostic agent built on an LLM. The agent was designed not merely to summarize individual metrics, but to reason about the temporal evolution and physiological interdependence of the signals. Rather than relying on rule-based logic, the agent interpreted the nightly sequences as a cohesive physiological narrative, producing structured, clinical-style assessments that integrated sleep quality, respiratory burden, autonomic reactivity, and oxygenation profiles. Results Across 124 nights, deep sleep coincided with autonomic stability and fewer apneas, while REM showed sympathetic activation. Apnea clusters often preceded SpO₂ declines, sleep transitions aligned with ANS shifts, and elevated ODI co-occurred with apnea and arousal bursts. The agent integrated these patterns into unified interpretations, associating desaturation with preceding instability and identifying sleep fragmentation as both contributor and consequence of respiratory burden. Conclusion This study shows multimodal sequence analysis of sleep, SpO₂, ANS, and apnea signals can support a diagnostic agent capable of high-level clinical reasoning. By recognizing temporal dependencies and cross-channel coupling, the agent transforms heterogeneous physiological sequences into cohesive assessments of sleep architecture, respiratory disturbance, hypoxemia risk, and autonomic regulation. The findings highlight the diagnostic agent—not merely the sensing hardware—as a central component of the system, enabling fingertip-based monitoring to serve as an interpretable, scalable alternative for home-based sleep evaluation and preliminary screening for sleep-disordered breathing. Support (if any)