Cardiovascular and Autonomic Phenotypes Reveal Distinct Mechanisms of Sepsis Decompensation via Deep Learning.

Sepsis heterogeneity reflects diverse etiologies and patient-specific physiological responses, motivating phenotype identification to enable precision therapeutics. However, most phenotyping approaches rely on intermittently sampled clinical variables, whereas continuously recorded physiological waveforms remain underutilized. We developed a deep-learning framework to derive physiological phenotypes from five-minute pre-onset electrocardiogram, photoplethysmogram and respiratory-impedance waveforms in 2,174 ICU patients meeting Sepsis-3 criteria. From these signals, 192 cardiorespiratory physiomarkers were extracted and embedded using a Feature Tokenizer Transformer encoder, which outperformed alternative representation methods. Consensus clustering identified four stable sepsis physio-phenotypes (SP-1-SP-4) associated with distinct autonomic and peripheral vascular signatures. Despite similar baseline severity and demographics, phenotypes differed significantly in mortality (19-29%), septic shock, vasopressor use and mechanical ventilation, with divergent 28-day survival trajectories (P<0.01). Explainable AI provided clinically interpretable characterizations, and a trained classifier enabled real-time bedside phenotyping. This framework establishes waveform-based phenotyping as a foundation for precision medicine in sepsis care.