Causal Multiorgan Dynamic Innovations for Six-Hour Sepsis Onset Warning: A Nested Evaluation of Koopman-Inspired Features on PhysioNet/CinC 2019

Early-warning models for sepsis often combine heterogeneous physiologic measurements and temporal summaries, yet more elaborate representations may not improve clinically relevant prediction. We evaluated whether causal multiorgan innovations derived from fold-specific lifted linear dynamics improve prediction of sepsis onset 1–6 h in advance using PhysioNet/CinC 2019 cohorts A and B (40,336 patients; 1,552,210 hourly rows). Four XGBoost representations were compared under five-fold patient-grouped nested validation, including a causal-state baseline with six 8-h coefficient-of-variation features (C0) and a Koopman-inspired innovation model (C3). All hyperparameters, lifts, calibration procedures, and alarm thresholds were selected without access to outer test folds. Among 39,910 eligible patients, patient-balanced AP was 0.0739 for C0 and 0.0732 for C3 (no-skill AP 0.0142; difference −0.0007; 95% CI −0.0025 to 0.0010). C3 achieved AUROC 0.8183 versus 0.8161 for C0, with identical Brier scores (0.0136). At fold-specific thresholds, false-alarm episodes decreased from 0.1701 to 0.1395 per patient-day (difference −0.0306; 95% CI −0.0339 to −0.0268), whereas useful sensitivity (20.11% versus 19.39%) did not differ significantly and was nearly identical at equal false-alarm budgets. Thus, Koopman-inspired innovations did not improve primary AP, and their lower alarm burden reflected the selected operating thresholds rather than a better sensitivity–alarm trade-off.