Physiology-informed deep learning for early identification of sepsis-induced cardiomyopathy from echocardiographic cine loops: a multicenter development and external validation study

Abstract Background Sepsis-induced cardiomyopathy (SICM) is heterogeneous and may involve abnormalities in cardiac contraction, relaxation, and beat-to-beat motion. Conventional echocardiographic measurements and cardiac troponin T (cTnT) may not fully capture dynamic changes in myocardial function during sepsis. We developed a physiology-informed deep-learning framework for early identification of SICM from apical four-chamber (A4C) echocardiographic cine loops. Methods Adult patients with sepsis were retrospectively identified from two centers. The first eligible A4C cine loop obtained within 24 h after sepsis onset was used for model development. SICM was defined by cardiac abnormalities observed during the septic episode. Cine loops were divided into physiologically aligned end-diastole-to-end-diastole cardiac cycles. The model combined EchoJEPA-based visual representation learning with phase-aware spatial and temporal modeling, LV area-derived physiological features, and multi-cycle aggregation. Internal validation was performed using patient-level grouped fivefold cross-validation. The final model was trained in the full development cohort and independently evaluated in MIMIC-IV-ECHO using a conventional cTnT-based SICM phenotype. Model performance was compared with direct EchoJEPA classification. Results The development cohort included 530 patients, including 305 with SICM and 225 sepsis-only controls. In patient-level grouped fivefold cross-validation, the physiology-informed model achieved an out-of-fold AUROC of 0.933 (95% CI, 0.910–0.954). The external cohort included 93 patients, including 39 with SICM and 54 sepsis-only controls. The model achieved an AUROC of 0.929 (95% CI, 0.864–0.980). The direct EchoJEPA baseline achieved an external AUROC of 0.735 (95% CI, 0.632–0.833). The paired difference in AUROC was 0.194 (95% CI, 0.095–0.294; DeLong test, P < 0.001). Results remained consistent in sensitivity analyses using alternative cTnT definitions. Conclusions A physiology-informed deep-learning framework enabled early identification of SICM from routinely acquired A4C cine loops and maintained discrimination in both internal cross-validation and an independent external cohort. Incorporating cardiac physiology into EchoJEPA improved discrimination compared with direct foundation-model classification. These findings support further prospective evaluation of physiology-informed echocardiographic video analysis for early identification of SICM.