Sepsis cardiomyopathy read from whole heartbeats: aligned cycles lift external AUROC to 0.929
Synopsis
Sepsis-induced cardiomyopathy is heterogeneous and lacks agreed diagnostic criteria, so conventional echo measures and troponin capture only part of the dynamic change. A physiology-informed model cuts apical four-chamber loops into end-diastole-to-end-diastole cycles, derives left-ventricular area features, and fuses them with EchoJEPA visual representations and phase-aware attention before aggregating beats to a patient-level probability. In a 93-patient MIMIC-IV-ECHO external cohort it reached an AUROC of 0.929, against 0.735 for direct EchoJEPA classification, a paired difference of 0.194.
Interpretation
Early identification of sepsis-induced cardiomyopathy now runs on a single routinely acquired apical four-chamber loop and returns a patient-level probability, rather than resting on ejection fraction, TAPSE or troponin alone. Earlier video models largely treated echo clips as image sequences, cardiac phase was not always explicitly standardized before feature extraction, and ventricular contraction and filling had to be learned indirectly from visual features. In the 93-patient MIMIC-IV-ECHO external cohort the model reached an AUROC of 0.929 (95% CI 0.864–0.980), while direct EchoJEPA classification reached 0.735 (95% CI 0.632–0.833), a paired difference of 0.194 (95% CI 0.095–0.294; DeLong test, P<0.001).
The model was developed on 530 sepsis patients from two centers, 305 with sepsis-induced cardiomyopathy and 225 sepsis-only controls, and checked with patient-level grouped fivefold cross-validation. Apparent performance in the development cohort was higher than the cross-validated figure, so the held-out-patient estimate is the comparable internal reference. Pooled out-of-fold AUROC in patient-level grouped fivefold cross-validation was 0.933 (95% CI 0.910–0.954), against an apparent development-cohort AUROC of 0.978 (95% CI 0.966–0.988).
External discrimination held when the troponin reference definition changed, though it fell when the threshold was tightened. External labels came from conventional troponin cutoffs, and changing the cutoff changes both the number of positives and the difficulty of the task. The primary >0.10 ng/mL definition gave 39/93 positives and AUROC 0.929; a >0.20 ng/mL cutoff gave 22/93 positives and AUROC 0.808 (95% CI 0.715–0.889); excluding the 0.08–0.12 ng/mL gray zone gave 33/84 positives and AUROC 0.923 (95% CI 0.853–0.979).
Perspective
The demonstrated scope is discrimination and calibration in a retrospective two-center development cohort and the MIMIC-IV-ECHO external cohort, with input limited to the first analyzable apical four-chamber loop within 24 hours of sepsis onset; the source proposes prospective evaluation to determine whether model-assisted identification improves patient care.
The source notes that the development cohort was retrospective and two-center, with examinations from routine care rather than a standardized imaging protocol; cardiomyopathy status was determined at the episode level while the model used only the first examination within 24 hours of onset, and prediction of later dysfunction was not separately evaluated; the two cohorts used different reference definitions, and the small external cohort limits the precision of performance and calibration estimates.