Machine learning models using data from the first 6 hours of ICU admission provided moderate early prediction of sepsis-induced cardiomyopathy, with GBM achieving an AUC of 0.717 (95% CI 0.660-0.775).
Cohort (n=1,359)
No
Can machine learning models using data from the first 6 hours of ICU admission accurately predict the development of sepsis-induced cardiomyopathy in critically ill patients?
Machine learning models utilizing early clinical data from the first 6 hours of ICU admission provide moderate but clinically meaningful prediction of sepsis-induced cardiomyopathy, driven primarily by markers of systemic inflammation, tissue hypoperfusion, and consumptive coagulopathy.
Effect estimate: AUC 0.717 (95% CI 0.660-0.775)
Background Sepsis‐induced cardiomyopathy (SIC) is a severe complication of sepsis associated with increased mortality, yet early prediction remains a clinical challenge. Machine learning models offer promise for early risk stratification, but their “black‐box” nature limits clinical adoption. Methods This retrospective cohort study utilized data from the MIMIC‐IV database. Adult patients meeting Sepsis‐3 criteria with ICU stays ≥ 24 h were included. SIC was defined as LVEF < 45% on echocardiography within 72 h of ICU admission. Predictor variables available within the first 6 h of ICU admission included demographics, comorbidity burden (Charlson Comorbidity Index), disease severity scores (APACHE III, SAPS II), vital signs, and laboratory values. Nine machine learning models were developed and evaluated using a 70%/30% train‐validation split with 10‐fold cross‐validation. Model performance was assessed by AUC, calibration curves, and decision curve analysis (DCA). SHAP was applied for model interpretation. Results A total of 1359 patients were included (SIC prevalence ∼22%). In the validation set, GBM and Regularized Logistic Regression (LR‐L1/EN) achieved the best discrimination (AUC 0.717, 95% CI: 0.660–0.775 and 0.658–0.776, respectively), followed by SVM‐RBF (0.714) and Decision Tree (0.710). Random Forest achieved the highest training AUC (0.786, 95% CI: 0.750–0.821) but lower validation AUC (0.660). DCA demonstrated net clinical benefit across threshold probabilities of 0.1–0.6. SHAP analysis identified albumin, lactate, APACHE III score, monocytes, platelets, and creatinine as the top predictors of SIC risk. Conclusion Machine learning models using data from the first 6 h of ICU admission provide moderate but clinically meaningful early prediction of SIC. SHAP interpretation revealed that markers of systemic inflammation, tissue hypoperfusion, and consumptive coagulopathy were key risk drivers. External validation in multicenter cohorts is needed before clinical deployment.
Yan et al. (Thu,) conducted a cohort in Sepsis-induced cardiomyopathy (n=1,359). Machine learning models (GBM and Regularized Logistic Regression) was evaluated on Early prediction of sepsis-induced cardiomyopathy (AUC 0.717, 95% CI 0.660-0.775). Machine learning models using data from the first 6 hours of ICU admission provided moderate early prediction of sepsis-induced cardiomyopathy, with GBM achieving an AUC of 0.717 (95% CI 0.660-0.775).