Machine learning predicts ICU mortality in cardiogenic shock: interpreting SOFA, lactate, and pH thresholds.

backgroundCardiogenic shock (CS) is a critical condition characterized by high mortality and clinical heterogeneity, creating a pressing need for advanced prognostic tools beyond traditional methods.

methodsWe analyzed 1827 CS patients from the MIMIC-IV database. Key predictive features were identified using a dual strategy combining the Boruta algorithm and LASSO regression. Eight machine learning models were developed and evaluated for discrimination, calibration, and clinical utility.

resultsAmong the included patients, 27.1% died during their ICU stay. The Logistic Regression model achieved the best overall performance among the evaluated algorithms (AUC 0.775 [95% CI: 0.722–0.821], Accuracy 0.712, Precision 0.787, F1-score 0.762) and demonstrated good calibration with a Brier score of 0.158. The classification threshold was optimized using Youden’s Index. Compared to complex ensemble models, Logistic Regression provided a robust and highly interpretable solution without requiring extensive hyperparameter tuning. Crucially, SHAP analysis provided model interpretability, identifying SOFA score, lactate, and pH as the most influential predictors. The model further quantified clinical risk gradients: mortality risk increased markedly when passing critical transition zones, specifically corresponding to SOFA scores of 8–10, lactate levels of 2.5–3.0 mmol/L, and deviations from a physiological pH range of 7.35–7.40.

conclusionThis study developed an interpretable machine learning model that predicts ICU mortality in CS patients with promising internal performance. By providing exploratory, cohort-specific risk ranges, the model identifies potentially relevant predictors that offer hypothesis-generating insights, which require further external validation before clinical implementation.