Predicting the Unpredictable: Machine Learning's Role in Sepsis Cardiac Arrest Mortality

Background Sepsis complicated by cardiac arrest (SCA) has a very high mortality rate. Traditional tools like the SOFA score inadequately address sepsis-specific factors. This study sought to create a machine learning model to predict in-hospital mortality in SCA early. Methods Adult SCA patients (n=1,431) from the MIMIC-IV 2.0 database showed a 39.6% in-hospital mortality rate. Predictors were gathered within 24 hours of ICU admission, covering demographics, comorbidities, vital signs, lab results, severity scores, and initial treatments. Using four feature selection methods, 12 key predictors were identified, including APSⅢ, age, CHF, lactate, pH, bicarbonate, sodium, ALT, ALP, AST, glucose, and GCS_min. Nine machine learning algorithms were trained, with Random Forest optimized via nested 10-fold cross-validation. Results In the independent test cohort (n=215), RF showed the highest predictive performance with an AUC-ROC of 0.84, accuracy of 0.78, sensitivity of 0.72, and specificity of 0.82. Calibration analysis (Brier score 0.155), decision curve analysis, and Kolmogorov–Smirnov statistics (0.54) supported its robustness and clinical relevance. SHAP analysis identified APSⅢ and GCS_min as key predictors, with metabolic and hepatic markers adding prognostic value. Conclusion Using only first-day ICU indicators, this study established and internally validated an interpretable RF-based model for predicting in-hospital mortality in SCA. The model demonstrated robust discrimination, reliable calibration, and clinical utility, supporting its potential use for early risk stratification. External multicenter validation is warranted to confirm generalizability and facilitate clinical translation.