Mortality Prediction and Risk Factor Identification in Heart Failure Patients Using 17 SHAP-Interpreted Machine Learning Algorithms

Background: Heart failure (HF) is a leading cause of morbidity and mortality in intensive care units (ICUs). While echocardiography is routinely used for prognosis, critical predictors distinguishing survivors from non-survivors remain poorly defined across diverse clinical settings. Machine learning (ML) offers potential to improve risk stratification by identifying key mortality determinants. Methods: In this retrospective study, we analyzed data from two large ICU databases: MIMIC-IV and eICU. Adult patients (>18 years) with HF who underwent echocardiographic evaluation during their ICU stay were included. After feature selection using the Boruta algorithm, we developed and compared 17 ML models to predict in-hospital mortality, including Random Forest (RF), Extreme Gradient Boosting (XGBoost), Logistic Regression (LR) and others. Models were internally validated on a test subset of MIMIC-IV data and externally validated on the eICU cohort. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) was applied to identify and rank key mortality predictors. Results: A total of 3,485 patients were included, with 2,738 from the MIMIC-IV database and 747 from the eICU database. Thirty-two potential predictors were identified through screening using the Boruta algorithm. Receiver operating characteristic (ROC) curves analysis showed that the RF model consistently demonstrated superior performance across all validation sets and achieved the highest AUC (0.940) in the combined MIMIC-IV and eICU datasets. SHAP algorithm analysis revealed that the clinical features with the greatest impact on model outputs were SAPS II score, OASIS score, SOFA score, age, white blood cell count within the first 24 hours, and first 24-hour blood urea nitrogen. Clinical interpretability analysis indicated that the positive contributions (SHAP values) of these features were 0.115, 0.090, 0.084, 0.046, 0.035, and 0.015, respectively, demonstrating the model's good granularity and transparency in individualized prediction. Conclusions: This study compared 17 prediction models, among which the RF model demonstrated the best predictive performance for mortality in HF patients within the ICU, and may assist clinicians in identifying and managing high-risk HF patients.