A Multicenter, Interpretable Machine Learning-Based Survival Model for Predicting 28-Day Mortality Risk in Sepsis Patients with Heart Failure: A Retrospective Cohort Study and Performance Comparison with the SOFA Score

Background Patients with sepsis complicated by heart failure face an extremely high risk of death, and the traditional Sequential Organ Failure Assessment (SOFA) score has limited efficacy in predicting 28-day outcomes. Therefore, this study aimed to develop and validate a predictive model for 28-day mortality in these patients that outperforms the SOFA score. Methods Data were retrieved from the Medical Information Mart for Intensive Care-IV (MIMIC-IV), the eICU Collaborative Research Database (eICU-CRD), and the Intensive Care Unit (ICU) of Cangzhou Hospital of Integrated Traditional Chinese and Western Medicine (CZHITCM) in Hebei Province. Features were screened using the random survival forest (RSF) algorithm, followed by the construction of four machine learning–based survival models. The concordance index (C-index) was used as the primary metric to identify the optimal model. The Shapley additive explanations (SHAP) approach was adopted for feature visualization, and this optimal model was further compared with a Cox proportional hazards model incorporating only the SOFA score (SOFA-Cox model). In addition, predictive results for individual cases were analyzed and interpreted by combining SHAP with the Locally Interpretable Model-agnostic Explanations (LIME) method. Results This study enrolled 6644 patients and identified nine core features. The Gradient Boosting Machine (GBM) survival model was the optimal model. The GBM model demonstrated discriminative ability for 28-day mortality risk (C-index: 0.815 (95% CI 0.789–0.839); AUC: 0.850), calibration (Brier Score: 0.119; intercept: 0.002; slope: 1.005), a net benefit (ENB₁₅₋₃₅) of 0.126 at a 15%–35% threshold, and risk stratification with a hazard ratio (HR) of 7.75 (95% CI 5.91–10.18) and log-rank P  < 0.0001. The in-hospital validation accuracy of the GBM model reached 87.3%, and it could identify 31.4% of high-risk cases missed by the SOFA-Cox model. Combined analysis using SHAP and LIME confirmed the advantages of this model in dynamic clinical decision-making. Conclusion The GBM model developed in this study demonstrated robust predictive performance for 28-day mortality in patients with sepsis and heart failure, consistently surpassing the SOFA-Cox model across all evaluation metrics. Furthermore, SHAP analysis enhanced model interpretability, offering precise, individualized insights to support clinical decision-making. Trial registration Not applicable.