Construction of a Model for In‐Hospital Death Prediction in Critically Ill Sepsis‐Induced Myocardial Injury Patients: An Interpretable Machine Learning Approach

Background To obtain the essential factors and construct an interpretable model for predicting in‐hospital death in critically ill patients with sepsis‐induced myocardial injury (SIMI) using machine learning (ML) methods. Methods This retrospective study collected data from the MIMIC‐IV database. We employed random forest (RF), XG Boost, and logistic regression (LR) algorithms to rank variable importance, identifying lactate as a core factor. Next, the association of lactate with in‐hospital death in SIMI was analyzed using LR analysis. A comprehensive lactate‐based LR model was subsequently constructed and interpreted by the Shapley additive explanation (SHAP). Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), decision curve analysis (DCA), integrated discrimination improvement (IDI), and calibration plots. Internal validation was performed to assess the model’s generalizability. Results Among 7309 enrolled patients, 1526 experienced in‐hospital death. Lactate elevation was independently associated with increased in‐hospital mortality in SIMI patients ( p < 0.05). The SHAP analysis indicated that lactate, vasopressor use, and SpO 2 were the top three contributors to predictions in the lactate‐based model. Both the comprehensive lactate‐based model and the simplified lactate–vasopressor–SpO 2 model demonstrated superior discriminative ability (AUCs: 0.767 and 0.727, respectively), compared to the Sequential Organ Failure Assessment (SOFA) score (AUC: 0.623), with consistently greater clinical net benefit and improved reclassification. All models showed good calibration. Conclusion Lactate, as the key factor, was positively associated with in‐hospital death in SIMI. The simplified lactate–vasopressor–SpO 2 model offers an optimal balance of predictive accuracy, interpretability, and clinical practicality for risk stratification.