Objective:

To develop an interpretable predictive model for in-hospital mortality in patients with sepsis-induced myocardial injury (SIMI) and validate it across multiple centers.

Approach:

• Feature Selection: LASSO regression was used for feature selection.
• Model Evaluation: Eight machine learning algorithms were evaluated, including Gradient Boosting Machine (GBM), XGBoost, and Logistic Regression, with performance assessed via AUC, sensitivity, specificity, and calibration curves.
• Interpretability: SHapley Additive exPlanations (SHAP) were employed to interpret feature contributions.
• Model Optimization: Recursive feature elimination was used to optimize the model for clinical usability.

Key Findings:

• The GBM model achieved an internal validation AUC of 0.751, outperforming other algorithms.
• External validation yielded AUCs of 0.924 in eICU and 0.703 in a Chinese cohort.
• Key predictors identified included APS III Score, Hypertension, Albumin, Diabetes, SOFA Score, ALT, RBC, and Lactate.
• A simplified model with five variables achieved an AUC of 0.789.

Interpretation:

The developed model provides a user-friendly platform for real-time risk assessment of in-hospital mortality in SIMI patients.

Limitations:

• The study is retrospective and relies on existing databases.
• The generalizability of findings may be limited to similar patient populations.

Conclusion:

An interpretable predictive model for in-hospital mortality in SIMI patients was developed and deployed, potentially aiding in early risk assessment.