Toward Clinically Interpretable Perioperative Decision Support: Explainable Machine Learning for 30-Day Postoperative Mortality Prediction.

Abstract

Postoperative mortality prediction remains a critical challenge in healthcare, demanding robust and interpretable predictive methods. This study developed and evaluated machine learning (ML) models using the novel Informative Surgical Patient dataset for Innovative Research Environment (INSPIRE), a comprehensive perioperative dataset, to predict 30-day postoperative mortality. Patient-specific features such as ASA physical status, BMI, and temporal medical parameters were incorporated. Data imbalance was addressed using SMOTEENN, and a hybrid feature selection approach was applied that combined Mutual Information (MI) and Recursive Feature Elimination (RFE). Three experimental setups: imbalanced, balanced with feature selection, and balanced without feature selection were used to assess multiple ML architectures including ensemble, boosting, and traditional classifiers. Logistic Regression (AUROC = 0.958), SVM (AUROC = 0.928), and XGBoost (AUROC = 0.953) achieved the best performance across experiments. Balancing and feature selection improved discrimination and model reliability. The clinically meaningful predictors identified by SHAP and LIME explainability analyses, such as preoperative albumin, ASA classification, and serum sodium, offered clear insights into the features contributing to increased mortality risk after surgery. The findings demonstrate the relevance of explainable machine learning for the identification of clinically relevant risk areas and may serve as a tool to classify risks in the future for perioperative use, as well as for patient counseling and optimization. The proposed framework illustrates the potential to increase the transparency of surgical risk assessment with explainable AI, and provides a foundation for decision support in the future to be integrated with the workflow.