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

In brief
The study developed and evaluated machine learning models on the INSPIRE perioperative dataset to predict 30-day postoperative mortality, testing imbalanced and balanced configurations with hybrid feature selection and multiple classifier families.
By achieving AUROCs above 0.92 and surfacing interpretable predictors such as albumin, ASA status, and sodium, the framework could make surgical risk assessment more transparent and actionable for counseling and optimization, though integration into clinical workflow remains prospective.
Main points
- Study used the Informative Surgical Patient dataset for Innovative Research Environment (INSPIRE) with patient-specific features such as ASA physical status, BMI, and temporal medical parameters.
- Three experimental setups were tested: imbalanced, balanced with feature selection, and balanced without feature selection, using SMOTEENN and hybrid MI plus RFE selection.
- Best models were Logistic Regression AUROC 0.958, XGBoost AUROC 0.953, and SVM AUROC 0.928, with balancing and feature selection improving discrimination and reliability.
The gain
Machine learning models trained on perioperative data achieved high discrimination for 30-day postoperative mortality and identified clinically meaningful predictors, supporting more transparent surgical risk assessment and patient counseling.
The rundown
Researchers built models on the INSPIRE perioperative dataset, incorporating ASA physical status, BMI, and temporal medical parameters, addressing imbalance with SMOTEENN and selecting features via Mutual Information combined with Recursive Feature Elimination.
They compared ensemble, boosting, and traditional classifiers across imbalanced and balanced setups, reporting top AUROCs for Logistic Regression, XGBoost, and SVM, and used SHAP and LIME to highlight predictors like preoperative albumin, ASA classification, and serum sodium.
Sources
- Peer-reviewedThe American Surgeon™2026-09-23
ace


The debate