An Explainable Artificial Intelligence Framework for Risk Prediction and Clinical Decision Support in Postoperative Complication Management

This is the authors' abstract. We don't add key points for this paper.

Postoperative complications remain a major concern in surgical care, creating a need for reliable risk prediction that can also provide clinically understandable explanations. This study developed an explainable artificial intelligence framework for postoperative complication prediction and decision support using data from 1,248 surgical patients, of whom 25.5% experienced complications. Logistic Regression, Random Forest, XGBoost, and CatBoost models were developed and evaluated using AUROC, accuracy, sensitivity, specificity, and F1-score. SHapley Additive exPlanations (SHAP) were applied to interpret the selected model, while predicted probabilities were translated into clinically meaningful risk categories. XGBoost demonstrated the strongest performance, achieving an AUROC of 0.902, accuracy of 84.8%, sensitivity of 80.0%, specificity of 86.5%, and F1-score of 0.728. SHAP analysis identified operative duration, ASA status, age, BMI, surgical complexity, and comorbidity burden as the leading contributors to predicted risk. Complication rates increased from 6.8% in low-risk patients to 58.9% in high-risk patients, demonstrating clear risk separation. The framework integrates predictive performance with individualized explanation and risk stratification, providing a transparent basis for postoperative clinical decision support. External and prospective validation is required before clinical implementation.