Explainable artificial intelligence for postoperative analgesia risk stratification and clinical decision support in abdominal surgery

Abstract

Background Early identification of patients at risk of inadequate postoperative analgesia remains a major challenge in perioperative medicine. Although machine learning approaches have shown potential for clinical outcome prediction, many existing models primarily emphasize discrimination performance while insufficiently addressing reliability, interpretability, and clinical applicability. This study aimed to develop and comprehensively evaluate an explainable artificial intelligence (XAI) framework for postoperative analgesia risk stratification and clinician-oriented decision support following abdominal surgery. Methods A retrospective cohort of 202 adult patients undergoing elective abdominal surgery and receiving postoperative analgesia with oxycodone or tilidine was analyzed. After exclusion of five patients with incomplete variables required for feature construction, 197 patient-level records were included in machine learning model development. Ten supervised learning algorithms were trained using demographic characteristics, surgical factors, analgesic information, and early postoperative pain trajectory features. Model performance was evaluated through stratified five-fold cross-validation using discrimination metrics, precision-recall analysis, calibration assessment, learning-curve analysis, and decision curve analysis (DCA). Model interpretability was assessed using SHAP-based feature attribution and cross-model explanation consistency analysis. Propensity score matching (PSM) was performed to evaluate the association between analgesic selection and postoperative outcomes after adjustment for measured baseline differences. A large language model (LLM) was integrated as a post hoc interpretation layer to translate model outputs into clinician-readable explanations. Results The evaluated models demonstrated moderate-to-good predictive performance, with validation AUC values ranging from 0.707 to 0.798. Early postoperative pain trajectory variables, particularly VAS-derived features, consistently represented the strongest predictors across different algorithms. Calibration analysis revealed differences in probability reliability that were not captured by discrimination metrics alone, while DCA demonstrated potential clinical utility of the best-performing models within clinically relevant decision thresholds. SHAP-based analyses showed stable feature attribution patterns across algorithms, supporting the robustness of identified predictors. After adjustment using PSM, analgesic type remained associated with postoperative analgesia outcomes, suggesting potential differences in real-world analgesic effectiveness while acknowledging the observational study design. Conclusions This study presents an integrated XAI framework combining predictive modeling, calibration assessment, decision-curve evaluation, explainability analysis, and language-based interpretation for postoperative analgesia risk stratification. The results highlight the importance of evaluating both predictive accuracy and clinical reliability when developing AI-assisted decision-support systems. Further multicenter prospective studies incorporating larger datasets and real-time clinical variables are required to validate model generalizability and evaluate implementation in perioperative care.