Interpretable machine learning for postoperative nausea and vomiting prediction in elderly orthopedic patients: a comparative study.

backgroundPostoperative nausea and vomiting (PONV) prolongs hospitalization and reduces patient satisfaction. Identifying high-risk elderly patients requires accurate absolute risk assessments, yet existing tools often lack probability calibration and transparency.

methodsWe included 1216 elderly patients undergoing elective hip or knee surgery. To strictly prevent data leakage, the dataset was partitioned into training, validation, and independent test sets in a 7:1:2 ratio prior to any imputation or feature selection. Following the systematic hyperparameter optimization of 12 distinct machine learning algorithms, a StackNet meta-model was developed by fusing optimal base-learner probabilities with raw clinical features. Clinical utility was evaluated via Brier scores and Decision Curve Analysis (DCA), alongside SHapley Additive exPlanations (SHAP) interpretability.

resultsOverall PONV incidence was 33%. The StackNet model achieved an AUC of 0.9338, significantly outperforming the conventional Logistic Regression baseline (AUC = 0.7564, p < 0.001) with superior calibration (Brier score = 0.102). On the independent test set, the StackNet model achieved an accuracy of 0.7860, sensitivity of 0.9250, specificity of 0.7178, and AUC of 0.9338, while the Logistic Regression baseline achieved an accuracy of 0.6584, sensitivity of 0.6750, specificity of 0.6503, and AUC of 0.7564. SHAP analysis identified preoperative frailty status and baseline hemoglobin levels as primary risk drivers.

conclusionThe StackNet framework offers highly calibrated absolute risk estimates for PONV in elderly orthopedic patients. Combined with SHAP transparency, it provides a clinically actionable tool to facilitate personalized antiemetic prophylaxis while avoiding unnecessary medical interventions due to overestimated risks.