Developing and Validating an Interpretable Machine Learning Model for Predicting Postoperative Nausea and Vomiting in pulmonary resection patients: A retrospective observational study

Background Postoperative nausea and vomiting (PONV) remains a prevalent complication hindering Enhanced Recovery After Surgery (ERAS) protocols, particularly after pulmonary resection. While established clinical risk factors exist, the Apfel score lacks precision for individualized prediction and ignores neuroanatomical biomarkers. This study aimed to develop and validate an interpretable machine learning (ML) model integrating clinical variables and structural brain MRI features to predict PONV risk. Methods In this retrospective study, 416 patients undergoing pulmonary resection were analyzed. Preoperative T1-weighted MRI and 19 perioperative clinical variables were collected. A total of 1,174 structural brain features were extracted using FreeSurfer. Feature engineering included handling missing data, univariate screening bootstrap-LASSO for stable feature selection, and forced retention of key clinical factors Final feature subset was optimized via forward selection with 5-fold cross-validation. Eleven ML models were trained and hyperparameters tuned. Model performance was evaluated via internal and external validation. Results The optimal model (ExtraTrees classifier) achieved an AUROC of 0.82 on internal validation cohort and 0.80 on external validation cohort, significantly outperforming the Apfel score. Key predictive features included female sex, reduced left parieto-occipital cortical curvature, and decreased right occipital pole curvature/cuneus folding index. SHAP analysis confirmed these neuroanatomical features, alongside clinical factors, as primary drivers of individual predictions. Conclusions This study demonstrates that integrating structural neuroimaging biomarkers with clinical data significantly enhances PONV prediction after lung surgery compared to traditional models. The identified neuroanatomical features, particularly in visual and vestibular processing regions, provide novel biological insights into PONV susceptibility. Our interpretable ML model offers a promising tool for personalized PONV risk stratification, potentially guiding targeted prophylaxis within ERAS pathways. Validation in larger, prospective cohorts is warranted.