Interpretable machine learning for predicting intraoperative oxygenation impairment during one-lung ventilation in thoracic surgery: a retrospective cohort study

Intraoperative oxygenation impairment during one-lung ventilation poses significant challenges in thoracic anesthesia. Traditional statistical models often fail to capture complex non-linear clinical relationships. This study aimed to evaluate machine learning algorithms for predicting this impairment and interpret potential feature associations using the SHAP method. Clinical data from 519 patients undergoing thoracic surgery were retrospectively analyzed. Five algorithms were evaluated. Model performance was measured using the area under the receiver operating characteristic curve. SHAP was applied to explore non-linear associations and feature interactions, without implying biological causality. The random forest model exhibited the highest area under the curve (0.924), performing better than logistic regression (0.790). SHAP analysis identified positive end-expiratory pressure, preoperative arterial oxygen partial pressure, and body mass index as the primary predictors. Moreover, the analysis revealed potential U-shaped protective associations for positive end-expiratory pressure (4 to 6 cm of water) and tidal volume, alongside synergistic interactions between baseline pulmonary conditions and mechanical ventilation settings. The random forest algorithm may provide superior discriminative performance for predicting oxygenation impairmen in this retrospective cohort. Interpretable machine learning provides exploratory hypothesis-generating insights into complex non-linear thresholds and interactions, which may serve as preliminary reference information for future prospective studies on individualized anesthetic management.