Development and validation of a machine learning model for the preoperative prediction of intractable hypoxemia in repeat lung surgery

Hypoxemia, particularly when prolonged, poses a significant challenge during one-lung ventilation (OLV) for repeat pulmonary resection (RPR), as preoperative pulmonary function tests (PFTs) alone perform poorly in predicting risk. In this multicenter retrospective study of 2040 RPR patients from seven Chinese hospitals (2013-2023), 19 machine learning learners incorporating wrapper-based feature selection were compared to develop predictive models using perioperative dual-time-point (pre-first procedure and pre-second procedure) data. Intractable hypoxemia occurred in 15.4% of RPR patients. Models based solely on PFTs showed inadequate performance (Area Under the Receiver Operating Characteristic curve [AUROC] < 0.6 and Area Under the Precision-Recall Curve [AUPRC] < 0.2 in the training cohort), whereas the finalized 5-variable Multivariate Adaptive Regression Splines (MARS) model demonstrated consistent discrimination, achieving an AUROC of 0.790 (AUPRC: 0.445; baseline incidence: 0.133) in the training cohort, and AUROCs of 0.750, 0.765, and 0.782 (AUPRCs: 0.641, 0.529, and 0.473, against baseline incidences of 0.349, 0.173, and 0.180, respectively). Calibration was acceptable, and decision curve analysis showed a positive net benefit. By integrating dual-time-point computed tomography (CT)-derived volumetric and clinical features, the MARS model may support preoperative risk assessment for intractable hypoxemia during RPR, pending prospective validation.