Machine learning-based prediction of prolonged air leak after uniportal video-assisted thoracic surgery segmentectomy

Prolonged air leak (PAL), defined as air leakage lasting more than 5 days, remains a prevalent complication following uniportal video-assisted thoracic surgery (uVATS) segmentectomy, contributing to extended hospitalization, elevated medical expenditures, and higher postoperative morbidity. This study was designed to develop and internally validate machine learning models for predicting PAL after uVATS segmentectomy using a comprehensive set of clinical, surgical, and physiological variables.

A retrospective cohort study was conducted including 610 consecutive patients who underwent uVATS segmentectomy between January 2019 and September 2025. Candidate predictors included demographics, comorbidities, pulmonary function, surgical parameters, and tumor characteristics. The dataset was randomly divided into training (70%) and internal test (30%) cohorts. Five machine learning algorithms-logistic regression, elastic net regression, random forest, gradient boosting machine (GBM), and extreme gradient boosting (XGBoost)-were constructed. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), overall accuracy, and decision curve analysis. The Shapley additive explanation (SHAP) algorithm was applied to quantify feature contributions and enhance model interpretability.

A total of 76 (12.46%) patients developed PAL. The XGBoost model achieved the highest discriminative performance, with an AUC of 0.874 [95% confidence interval (CI): 0.833-0.906] in the internal test set, outperforming conventional regression and other machine learning models. SHAP analysis identified five dominant predictive factors: low body mass index (BMI), prolonged operative time, reduced percentage of carbon monoxide lung diffusion capacity (DLCO%), diabetes, and complex segmentectomy.

Machine learning models, especially the XGBoost algorithm, show promising internal performance for predicting PAL following uVATS segmentectomy by integrating multidimensional perioperative variables. The key predictors identified in this study offer evidence-based insights for perioperative risk stratification and individualized care strategies. External multicenter validation, calibration assessment, and prospective clinical evaluation are required before routine clinical implementation.