Machine learning-based prediction of prolonged air leak after video-assisted thoracoscopic anatomical lung resection in older patients using cardiopulmonary exercise testing: a dual-centre development and validation study

Prolonged air leak (PAL) is among the most common complications after video-assisted thoracoscopic anatomical lung resection and significantly prolongs hospital stay and increases medical costs. However, existing prediction models often lack comprehensive verification and interpretability. The purpose of this study was to develop and validate an interpretable machine learning model combined with a cardiopulmonary exercise test (CPET) for predicting the risk of postoperative prolonged air leak in older patients with lung cancer. This was a dual-centre, retrospective and prospective cohort study. Patients (≥ 60 years old) with lung cancer who underwent thoracoscopic anatomical lung resection (lobectomy or segmentectomy) were enrolled from two centers: those treated at Qinhuangdao First Hospital (January 2024–June 2025) were retrospectively included in the model development cohort, while those from Tianjin Chest Hospital (June–August 2025) were prospectively enrolled for external validation. After preliminary data processing, the patient cohort was randomly divided into a training cohort (70%) and an internal validation cohort (30%). LASSO regression was used to screen the characteristic factors to determine the best features related to prolonged air leak (more than 5 days) after lung surgery. The SMOTE method was applied to correct the unbalanced datasets in the training set. Six machine learning models were constructed and comprehensively analysed to verify and evaluate their predictive ability. The optimal model was selected through rigorous evaluation. The Shapley Additive Explanation (SHAP) method was utilized to interpret the predictive model. A web-based risk calculator was developed for clinical use. A total of 471 patients were included in the development cohort and 140 in the external validation cohort. Four predictors were retained by LASSO: history of emphysema, smoking index, previous history of thoracic surgery or trauma, and VE/VCO₂. In internal validation, XGBoost achieved an AUC of 0.903 (95% CI: 0.834–0.960) with the highest F1 score (0.717) and lowest Brier score (0.092). In external validation, XGBoost achieved an AUC of 0.800 (95% CI: 0.656–0.913), sensitivity of 0.700, specificity of 0.833, and Brier score of 0.133. The most important feature was VE/VCO 2 , and four features of prolonged air leak after lung surgery were identified using SHAP technology on the basis of correlation ranking. A machine learning model based on cardiopulmonary exercise test parameters showed promising performance in predicting the risk of prolonged air leak in older patients after lung surgery.