The research was led by Zixiao Wang, Jiabao Ding, Huagang Liang and colleagues at the First Hospital of Qinhuangdao, in collaboration with Kai Wang of Tianjin Chest Hospital. It addressed a well-recognized gap in thoracic surgical practice. While prolonged air leak — typically defined as air leaking from the chest drainage system for more than five days after surgery — is among the most frequent complications following video-assisted thoracoscopic anatomical lung resection, the prediction models available to clinicians have historically lacked rigorous external validation and, critically, interpretability. Surgeons have been asked to trust risk scores without understanding where those risks come from, a barrier that has limited clinical uptake of computational tools in the operating room planning process.

To build their model, the investigators enrolled patients aged 60 and older with lung cancer who underwent thoracoscopic anatomical lung resection — either lobectomy, the removal of an entire lung lobe, or segmentectomy, the removal of a smaller anatomical segment. The study was designed as a dual-centre cohort investigation with both retrospective and prospective components. A total of 471 patients treated at Qinhuangdao First Hospital between January 2024 and June 2025 formed the model development cohort. A separate group of 140 patients from Tianjin Chest Hospital, enrolled prospectively between June and August 2025, provided the external validation set — a crucial test of whether the model generalized beyond the institution where it was built.

The methodology followed the emerging best-practice playbook for clinical machine learning. After preliminary data processing, the development cohort was randomly split into a training set containing 70 percent of patients and an internal validation set holding the remaining 30 percent. Because prolonged air leak affects only a minority of surgical patients, the training data were inherently imbalanced, with far more non-leak cases than leak cases — a condition that can fool machine learning algorithms into simply predicting that no one is at risk. The researchers countered this with SMOTE, the Synthetic Minority Over-sampling Technique, which generates synthetic examples of the underrepresented class to rebalance the dataset and force the algorithms to learn the genuine signature of risk.

Feature selection was performed using LASSO regression — least absolute shrinkage and selection operator — a statistical technique that penalizes model complexity and drives the coefficients of irrelevant variables to zero, effectively filtering a large panel of candidate variables down to the most predictive few. From this process, four predictors of prolonged air leak emerged: a history of emphysema, the smoking index, a previous history of thoracic surgery or chest trauma, and the ventilatory equivalent for carbon dioxide, abbreviated VE/VCO₂. This last variable is drawn directly from cardiopulmonary exercise testing, a stress test performed on a bicycle or treadmill that measures how the heart, lungs, and circulatory system respond together to increasing exertion. VE/VCO₂ reflects the efficiency of ventilation during exercise — how much air a patient must move to eliminate a given amount of carbon dioxide — and elevated values signal impaired pulmonary gas exchange and often underlying parenchymal lung disease.

With the four features in hand, the team constructed six competing machine learning models: logistic regression, support vector machine, decision tree, extreme gradient boosting, and artificial neural networks, among others. Each was evaluated using a battery of metrics including the area under the receiver operating characteristic curve, or AUC, the F1 score, and the Brier score, which measures the accuracy of probabilistic predictions. The winner was XGBoost, an ensemble method that builds a sequence of decision trees, with each new tree trained to correct the residual errors of its predecessors. In internal validation, XGBoost achieved an AUC of 0.903 (95 percent confidence interval, 0.834 to 0.960) — well within the range conventionally considered excellent discrimination — along with the highest F1 score of 0.717 and the lowest Brier score of 0.092 among all six candidates.

The external validation was arguably the study’s most important test. A model can memorize the quirks of its home institution’s patient population, equipment, and surgical practices, only to falter elsewhere. Yet when applied to the 140 previously unseen patients from Tianjin Chest Hospital, the XGBoost model maintained an AUC of 0.800 (95 percent confidence interval, 0.656 to 0.913), with a sensitivity of 0.700, a specificity of 0.833, and a Brier score of 0.133. In practical terms, the model correctly flagged seven out of ten patients who went on to develop prolonged air leak while correctly reassuring more than eight out of ten of those who did not — performance that, if replicated in routine practice, would give surgeons a meaningful window to intervene preemptively.

Interpretability was treated not as an afterthought but as a core design requirement. The researchers applied SHAP — Shapley Additive Explanations — a technique borrowed from cooperative game theory that assigns each feature a contribution value for every individual prediction. Rather than presenting clinicians with an opaque probability, SHAP reveals how much each of the four variables pushed a given patient’s risk up or down, and in which direction. Through this analysis, VE/VCO₂ emerged as the single most influential feature, followed by the other three predictors ranked by correlation. The finding is physiologically plausible: patients whose lungs ventilate inefficiently under the stress of exercise are precisely those whose lung tissue is most fragile and least likely to seal spontaneously after surgical manipulation. A history of emphysema, a heavy smoking history, and prior chest surgery or trauma each independently compound this tissue vulnerability.

The clinical utility of the work extends beyond prediction to prevention. Armed with a preoperative risk estimate, surgical teams can tailor their approach for high-risk patients. Options include more meticulous staple-line reinforcement, the use of sealants or pledgeted sutures on fragile lung tissue, choosing segmentectomy over lobectomy where oncologically appropriate, planning for ambulatory drainage systems on discharge, or scheduling earlier follow-up to catch and manage leaks before they become complicated. To lower the barrier to bedside use, the team built a web-based risk calculator that allows clinicians to enter a patient’s four variables and receive an individualized probability of prolonged air leak — no programming expertise required.

The significance of incorporating cardiopulmonary exercise testing into a surgical prediction model deserves particular emphasis. Traditional preoperative pulmonary assessment in lung cancer surgery has relied on static spirometric measures such as forced expiratory volume in one second, or FEV1, and the diffusing capacity for carbon monoxide, DLCO. These tests measure lung function at rest and provide only a partial picture of cardiopulmonary reserve. CPET, by contrast, captures the integrated response of multiple organ systems under physiological stress, and the VE/VCO₂ slope in particular has gained recognition as a sensitive marker of ventilatory inefficiency in chronic obstructive pulmonary disease and heart failure. The present study suggests that this dynamic measure carries prognostic information for surgical outcomes that static tests do not fully capture — a finding with implications well beyond air leak prediction.

Several caveats frame the work appropriately. The study population was limited to Chinese patients aged 60 and above undergoing video-assisted thoracoscopic resection, and generalization to younger patients, other surgical techniques such as robotic-assisted resection, or different health care systems remains to be demonstrated. The external validation cohort, while methodologically essential, comprised 140 patients, and wider multicenter prospective trials would strengthen confidence in the model’s performance. The published article is an accepted, peer-reviewed author manuscript shared early ahead of the final version of record. The study was funded by the Beijing Kechuang Medical Development Foundation, with the funder playing no role in study design, analysis, or manuscript preparation, and the authors declare no competing interests.

Even so, the study stands as a compelling example of where predictive medicine is heading. By pairing a rigorous dual-centre validation strategy with explainable artificial intelligence and a physiologically grounded predictor, the researchers have moved machine learning in thoracic surgery closer to genuine clinical utility rather than proof-of-concept novelty. For the growing population of older patients facing lung cancer surgery, a simple preoperative test and a few clicks in a web calculator may soon determine who needs extra surgical vigilance — and spare many of them the days of tethered chest drainage that have long been an accepted, if unwelcome, part of recovery.