Background/Objectives: Preoperative pulmonary function tests (PFTs) contain numerous physiologic parameters, yet surgeons typically rely on forced expiratory volume in one second (FEV1) and diffusing capacity of the lung for carbon monoxide (DLCO) to assess surgical risk. This study aimed to evaluate whether artificial intelligence (AI) could utilize more PFT data to predict the occurrence of prolonged air leak (PAL) following lung resection. Methods: An optical character recognition (OCR) model was used to extract structured data from PFT reports. These data were combined with clinical and demographic features from our institutional Society of Thoracic Surgeons General Thoracic Surgery Database (STS-GTSD) between 2016 and 2023. A feature selection algorithm was used to select the most predictive features, and a neural network was trained and tested on an internal validation cohort to predict PAL. Model performance was compared to previously published models. Results: There were 410 patients undergoing lung resection who had PFTs successfully digitized by the OCR system. A total of 76 available PFT features were extracted per patient. The final AI model included 10 key input variables, including three PFTs and seven clinical variables. On validation, the model achieved a specificity of 73%, sensitivity of 60%, overall accuracy of 72%, and an area under the curve of 0.74. This performance exceeded most existing PAL prediction models. Conclusions: AI-driven models using structured PFT and clinical data can enhance prediction of prolonged air leak after lung resection and outperform conventional regression-based models. Further research may focus on external validation and integration into clinical workflows.
Zahra et al. (2026) studied this question.