Artificial intelligence-assisted risk prediction of postoperative pulmonary complications in non-small cell lung cancer surgery.

BACKGROUND

Pulmonary complications are the most frequent adverse events following surgery for non-small cell lung cancer (NSCLC), influencing both short-term clinical outcomes and long-term prognosis. This study aimed to develop and evaluate artificial intelligence (AI) models to predict postoperative pulmonary complications in patients undergoing surgical treatment for NSCLC.

MATERIALS AND

METHODS

A total of 953 patients who underwent lung resection and mediastinal lymph node dissection for NSCLC between 2001 and 2023 were retrospectively analyzed. Clinical, laboratory, respiratory function, tumor-related radiological, surgical, and pathological parameters served as input variables, while the occurrence of postoperative pulmonary complications constituted the output variable. A fully connected deep neural network was employed, using 10-fold cross-validation. Model performance was evaluated via 10-fold cross-validation using specificity, sensitivity, NPV, PPV, accuracy, and F1 score. Additionally, area under the receiver operating characteristic (ROC) curve (AUC) and Decision Curve Analysis (DCA) were utilized to assess discriminative ability and clinical net benefit, respectively.

RESULTS

The model achieved an accuracy of 88.6% and an average F1 score of 84.4% on the training dataset. In the test dataset, the model demonstrated robust performance with an accuracy of 90.4%, an average F1 score of 86.4%, and an area under the receiver operating characteristic curve (AUC) of 0.84. These results indicate high discriminative power and reliability in predicting postoperative pulmonary complications.

CONCLUSIONS

Accurate prediction of postoperative pulmonary complications in NSCLC surgery is crucial for optimizing perioperative care and reducing morbidity. The proposed deep learning model demonstrates promising predictive performance, enabling stratification of patients into high- and low-risk groups, and may serve as a valuable decision-support tool for clinicians.