Collaborative assessment of the risk of postoperative progression in early-stage non-small cell lung cancer: a robust federated learning model.

Background: While the TNM staging system provides valuable insights into the extent of disease, predicting postoperative progression in early-stage non-small cell lung cancer (NSCLC) remains a significant challenge. An effective bioimaging prognostic marker for early-stage NSCLC, powered by artificial intelligence, could greatly assist clinicians in making informed treatment decisions.

Methods: A total of 926 patients from four centers (A, B, C, and D) with histologically confirmed stage I or II solid non-small cell lung cancer (NSCLC) who underwent surgical resection were retrospectively reviewed. In this study, we propose a robust federated learning model (RFed) designed to predict the risk of postoperative progression in early-stage NSCLC patients. The diagnostic efficiency of the RFed model was evaluated using the area under the curve (AUC) and Decision Curve Analysis (DCA). Additionally, the model's performance was further validated through Kaplan-Meier survival analysis, with statistical significance assessed using the log-rank test. Finally, the robustness, generalizability, and interpretability of the RFed model were comprehensively evaluated to confirm its clinical applicability.

Results: Experimental results demonstrated the superior performance of the RFed model. Specifically, RFed achieved AUC values of 0.936, 0.861, 0.925, and 0.970 on the test sets from the four centers. DCA further revealed that RFed provided a greater net benefit compared to the clinical model across a threshold probability range of 0.02 to 0.99. Moreover, Kaplan-Meier curves showed improved discrimination between high-risk and low-risk groups when compared to other models, highlighting its enhanced predictive capability.

Conclusions: The RFed model demonstrates significant effectiveness in predicting the risk of postoperative progression in early-stage NSCLC patients. Its clinical application value lies in its potential to enhance stratified management and support the development of precise treatment strategies for this patient population.