Risk stratification for early-stage NSCLC progression: a federated learning framework with large-small model synergy.

[PURPOSE] Accurate prediction of non-small cell lung cancer (NSCLC) progression is crucial for guiding clinical decision-making and alleviating patients' psychological burden. Overtreatment risks unnecessary trauma for low-risk patients, while undertreatment may delay timely intervention for high-risk cases, worsening prognosis. This study aimed to develop a precise risk stratification system. The aim of this study is to develop an accurate risk stratification system to improve prediction accuracy and stratified management. To validate the proposed framework's versatility and robustness, we conducted multiple cross-task validation experiments.

[MATERIALS AND METHODS] This retrospective study included 926 patients with resected stage I-IIA NSCLC who underwent radical surgery at four centers between January 2014 and September 2019. A multi-center intelligent risk stratification model FesCPI (Federated cross-scale Common-Personal-Interactive learning) was developed. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Model stability was assessed with five-fold cross-validation, and clinical utility was further validated through decision curve analysis (DCA). Additionally, cross-task experiments for gastric cancer and endometrial cancer were conducted to validate the model's performance.

[RESULTS] FedCPI consistently outperformed clinical stratification ((DeLong test, p < 0.05)) and federated learning baselines across multicenter tasks. In early-stage NSCLC, it achieved AUCs up to 0.9255 and ACCs up to 0.8909, with 4.23-15.95% gains over competing models. The framework continues to demonstrate outstanding performance in various tasks, including predicting gastric cancer recurrence and endometrial cancer infiltration, thereby validating the effectiveness of this methodology. Feature analyses confirmed complementary roles of VFMs and ResNet18, while ablation studies showed that both Large-Small Model Feature Decomposition and Fusion (LMSF) and Federated Adaptive Communication Mechanism (FACM) were indispensable for optimal performance.

[CONCLUSION] Our findings suggest that the DL(deep learning)-based FedCPI framework provides a non-invasive, accurate, and reliable tool for early-stage lung cancer risk stratification. Furthermore, the methodology demonstrated excellent performance in independent validation experiments for clinical tasks involving gastric cancer and endometrial cancer. By improving diagnostic precision, this approach has the potential to optimize clinical decision-making and reduce the burden of overtreatment for patients.