Development and validation of a machine learning model to predict early recurrence after surgery in NSCLC patients

To develop and validate a machine learning (ML) model for predicting early recurrence (ER) within two years post-surgery in non-small cell lung cancer (NSCLC) patients. This multicenter cohort study included 3, 171 NSCLC patients who underwent radical surgery between January 2015 and January 2021. Patients were randomly allocated to training and testing cohorts, with nine machine learning algorithms employed to construct prediction models for early recurrence, and a stacking method utilized to combine the three best-performing models. Furthermore, external validation was performed from a single institution (n = 619). Model performance was evaluated using various metrics, including Area Under the Receiver Operating Characteristic Curve (AUC). SHapley Additive exPlanations (SHAP) methodology was used to interpret predictions. Among the cohort, 553 patients (17. 4%) experienced ER, with common recurrence sites including the lung (35. 2%), brain (24. 1%), and bone (18. 5%). Significant predictors identified included pathological T stage (pT3-4: 77% vs. 50%, p < 0. 001), pathological N stage (pN1-2: 51% vs. 31%, p < 0. 001), and tumor differentiation grade (poorly differentiated: 68% vs. 54%, p < 0. 001). The stacking model achieved superior predictive performance (AUC = 0. 81, accuracy = 0. 83, Brier score = 0. 03), outperforming individual ML models, whose AUC values ranged from 0. 72 to 0. 79. SHAP analysis revealed pT stage, maximum tumor diameter, and tumor markers as key determinants of ER risk. An online computing platform (https: //nsclc-risk. shinyapps. io/NSCLCₑarlyᵣecurrence/) for this stacking model is publicly available and free-to-use by doctors and patients. This study developed an interpretable ML model with high predictive performance for ER in post-operative NSCLC patients. The model, based on readily available clinical data, offers a valuable tool for personalized treatment decisions and follow-up strategies. Future prospective studies across multiple centers are needed to further validate the model's generalizability and accuracy.