e20053 Background: Recurrence risk after curative-intent surgery in early-stage (I–II) NSCLC remains heterogeneous. We developed and externally validated a machine-learning model to improve recurrence risk stratification for personalized postoperative management. Methods: We retrospectively evaluated a cohort of 724 patients with stage I–II NSCLC who underwent curative-intent resection. Recurrence was defined as radiologic and/or pathologic evidence of relapse, and a total of 52 patients experienced disease recurrence. A Random Forest (RF) classifier was trained using 11 routinely collected variables and benchmarked against Logistic Regression (LR) within a nested cross-validation pipeline to prevent data leakage. MICE imputation was performed strictly within the cross-validation folds. The independent external cohort (n = 50) was intentionally enriched (25 recurrence/25 no recurrence) to enable stable discrimination testing. Clinical utility was assessed via Decision Curve Analysis (DCA), and interpretability was established using SHapley Additive exPlanations (SHAP). Results: Median follow-up was 4.2 years. The Random Forest model outperformed Logistic Regression in both internal (ROC-AUC 0.704 vs 0.64) and external validation (ROC-AUC 0.71 vs 0.57). In internal out-of-fold analysis, the RF model achieved a sensitivity of 73.1% and specificity of 63.5% at a Youden-optimal threshold of 0.41. External validation demonstrated a consistent ROC-AUC of 0.71 (95% CI: 0.56–0.85). SHAP analysis identified tumor size, FEV1/FVC ratio, STAS, T-stage, and necrosis as the most influential predictors of recurrence. DCA showed superior net clinical benefit compared to "treat-all" or "treat-none" strategies. Conclusions: Our interpretable Random Forest model demonstrates stable discrimination and external validation using standard clinical data. This approach may support risk-adapted surveillance and guide postoperative decision-making, warranting prospective multicenter validation.
Kemik et al. (Thu,) studied this question.