Prediction of postoperative recurrence in non–small cell lung cancer using machine learning approaches.

e20020 Background: Early-stage NSCLC can have favorable survival outcomes; however, postoperative recurrence remains common and clinically challenging. This study aimed to develop machine learning (ML) models to predict recurrence after curative resection using routinely available clinicopathologic variables, to test the stability of model performance across different feature-selection strategies, and to interpret ML findings alongside conventional Cox regression results. Methods: We retrospectively analyzed 265 patients who underwent curative-intent surgical resection at Ankara Bilkent City Hospital (January 2020–June 2025). The primary endpoint was recurrence (yes/no). Seventeen clinical, pathological, and treatment-related variables were evaluated. Multiple supervised ML classifiers were trained using (i) the full feature set and (ii) reduced feature subsets generated by ANOVA, chi-square, and Kruskal–Wallis–based selection. Performance was assessed using accuracy, AUC, F1 score, and confusion matrices. Prognostic associations were examined using univariate and multivariate Cox regression, and model interpretability was explored using feature importance and SHAP analyses. Results: Recurrence occurred in 82/265 (30.9%) patients. Using all variables, AdaBoost achieved the highest accuracy (≈0.79), with SVC-RBF and several neural network/Naïve Bayes models showing comparable accuracy (≈0.77). Across ANOVA-, chi-square–, and Kruskal–Wallis–selected feature sets, overall performance remained largely consistent, and no single selection method was uniformly superior. Discriminative performance was highest for SVC-RBF (AUC ≈0.81), whereas AdaBoost yielded the best balance between precision and recall (F1 ≈0.87). Across models, key predictors repeatedly included histologic subtype, tumor size, tumor grade, and ECOG performance status. SHAP analysis identified ECOG performance status and tumor size as the dominant contributors to individual recurrence predictions. Several of these factors were also supported as prognostic variables in Cox regression analyses. Conclusions: In conclusion, ML models can reliably predict postoperative recurrence in resected NSCLC using standard clinicopathologic data. Ensemble (AdaBoost) and kernel-based (SVC-RBF) approaches showed the most favorable and stable performance, offering complementary strengths that may support individualized postoperative risk assessment.