Externally validated explainable machine learning for postoperative recurrence prediction in early-stage NSCLC

Creative Commons License

Kemik F., Goren H. K., Koylu B., Kikili C. I., Demir N., KARATAŞ B., ...Daha Fazla

FRONTIERS IN ONCOLOGY, cilt.16, 2026 (SCI-Expanded, Scopus)

Özet

Background Postoperative recurrence remains a major challenge in early-stage non-small cell lung cancer (NSCLC), and pathological TNM staging does not fully capture within-stage heterogeneity. We aimed to develop and externally validate an explainable machine learning model for recurrence prediction after curative resection.Methods This retrospective study included 723 patients with stage I-II NSCLC, including 52 recurrences, with independent external validation in a separate cohort (n=50). A Random Forest model using routinely available clinical and pathological variables was developed within a nested cross-validation framework and compared with logistic regression. Performance was evaluated using ROC-AUC, calibration, Decision Curve Analysis, and SHAP-based interpretation.Results The Random Forest achieved a mean internal ROC-AUC of 0.70 versus 0.64 for logistic regression, although no formal paired statistical comparison was performed. In an independent case-control external validation cohort with an artificially balanced outcome distribution, the Random Forest achieved an ROC-AUC of 0.68 (95% CI 0.53-0.83); the balanced design precluded assessment of calibration or absolute risk at natural prevalence. Sigmoid recalibration of pooled out-of-fold predictions yielded an apparent Brier score reduction to 0.063, although post-calibration performance was not independently evaluated. Decision Curve Analysis demonstrated positive net clinical benefit across clinically relevant thresholds. SHAP analysis highlighted tumor size, pathological T stage, and STAS among the principal tumor-related predictors, whereas fold-level analysis showed greater stability for tumor size, tumor necrosis, and STAS. Complementary time-to-event analyses accounting for right censoring included 715 patients and 44 recurrence events.Conclusion An explainable machine learning model based on routinely available variables demonstrated promising but preliminary discrimination in an independent case-control external validation cohort. Larger consecutive cohorts with natural outcome prevalence are required before clinical application.