Development and internal validation of a laboratory test-based model for predicting recurrence in patients with lung adenocarcinoma treated with third-generation EGFR-TKIs

We aimed to develop and internally validate a laboratory-test-based model for predicting recurrence risk in patients with lung adenocarcinoma receiving third-generation epidermal growth factor receptor tyrosine kinase inhibitor (EGFR-TKI) therapy. This retrospective single-center study included 320 patients. The cohort was stratified by recurrence status and randomly divided into a training set (70%, n = 224) and an independent held-out test set (30%, n = 96). Ninety-three candidate predictors were considered. Random-forest-based feature screening was repeated within the training portion of each stratified 10-fold cross-validation fold to reduce information leakage. Six statistical and machine-learning algorithms, including logistic regression, LASSO regression, random forest, gradient boosting machine, support vector machine, and XGBoost, were compared. The model with the highest mean cross-validated area under the receiver operating characteristic curve (AUC) was selected and subsequently evaluated in the independent test set. Model performance was assessed using discrimination, classification metrics, calibration, decision curve analysis, and full-pipeline bootstrap validation. Individual laboratory and exploratory immune-cell indicators showed limited discriminatory performance for recurrence when evaluated separately. Among the six candidate models, XGBoost achieved the highest mean AUC during stratified 10-fold cross-validation in the training set (0.700) and was therefore selected as the final model. After refitting the complete training set, random-forest screening retained 30 predictors. In the independent held-out test set, XGBoost achieved an AUC of 0.812 (95% CI, 0.727–0.896), with a sensitivity of 0.790, specificity of 0.647, accuracy of 0.740, positive predictive value of 0.803, negative predictive value of 0.629, and Brier score of 0.183. Full-pipeline bootstrap validation yielded an optimism-corrected AUC of 0.760. Calibration analysis showed some deviation from ideal calibration, while decision curve analysis indicated potential clinical net benefit across a broad range of threshold probabilities. A random-forest-screened XGBoost model integrating routinely available clinical and laboratory variables demonstrated good discrimination for recurrence-risk stratification in patients with EGFR-mutant lung adenocarcinoma receiving third-generation EGFR-TKI therapy. The model may provide complementary information for recurrence risk stratification, but external validation and prospective evaluation are required before clinical implementation.