Objective To investigate the utility of a computational pathology-based tumor microenvironment (TME) score derived from whole slide images (WSIs) in predicting the efficacy of epidermal growth factor receptor tyrosine kinase inhibitors (EGFR-TKIs) in patients with EGFR mutation-positive non-small cell lung cancer (NSCLC).

Methods This retrospective study collected 240 EGFR-mutant NSCLC patients treated with EGFR-TKIs at The First Affiliated Hospital of Wannan Medical College and analyzed hematoxylin-eosin (H&E)-stained WSIs of biopsy specimens, along with clinical and imaging data. The patients were randomly assigned into a training cohort (n=160) and an independent validation cohort (n=80) in a 2:1 ratio. Treatment response was assessed based on CT findings at 3 months after EGFR-TKIs initiation. Computational pathology was employed to automatically quantify the proportions of four TME components (tumor epithelium, stroma, lymphocytes, and vasculature) within the tumor regions of WSIs. Multivariate Logistic regression in the training cohort identified TME components independently predictive of treatment response (P<0.05), which were then integrated into a TME-score. The predictive performance was evaluated using receiver operating characteristic (ROC) curve analysis and area under the curve (AUC). The TME-score model was compared with a clinical-feature-based model and a combined model (TME-score+clinical features). Finally, the models were validated in the independent cohort.

Results In the training cohort, the TME-score, incorporating epithelial and stromal proportions, achieved an AUC of 0.827 (95%CI: 0.749–0.892) for predicting treatment response, while the validation cohort yielded an AUC of 0.845 (95%CI: 0.735–0.937). Both outperformed the clinical model (AUCs=0.730 95%CI: 0.645–0.804 and 0.712 95%CI: 0.586–0.824, respectively). The combined model (TME-score+clinical features, including cytokeratin 19 fragment and non-contrast CT values) further improved predictive performance (AUCs=0.884 95%CI: 0.827–0.932 and 0.882 95%CI: 0.798–0.950, respectively). Delong’s test for pairwise model comparisons showed significant differences (all P<0.05) except TME-score and the combined model in the validation cohort (P=0.289).

Conclusions TME-score outperformed clinical models in predicting EGFR-TKIs efficacy in EGFR mutation-positive NSCLC patients and may serve as a novel tool for identifying patients likely to benefit from targeted therapy.