Abstract / Summary
Abstract Background Patients with stage T1 non‑small cell lung cancer (NSCLC) still face a non‑negligible risk of early recurrence after curative surgery. Spread through air spaces (STAS) is a valuable prognostic indicator but can only be confirmed postoperatively. Conventional radiomics often ignores intratumoral spatial heterogeneity and the peritumoral microenvironment, limiting predictive accuracy. This study aimed to develop a multimodal machine‑learning model that integrates habitat imaging, multiscale peritumoral radiomics, and STAS status to improve recurrence prediction. Methods We retrospectively enrolled 216 stage T1 NSCLC patients (training : validation = 7 : 3). Intratumoral habitats were generated by K‑means clustering optimized with the Calinski‑Harabasz criterion. Radiomic features were extracted from the whole tumor, five habitat subregions, and 1‑mm, 3‑mm, 5‑mm peritumoral zones. Three classifiers (Random Forest, Extra Trees, XGBoost) were built. Feature selection used t‑tests, Pearson correlation, mRMR, and LASSO with 10‑fold cross‑validation. Independent risk factors were identified by multivariate logistic regression. Model performance was evaluated by ROC curves, DeLong tests, calibration curves, and decision curve analysis. Results Vacuole sign, gender, STAS status, pleural traction, and maximum diameter were independent predictors. The combined model (clinical + habitat + 3‑mm peritumoral features) achieved the highest AUCs: 0.910 (95% CI: 0.8484–0.9711) in training and 0.865 (95% CI: 0.7772–0.9532) in validation. It showed good calibration and superior net clinical benefit over single‑modality models. DeLong tests confirmed significant improvements over peritumoral‑only models in validation (P < 0.05). Conclusions The proposed noninvasive multimodal model accurately predicts early postoperative recurrence in stage T1 NSCLC and may assist in individualized risk stratification and adjuvant therapy decisions, offering a practical imaging‑informatics tool for early‑stage lung cancer management.