An interpretable machine learning model for predicting visceral pleural invasion in cT1 lung adenocarcinoma based on habitat analysis.

Background: Visceral pleural invasion (VPI) is a critical prognostic factor in early-stage lung adenocarcinoma, influencing T staging and treatment decisions, yet its preoperative assessment remains challenging due to reliance on postoperative pathological confirmation. This study evaluates interpretable machine learning models to predict VPI in solid cT1 pulmonary nodules using preoperative computed tomography (CT)-based habitat analysis, offering a non-invasive tool to optimize surgical planning. Methods: This retrospective study analyzed data of 802 patient with invasive pulmonary adenocarcinoma and solid lung nodules ≤3 cm from two centers and a public database. CT scans were preprocessed and regions were delineated for feature extraction. Habitat analysis using K-means clustering identified tumor heterogeneity. Radiomic features were extracted and modeled using light gradient boosting machine (LightGBM), with interpretability via SHapley Additive exPlanations (SHAP). Statistical analyses included logistic regression for predictive modeling, Kaplan-Meier for survival assessment, and evaluation of area under the curve (AUC), calibration, and decision curve analysis (DCA) for model performance. Results: The intra-tumoral region and the P5 region, which encompasses both the tumor and peri-tumor areas, were analyzed for habitat characterization. K-means unsupervised clustering was employed to generate habitat subregions. From each subregion, 1,834 radiomic features were extracted. The P5-model, demonstrated superior VPI prediction with an AUC of 0.787 [95% confidence interval (CI): 0.728-0.846] in the external test cohort (ETC), significantly outperforming intra-model (P=0.008). Calibration curves confirmed prediction consistency, and DCA showed a higher net benefit across thresholds. Kaplan-Meier survival analysis stratified patients into high- and low-risk groups effectively (P=0.042). SHAP analysis highlighted the significant role of peritumoral features, providing both global and local interpretability. Conclusions: The model demonstrates robust predictive accuracy for VPI status and effective prognosis stratification in cT1 stage lung adenocarcinoma patients, providing a non-invasive technique that significantly aids surgeons in preoperative planning and enhances clinical decision-making and patient outcomes.