Interpretable machine learning model integrating contrast-enhanced CT environmental radiomics and clinicopathological features for predicting postoperative recurrence in lung adenocarcinoma: a retrospective pilot study.

Purpose: This study aims to develop an interpretable predictive model combining contrast-enhanced CT (CECT) radiomics features with clinicopathological parameters to assess 3-year recurrence risk after surgery for lung adenocarcinoma (LA).

Methods: A retrospective cohort of 350 LA patients (126 recurrence, 224 non-recurrence) from Xiangyang NO.1 People's Hospital (2016-2023) was included. Radiomics features were extracted from arterial and venous phase CECT images using 3D Slicer's Radiomics plugin. Features with intraclass correlation coefficient (ICC > 0.75) were selected, followed by LASSO regression with cross-validation to generate radiomics scores (Radscore3 for intratumoral and Radscore4 for peritumoral regions). Clinical variables (sex, heterogeneous enhancement, pleural invasion, Ki67) were integrated via chi-square/t-test analysis. Ten machine learning algorithms (e.g., XGBoost, CatBoost, Random Forest) were trained on a stratified 7:3 split (training: n=245; testing: n=105) with five-fold cross-validation. Model performance was evaluated using ROC curves (AUC), calibration curves, decision curve analysis (DCA), and a nomogram.

Results: Univariate analysis identified sex (OR=1.66, p=0.02), heterogeneous enhancement (OR=4.32, p<0.05), visceral pleural invasion (OR=4.75, p<0.05), Radscore3 (OR=356.17, p<0.05), Radscore4 (OR=1529.16, p<0.05), and Ki67 (OR=1.09, p=0.01) as significant predictors. Among machine learning models, CatBoost achieved superior performance (AUC=0.883, 95% CI:0.811-0.955) compared to logistic regression (AUC=0.877, 95% CI:0.804-0.949) in test set. Calibration curves demonstrated high consistency between predicted and observed recurrence risks, while DCA indicated clinical utility at threshold probabilities >0.17. SHAP analysis highlighted heterogeneous enhancement, visceral pleural invasion, Radscore3/4, and Ki67 as key contributors. The nomogram integrated these factors, enhancing model interpretability and clinical applicability.

Conclusion: The CatBoost model integrating CECT environmental radiomics and clinicopathological parameters effectively predicts postoperative LA recurrence, supporting personalized adjuvant therapy decisions. Its interpretable framework emphasizes tumor heterogeneity (Radscore3/4) as a critical prognostic biomarker, providing mechanistic insights into LA recurrence.