Interpretable machine learning model integrating delta-radiomics enhances postoperative recurrence prediction in...

Despite curative resection for early-stage lung adenocarcinoma (LUAD), postoperative recurrence remains a significant concern, with substantial outcome heterogeneity even among patients with stage IA disease. An accurate tool to identify individuals at high risk of early postoperative recurrence after curative resection is urgently needed to guide intensified surveillance and adjuvant strategies. Therefore, the objective of this study was to develop and validate an interpretable machine learning model integrating clinical-radiological features, radiomics, and delta-radiomics to predict early postoperative recurrence in stage IA LUAD.

This retrospective study included 463 patients with pathologically confirmed stage IA LUAD who underwent curative surgery. Preoperative serial computed tomography (CT) scans (baseline and follow-up) were used to extract radiomics and delta-radiomics features, the latter quantifying temporal changes in tumor phenotype. Clinical and conventional radiological variables were also collected. After random allocation to training and validation cohorts, the Synthetic Minority Oversampling Technique (SMOTE) was applied to address class imbalance. Feature selection was performed using univariate analysis, correlation analysis, and Random Forest with five-fold cross-validation. A fusion model integrating clinical-radiological features, radiomics score (Rad-score), and Delta Rad-score was developed using a Random Forest algorithm. Model performance was assessed by area under the curve (AUC), calibration, and decision curve analysis. Shapley additive explanations (SHAP) were used for model interpretability. Prognostic value for recurrence-free survival (RFS) and overall survival (OS) was evaluated using Kaplan-Meier analysis and compared with traditional tumor-node-metastasis (TNM) staging.

Among 463 patients (median age 57 years; 156 males), 23 (5.0%) experienced early recurrence. The fusion model demonstrated excellent discrimination in the validation cohort (AUC: 0.881), significantly outperforming the clinical model (AUC: 0.725), with comparable discrimination to the radiomics model (AUC: 0.845) and delta-radiomics model (AUC: 0.827). Calibration was good (Brier score: 0.021), and decision curve analysis confirmed the highest net clinical benefit for the fusion model. SHAP analysis identified Rad-score and Delta Rad-score as the most important predictors, with higher values associated with increased recurrence risk. The fusion model stratified patients into high- and low-risk groups with significant differences in both RFS and OS (log-rank P<0.001 for both), whereas TNM staging (T1c vs. T1a/T1b) was associated with RFS (P=0.02) but not OS (P=0.32).

The interpretable machine learning model integrating clinical-radiological features, radiomics, and delta-radiomics demonstrated promising predictive performance in internal validation of early postoperative recurrence in stage IA LUAD compared with traditional models and TNM staging. Delta-radiomics captures dynamic tumor evolution and emerges as a key prognostic biomarker. This model offers a non-invasive tool for individualized risk stratification to guide postoperative management.