Interpretable Prognostic Modeling for Postoperative Pancreatic Cancer Using Multi-machine Learning and Habitat Radiomics: A Multi-center Study.

Rationale And Objectives: Accurate risk stratification is critical for guiding personalized treatment in resectable pancreatic cancer (RPC). This retrospective study assessed the utility of habitat radiomics for predicting recurrence-free survival (RFS) in RPC patients.

Materials And Methods: A total of 455 RPC patients were divided into training and external test sets from January 2018 to July 2024. Tumors were segmented into subregions using habitat radiomics to capture localized heterogeneity. Seven machine learning models, including random survival forest (RSF), were compared using Harrell's C-index. The optimal model underwent further validation through time-dependent ROC and Kaplan-Meier (KM) analyses. Shapley additive explanations (SHAP) and survival local interpretable model-agnostic explanations (SurvLIME) were applied to enhance model interpretability.

Results: The RSF model based on habitat radiomics achieved a C-index of 0.828 in the training cohort and 0.702, 0.680 in external test sets, outperforming whole-tumor radiomics (p<0.05). Time-dependent ROC analysis showed AUCs of 0.71, 0.83, and 0.79 at 0.5, 1, and 2 years in the first test set, and 0.65, 0.79, and 0.75 in the second test set. KM analysis revealed that the predicted low-risk groups had significantly longer RFS compared to the predicted high-risk groups in both external test sets (all p<0.05). Interpretability analysis identified key variables, including Feature 1, Feature 5, Feature 2, and Feature 4 from Habitat Subregion 1, and Feature 3 from Habitat Subregion 3.

Conclusion: The habitat radiomics RSF machine learning model improves prognostic accuracy and interpretability for postoperative RPC, providing a promising tool for personalized management.