Development and external validation of an interpretable multimodal deep learning model for 5-year mortality in high-risk stage ii colorectal cancer.

Abstract

Purpose: High-risk stage II colorectal cancer (CRC) shows heterogeneous outcomes despite adjuvant chemotherapy. We developed and validated an interpretable multimodal deep learning model integrating clinical data, serum biomarkers, and venous-phase CT to predict 5-year CRC-specific mortality in high-risk stage II CRC.

Methods: This retrospective, multicenter cohort included 778 high-risk stage II CRC patients from three centers, all treated with adjuvant chemotherapy and with complete preoperative clinical, biomarker, and venous-phase CT data. Patients were split into a development cohort (Centers A + B, n = 720) and an external testing cohort (Center C, n = 58). A multimodal model combining numerical (clinical + biomarker) and imaging (CT) inputs was developed and internally validated using tenfold cross-validation in the development cohort and evaluated in the external cohort. Interpretability was assessed using SHAP and Grad-CAM.

Results: In the development cohort, the multimodal model showed superior discrimination (AUC 0.89; 95% CI, 0.87-0.91) versus numerical-only (AUC 0.76) and imaging-only (AUC 0.69). In the external testing cohort (9/58 CRC-specific deaths), the multimodal model achieved an AUC of 0.88 (95% CI, 0.76-0.96). SHAP and Grad-CAM consistently highlighted age, CA125, and tumor regions on CT as key contributors.

Conclusion: This interpretable multimodal approach, using routine clinical, biomarker, and CT data, improves 5-year mortality risk stratification in high-risk stage II CRC and may inform risk-adapted surveillance and clinical decision support; prospective validation is warranted before treatment modification.