From predictive model to clinical application: interpretable prediction of metachronous liver metastasis in colorectal cancer

Abstract

Metachronous liver metastasis (MLM) after curative surgery is a major determinant of poor prognosis in colorectal cancer (CRC), yet existing approaches based on clinicopathological indicators or imaging alone have limited predictive performance and clinical applicability. This retrospective multicenter study included 2983 CRC patients from four Chinese medical centers, including 161 MLM cases and 2822 non-MLM cases with definite 5-year postoperative outcome ascertainment. We developed CMLM, an interpretable multimodal prediction framework integrating preoperative venous-phase CT images, clinicopathological variables, and perioperative laboratory features through multi-scale image extraction, self-attention numerical modeling, and cross-attention fusion. Tumor ROIs were extracted using MA-YOLO with radiologist quality control, and numerical features were selected within the training cohort. CMLM achieved strong performance, with an AUROC of 0.95 and an AUPRC of 0.81 in internal validation, and AUROCs of 0.93, 0.93, and 0.94 in three locked external cohorts. The model outperformed unimodal models, showed reasonable calibration, provided clinical net benefit, and stratified patients into prognostically distinct risk groups. SHAP and Grad-CAM supported model interpretability. Deployed as an online application supporting complete and missing-modality inputs, CMLM may facilitate individualized postoperative surveillance and risk-adapted management for CRC patients.