An Interpretable Clinical-Radiomics Model for Risk Stratification of Postoperative Liver Metastasis in Colorectal Cancer Patients

Objective To evaluate the value of a Clinical-radiomics model incorporating Shapley Additive exPlanations (SHAP) interpretability analysis in predicting postoperative liver metastasis in patients with colorectal cancer (CRC). Methods A retrospective analysis was conducted on 270 CRC patients who underwent surgical resection and had complete clinical, pathological, and imaging data, and patients were randomly divided into the training set and the test set in a ratio of 7:3. Univariate and multivariate logistic regression analyses were performed to identify independent clinical predictors of postoperative liver metastasis in CRC patients. Preoperative enhanced CT venous phase images were used for radiomics analysis, and key radiomics features were selected through pearson correlation, univariate and least absolute shrinkage and selection operator (LASSO) regression analyses. Predictive models were developed through five machine learning algorithms: Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), Stochastic Gradient Descent (SGD), and XGBoost. Model performance was assessed by the area under the ROC curve (AUC), calibration curves, and decision curve analysis (DCA). SHAP analysis was applied to the optimal model to enhance interpretability by quantifying the contribution of each feature to the prediction of postoperative liver metastasis. Results The RF-based clinical-radiomics model demonstrated superior predictive performance with an AUC of 0.921 (95% CI: 0.884–0.958) in the training set and 0.880 (95% CI: 0.803–0.956) in the test set. Meanwhile, the SHAP interpretability analysis yielded a ranking of feature importance for predicting postoperative liver metastasis, accompanied by the corresponding SHAP values for each patient. Conclusion The RF-based clinical-radiomics model is effective in predicting postoperative liver metastasis in CRC patients.SHAP analysis further enhances the model’s interpretability by clarifying the contribution of individual features to predictions.