Machine learning-based survival analysis and risk stratification for patients with cholangiocarcinoma

<title>Abstract</title> <p>Background/Aims Cholangiocarcinoma (CCA) has heterogeneous biology and poor survival, while existing prognostic tools show limited discrimination. We aimed to develop and internally validate interpretable machine-learning (ML) survival models to improve risk stratification in treatment-naïve patients with CCA. Methods We retrospectively assembled a single-center cohort from Taipei Veterans General Hospital of 693 consecutive treatment-naïve patients diagnosed with CCA between 2014 and 2023. The primary outcome was overall survival. Predictors included demographics, cancer stage, treatment modality, surgical resection, and serum biomarkers. Preprocessing used log-transformation, imputation, and encoding. Patients were split into training-validation sets as 70:30, and five-fold cross-validation tuned models. We compared Cox proportional hazards with random survival forest (RSF), XGBoost, survival decision tree, and logistic regression. Discrimination was assessed by Harrell’s C-index and time-dependent AUROC, and interpretability used permutation importance and Shapley Additive Explanations (SHAP). Results Multivariable Cox analysis identified nine independent mortality predictors: cancer stage, tumor location, treatment status, absence of surgical resection, International Normalized Ratio (INR), carbohydrate antigen 19-9 (CA19-9), gamma-glutamyl transferase (GGT), ALBI grade, and FIB-4. RSF outperformed other comparators (validation C-index 0.768; AUROC 0.815). RSF-derived risk tertiles produced well-separated survival (log-rank P&amp;lt;0.001): 3-year survival was 54%, 19%, and 3%, in low-, intermediate-, and high-risk groups. Variable importance and SHAP consistently prioritized surgery, cancer stage, CA19-9, ALBI, and treatment status. Conclusions In a real-world, treatment-naïve cohort of CCA, RSF provided accurate, interpretable risk prediction and clinically meaningful stratification, outperforming Cox and other ML approaches. External validation is warranted before clinical deployment.</p>