Development and temporal validation of nomograms and machine learning models for predicting overall and cancer-specific survival in primary distal cholangiocarcinoma: a SEER-based study.

Background: Distal cholangiocarcinoma (dCCA) differs from intrahepatic and perihilar cholangiocarcinoma in anatomical origin, surgical management, and molecular profile, yet most existing prognostic tools either pool all biliary tract cancers together or rely on anatomical Tumor Node Metastasis (TNM) staging alone. A dCCA-specific model integrating clinical, treatment, and tumor-burden variables is therefore needed. We aimed to develop and evaluate nomograms and machine learning models for predicting overall survival (OS) and cancer-specific survival (CSS) in patients with primary dCCA. Methods: We identified 1,515 patients with primary dCCA diagnosed between 2010 and 2017 in the Surveillance, Epidemiology, and End Results (SEER) database. Patients diagnosed in 2010-2015 were randomly divided into a training cohort (70%, n=773) and an internal testing cohort (30%, n=331); patients diagnosed in 2016-2017 (n=411) formed a temporal validation cohort with no calendar-year overlap. Independent predictors were selected by univariable and multivariable Cox regression for OS and by Fine-Gray subdistribution hazard modeling for CSS, and were used to construct nomograms predicting 1-, 3-, and 5-year survival. Four machine learning survival models-decision tree, random survival forest, gradient boosting survival analysis, and an XGBoost-based accelerated failure time model (XGBAFT)-were trained for OS prediction and interpreted with SHapley Additive exPlanations (SHAP). Performance was assessed by Harrell's concordance index (C-index), time-dependent area under the curve (AUC), calibration plots, and decision curve analysis, and benchmarked against the eighth-edition American Joint Committee on Cancer (AJCC) TNM staging system. Results: The development cohort was predominantly male (53.2%) and aged over 65 years (61.6%); 28.9% underwent radical resection of the primary site and 47.6% received chemotherapy, with a median observed survival of 8 months. For OS, the nomogram achieved a C-index of 0.754 and 1-, 3-, and 5-year AUCs of 0.854, 0.868, and 0.871 in the internal testing cohort, and a C-index of 0.742 with AUCs of 0.830, 0.836, and 0.843 in the temporal validation cohort. For CSS, the corresponding C-indices were 0.726 and 0.710, with AUCs of 0.813/0.797/0.760 and 0.792/0.712/0.702. Both nomograms outperformed the eighth-edition TNM system (OS C-index, 0.585; CSS C-index, 0.635) and showed good calibration and net clinical benefit. Among the machine learning models, XGBAFT achieved the highest discrimination (AUCs, 0.861/0.834/0.831 internally and 0.807/0.803/0.789 temporally). SHAP identified chemotherapy, primary-site radical surgery, regional lymph node surgery as the most influential predictors. The nomogram stratified patients into low-, intermediate-, and high-risk groups with significantly divergent survival curves (P<0.001). Conclusions: Within the SEER population, our dCCA-specific nomograms and the XGBAFT model provided accurate, well-calibrated, and interpretable predictions of OS and CSS, outperformed eighth-edition TNM staging, and supported effective risk stratification. Because both cohorts were drawn from SEER, these results constitute internal and temporal validation; prospective multicenter studies in geographically and ethnically diverse cohorts are required before clinical implementation.