A practical nomogram for predicting survival in patients with extrahepatic cholangiocarcinoma: a population-based cohort study

Introduction

Extrahepatic cholangiocarcinoma (ECC) represents one of the most aggressive malignancies of the biliary tract, characterized by diagnostic challenges, rapid disease progression, and dismal prognosis (1). As the second most common primary hepatobiliary malignancy, ECC has exhibited a steadily rising incidence worldwide, particularly among elderly populations. With a median age at diagnosis of approximately 70 years and slight male predominance, ECC often remains clinically silent during its early stages due to its longitudinal infiltrative growth pattern along the bile duct wall and deep anatomical location. Consequently, over 70% of patients present with locally advanced or metastatic disease at diagnosis, precluding curative surgical intervention. Even following radical resection, the 5-year survival rate remains disappointing at merely 20–30%, while patients with advanced disease face a median overall survival (OS) of only 6–9 months, underscoring the highly aggressive biological behavior of this malignancy (2-4).

Anatomically, ECC is classified into perihilar and distal subtypes. Perihilar tumors, arising at the confluence of the hepatic ducts, demonstrate early propensity for portal vein and hepatic parenchymal invasion, whereas distal tumors, located within the pancreatic portion of the common bile duct, behave similarly to pancreatic head carcinoma with early lymphatic dissemination. This anatomical complexity renders only 20–30% of patients candidates for surgical resection at presentation (5). For the majority with advanced disease, treatment relies on systemic chemotherapy (primarily gemcitabine-cisplatin regimens) combined with palliative biliary drainage to relieve obstructive jaundice, prevent cholangitis, and improve quality of life. However, selecting the optimal biliary drainage modality—between endoscopic retrograde cholangiopancreatography (ERCP) and percutaneous transhepatic biliary drainage (PTBD), or plastic versus metal stents—currently lacks objective individualized guidance and depends largely on physician preference. This empirical approach exposes patients with extremely limited life expectancy to unnecessary risks of invasive procedures, while those with better prognoses may be deprived of the durable patency benefits afforded by metal stents (6-10).

Accurate risk assessment is a prerequisite for individualized therapeutic decision-making. Currently, the American Joint Committee on Cancer (AJCC) 8th edition TNM (tumor-node-metastasis) staging system serves as the primary prognostic tool for ECC; however, it relies exclusively on anatomical tumor extent without considering treatment responsiveness or host factors, resulting in substantial survival heterogeneity among patients within identical stage categories. Recently, nomograms integrating multiple variables to predict individual survival probabilities have emerged as promising alternatives, demonstrating superior predictive accuracy over conventional staging in various solid malignancies. Nevertheless, existing prognostic models for ECC remain limited: most are based on single-center retrospective cohorts (typically <500 patients) with insufficient external validation and restricted generalizability. More importantly, previous models have focused predominantly on demographic and pathological characteristics while failing to incorporate critical therapeutic interventions—specifically surgery and systemic chemotherapy—which represent the strongest independent determinants of patient survival (11-13).

Therefore, our study aimed to develop a prognostic nomogram incorporating treatment data and validate it through split-sample internal validation using the Surveillance, Epidemiology, and End Results (SEER) database. Furthermore, we established a risk stratification system based on nomogram scores, categorizing patients into three distinct prognostic tiers: high-risk (<3 months), intermediate-risk (3–12 months), and low-risk (>1 year). This stratification system divides patients into discrete prognostic subgroups according to their predicted survival trajectories, providing reference for determining the overall intensity of palliative care allocation. Nevertheless, the selection of specific interventional strategies relies on comprehensive clinical judgment and institutional experience, with full consideration of patients’ performance status, anatomical characteristics, and regional medical resources. These key variables lie outside the analytical scope of the current registry-derived model. We present this article in accordance with the TRIPOD reporting checklist (available at https://jgo.amegroups.com/article/view/10.21037/jgo-2026-0258/rc).

Methods

Study population

Data were retrieved from the SEER program database (https://seer.cancer.gov/), specifically the SEERStat Database: Incidence-SEER Research Data, 17 Registries, Nov 2024 Sub (2000–2022), released by the National Cancer Institute in April 2025, by using the National Cancer Institute’s SEERStat software (version 9.0.42.0). SEER is a population-based cancer registry covering approximately 28% of the US population.

Patients diagnosed with ECC [site code C24.0, International Classification of Diseases for Oncology, Third Edition (ICD-O-3)] between 2018 and 2022 were included, and only malignant tumors (behavior code 3) and first primary cancers [sequence number 00 (one primary only) or 01 (first of two or more primaries)] were eligible for analysis (14). Given the advances in oncological treatment in recent years, only patients diagnosed within the most recent 5-year period were included to better reflect real-world clinical practice. Tumor staging data (TNM classification) were retrieved according to the relevant derived AJCC staging fields (8th edition) (14). Exclusion criteria included: (I) patients diagnosed by death certificate only or autopsy reports; (II) patients with incomplete clinicopathological data, including missing survival information, unknown T/N/M stage, or indeterminate treatment records. Finally, a total of 2,254 patients were included in the final analysis, and randomly divided into a training cohort (n=1,577) and a validation cohort (n=677) in a 7:3 ratio using stratified random sampling by vital status to ensure a balanced distribution of survival status between the two groups (Figure 1).

Figure 1 The flowchart of the study.

This study used publicly accessible de-identified data from the SEER Program. Research utilizing aggregated, de-identified SEER data is exempt from informed consent requirements. Consequently, neither additional informed consent nor institutional review board approval was required for this study. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was registered with the Research Registry (Unique Identifying Number: researchregistry11805).

Study variables

Demographic variables included age at diagnosis (continuous, years), sex (male versus female), and race (White, Black, Asian/Pacific Islander, or other). Tumor characteristics were staged according to the AJCC 8th edition (14). The T, N, and M categories were derived from the SEER “Derived EOD 2018” fields, which provide standardized AJCC 8th edition staging for cases diagnosed from 2018 onward. Overall summary stage (localized, regional, distant) was determined using the Combined Summary Stage system. Tumor location (perihilar versus distal bile duct) was classified using the AJCC ID (2018+). Treatment variables included surgical intervention, radiation, and systemic therapy. Surgical resection of the primary site was categorized into three groups based on “RX Summ-Surg Prim Site”: no surgery, partial resection (local excision or partial removal), and radical surgery (radical cholecystectomy, bile duct resection, or Whipple procedure). Additional surgery to regional or distant sites was indicated by “RX Summ-Surg Oth Reg/Dis” (yes versus no). Radiation therapy (yes versus no/unknown) and chemotherapy (yes versus no/unknown) were obtained from the respective recode fields. Systemic therapy sequence relative to surgery (preoperative, postoperative, or none) was determined using “RX Summ-Systemic/Sur Seq”.

The primary endpoint was OS, defined as the time from diagnosis to death from any cause or last follow-up. Survival time was extracted from the “Survival months” field, and vital status was obtained from “Vital status recode”. For regression analysis, categorical variables were encoded as dummy variables, while continuous variables (age, survival time) were maintained in their original scale.

Study design

This retrospective cohort study was designed in accordance with the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis statement guidelines for developing and validating a multivariable prediction model. The study comprised six sequential analytical phases: (I) variable screening through univariate Cox regression analysis to identify potential prognostic factors; (II) model development employing both full multivariate Cox regression and Akaike Information Criterion (AIC)-based backward stepwise selection to determine the optimal combination of predictors; (III) nomogram construction and visualization of variable-specific Kaplan-Meier survival curves to illustrate the prognostic impact of individual variables; (IV) model validation including discrimination, calibration, and clinical utility assessment in both training and validation cohorts; (V) subgroup analysis stratified by age, sex, tumor stage, and treatment modalities to evaluate model robustness and generalizability; (VI) risk stratification with optimal cutoff determination and classification performance evaluation for clinical decision-making (Figure 1).

Statistical analysis

According to the rule of thumb for Cox regression, a minimum of 10 events per variable (EPV) is recommended to avoid overfitting. With 6 predictor variables and approximately 1,500 death events in the training set, our EPV ratio exceeded 200:1, well above the recommended threshold. Continuous variables were expressed as median [interquartile range (IQR)] and compared using the Mann-Whitney U test, and categorical variables were presented as frequencies (percentages) and compared using Chi-square or Fisher’s exact test. Univariate Cox proportional hazards regression identified potential prognostic factors associated with OS. Variables with P<0.05 were entered into multivariate modeling. Two multivariate approaches were employed: (I) a full model including all univariate-significant variables; (II) an AIC-based model using backward elimination to optimize parsimony. Based on the AIC-selected model, a prognostic nomogram was constructed to estimate 6-month, 1-year, and 3-year survival probabilities using the rms package in R. Variable scores were derived from regression coefficients and linearly transformed to a 0–100 scale.

Model performance was assessed through discrimination [consistency index (C-index) with 95% confidence interval (CI), Brier score], calibration (calibration curves, slope at 6 months and 1 year, R-squared), and clinical utility [decision curve analysis (DCA)]. Subgroup analyses examined model consistency across age (<65 versus ≥65 years), sex, N/M stage, and treatment modalities. Forest plots displayed hazard ratios with interaction tests for heterogeneity. Risk stratification was performed using cutoff points determined by the receiver operating characteristic (ROC) curve analysis based on 3-month and 1-year survival in the training cohort (Youden’s index). Sensitivity, specificity, positive predictive value, and negative predictive value were calculated for each risk group. All statistical analyses were performed using R software (version 4.5.0) with packages. Two-sided P<0.05 indicated statistical significance.

Results

Patient characteristics

SEER data from 2000–2022 indicated a rising incidence of ECC, from approximately 400 to over 800 cases annually, peaking in the 70–74 years age group. Notably, 50% of patients had survival <6 months, including 24.7% with survival <1 month (Figure 2). For the survival analysis, 2,254 patients diagnosed between 2018 and 2022 were enrolled and randomly assigned to training (n=1,577) and validation (n=677) cohorts. Baseline characteristics were well balanced between cohorts (all P>0.05) (Table S1). The study population had a median age at diagnosis of 71.0 years, comprising 56.3% male and 72.9% White patients. The majority of tumors were located in the distal bile duct (71.6%), with most patients presenting with advanced-stage disease (regional or distant, 83.4%). Regarding treatment modalities, 66.9% of patients did not receive surgical intervention, whereas 51.0% underwent chemotherapy (Table S1).

Figure 2 Epidemiological characteristics and survival patterns. (A) Temporal trend of annual incidence from 2000 to 2022 in the SEER database, illustrating progressive increase in case numbers. (B) Age distribution at diagnosis (2000–2022) showing peak incidence in the 70–74 years age group. (C) Distribution of overall survival time intervals (2000–2022), with the largest proportion (24.7%) having survival time of 0–1 month. SEER, Surveillance, Epidemiology, and End Results.

Nomogram development

In the training cohort, univariate Cox analysis identified 11 potential prognostic factors (P<0.05). Multivariate analysis adjusting for all covariates showed that sex, T stage, tumor location, other surgery, and systemic therapy lost significance, likely due to collinearity. Using backward stepwise selection based on AIC, the optimal model included six variables: age (continuous), N stage (N1, N2, NX versus N0), M stage (M1 versus M0), surgery (partial and radical versus none), radiation, and chemotherapy (AIC: 13,596.19 versus 13,612.09 for the full model) (Table S2).

Figure 3 presents Kaplan-Meier curves for the six variables. Notably, N1/N2 stages paradoxically showed better survival than N0 in unadjusted analysis, likely reflecting treatment selection bias (patients with nodal disease more likely to receive curative-intent therapy). However, after multivariate adjustment, N1 and N2 remained independent risk factors [hazard ratio (HR) 1.423 and 2.580, respectively], indicating worse prognosis when controlling for treatment.

Figure 3 Kaplan-Meier survival curves stratified by AIC model variables in the training cohort. (A) Age. (B) N stage. (C) M stage. (D) Surgery. (E) Chemotherapy. Shaded areas represent 95% confidence intervals. Statistical significance was assessed using the log-rank test. All variables showed statistically significant differences in survival (P<0.001). AIC, Akaike Information Criterion; M, metastasis; N, node.

The nomogram incorporated these six variables, with scores derived from hazard ratios (Figure 4). Surgery showed the strongest protective effect (partial: HR 0.263, 95% CI: 0.210–0.329; radical: HR 0.278, 95% CI: 0.218–0.353; both P<0.001). Advanced nodal involvement (N2: HR 2.580, 95% CI: 1.902–3.500; N1: HR 1.423, 95% CI: 1.146–1.766) and distant metastasis (M1: HR 1.406, 95% CI: 1.218–1.623) conferred the highest risk scores. Each year of age increased risk (HR 1.010, 95% CI: 1.005–1.016), while chemotherapy reduced mortality (HR 0.420, 95% CI: 0.367–0.482).

Figure 4 Nomogram for predicting overall survival in patients with extrahepatic cholangiocarcinoma. M, metastasis; N, node.

Nomogram validation

The nomogram demonstrated good discrimination, with C-indices of 0.770 (95% CI: 0.750–0.790) in the training cohort and 0.789 (95% CI: 0.768–0.802) in the validation cohort (Table S3). Calibration curves showed good agreement between predicted and observed survival probabilities at 6 months and 1 year, with calibration R-squared values exceeding 0.93 in both cohorts. Brier scores remained below 0.2, indicating low prediction error. Calibration slopes slightly exceeded 1.0, suggesting marginally optimistic predictions at higher probabilities (Table S3 and Figure 5).

Figure 5 Calibration curves and decision curve analysis of the nomogram. (A) The calibration curves of overall survival nomograms in the training cohort. (B) The calibration curves of overall survival nomograms in the validation cohort. The dashed diagonal line represents perfect calibration. Error bars indicate 95% confidence intervals. The proximity of the data points to the diagonal line demonstrates excellent agreement between predicted and observed survival probabilities in both cohorts. (C) Decision curve analysis for the nomogram in the training cohort. (D) Decision curve analysis for the nomogram in the validation cohort. The shaded area indicates the net benefit gained by using the nomogram compared to treating all or no patients.

DCA demonstrated positive net benefit across threshold probabilities of 0.05–0.80 in the training cohort and 0.10–0.85 in the validation cohort, outperforming both treat-all and treat-none strategies (Figure 5). These results confirm good discrimination, excellent calibration, and clinical utility of the nomogram for predicting survival in ECC.

Subgroup analysis of the nomogram

Subgroup analyses showed consistent discrimination across patient subsets, with C-indices ranging from 0.770 to 0.789. The nomogram performed comparably across age and sex subgroups (Figure 6). However, discrimination decreased in patients with advanced nodal disease (N2/NX: 0.676–0.737) and distant metastasis (M1: 0.697–0.740) compared to earlier-stage disease. Chemotherapy status showed the strongest effect on model performance, with lower C-indices in patients not receiving chemotherapy (0.634–0.678) than in those who did (0.730–0.736), possibly due to greater heterogeneity in disease progression (Table S4). These patterns persisted in the validation cohort, supporting the model’s generalizability while suggesting careful application in advanced-stage or chemotherapy-untreated patients.

Figure 6 Subgroup analysis of nomogram discriminative performance across patient subpopulations in the training (blue circles) and validation (red squares) cohorts. The x-axis represents the C-index with 95% confidence intervals (horizontal bars). Sample sizes for each subgroup are indicated (training/validation). The green shaded zone denotes good discrimination (C-index 0.70–0.80). Dashed vertical lines indicate the overall C-indices for the training (0.770) and validation (0.789) cohorts, respectively. C-index, consistency index.

Risk group classification

Based on actual survival duration, patients were stratified into three risk groups: high-risk (<3 months, n=343), intermediate-risk (3–12 months, n=631), and low-risk (>1 year, n=603). Nomogram scores differed significantly across these categories, with median scores of 82 (IQR, 72–88), 63 (IQR, 50–78), and 38 (IQR, 22–52) points, respectively (P<0.001) (Figure 7A,7B). ROC curve analysis identified optimal cutoff values (Figure 7C,7D). For high-risk identification, the AUC was 0.792 at a cutoff of 69.9 points (sensitivity 0.76, specificity 0.76). For low-risk identification, the AUC was 0.794 at a cutoff of 58.5 points (sensitivity 0.75, specificity 0.72). Based on these thresholds, risk categories were defined as: high-risk (≥70 points), intermediate-risk (59–69 points), and low-risk (≤58 points).

Figure 7 Risk stratification performance of the nomogram in the training and validation cohorts. (A) Box plot of nomogram scores across risk groups. The box represents the IQR, with the horizontal line inside indicating the median value. The whiskers extend to 1.5× IQR, and individual circles represent outliers. Scores increased progressively from low-risk to intermediate-risk and high-risk groups. (B) Histogram of nomogram scores. Histogram displays the density distribution of nomogram total scores stratified by risk group. Different colors represent high-risk (red), intermediate-risk (orange), and low-risk (green) patients. The overlapping distributions demonstrate clear separation between risk categories, with low-risk patients concentrated at lower scores, intermediate-risk patients distributed across middle ranges, and high-risk patients clustered at higher scores. (C) ROC curve for identifying high-risk patients (survival <3 months) versus all others in the training cohort. The AUC was 0.792, indicating good discriminative performance. The optimal cutoff value was determined to be 69.9 points (red dot). (D) ROC curve for identifying low-risk patients (survival >1 year) versus all others in the training cohort. The AUC was 0.794, demonstrating comparable discriminative ability to the high-risk model. The optimal cutoff value was 58.5 points (green dot). (E) Confusion matrix showing the classification performance of the nomogram-based risk stratification in the training cohort. Rows represent actual risk groups based on observed survival duration, and columns represent predicted risk groups based on nomogram scores. The diagonal values indicate correct classification rates. Color intensity corresponds to the proportion of patients in each category. (F) Classification performance metrics for each risk group in the training cohort. The bar chart compares sensitivity, specificity, PPV, and overall accuracy across high-risk (red), intermediate-risk (orange), and low-risk (green) categories. (G) Confusion matrix showing the classification performance of the nomogram-based risk stratification in the validation cohort. (H) Classification performance metrics for each risk group in the validation cohort. AUC, area under the curve; IQR, interquartile range; PPV, positive predictive value; ROC, receiver operating characteristic; Sens, sensitivity; Spec, specificity.

Classification performance was evaluated using confusion matrices. In the training cohort (Figure 7E,7F), the model correctly identified 75.9% of high-risk patients and 61.5% of low-risk patients, with an overall accuracy of 58.0%. The intermediate-risk group showed lower classification accuracy (39.5%), likely due to overlapping prognostic features. In the validation cohort (Figure 7G,7H), performance was consistent: high-risk identification rate of 76.4%, low-risk identification rate of 63.7%, and overall accuracy of 56.4%. This consistency supports the robustness of the risk stratification system for clinical application in ECC.

Discussion

In this study, we developed and internally validated a prognostic nomogram incorporating treatment data to predict 6-month and 1-year OS in patients with ECC using SEER data (for the period from 2018 to 2022). This recent timeframe was selected to reflect contemporary treatment patterns, including the standardization of gemcitabine-cisplatin regimens and advanced biliary interventions. Our model, integrating age, N stage, M stage, surgery, and chemotherapy status, demonstrated good discrimination [C-index 0.770 (95% CI: 0.750–0.790) in the training cohort and 0.789 (95% CI 0.768–0.802) in the validation cohort] and favorable clinical net benefit across relevant threshold probabilities, indicating utility for risk stratification in clinical practice. Our model addresses the limitations of prior single-center models by leveraging a large, unselected, population-representative cohort with complete treatment information, while maintaining the clinical interpretability necessary for bedside use. Notably, this prognostic model is designed for survival prediction and risk stratification, not for causal inference regarding treatment efficacy. The inclusion of treatment variables (surgery and chemotherapy) serves to improve prognostic accuracy by capturing real-world treatment patterns and patient selection effects, rather than to estimate unbiased treatment effects. Users should not interpret the hazard ratios for these variables as causal effect estimates.

The epidemiological characteristics of our cohort underscore the aggressive nature of ECC. The median age at diagnosis (71 years) and male predominance (56.3%) align with previous population-based reports (3,15). Importantly, over 30% of patients survived less than 3 months, with an additional 24.7% surviving less than 1 month, resulting in a median survival of only 7 months. These figures are consistent with prior SEER-based analyses reporting 6- to 9-month median survival (3,16). However, previous studies predominantly relied on conventional AJCC staging or limited demographic variables without adequately considering therapeutic interventions (16-20). By systematically integrating surgery and chemotherapy status, the present model enhances prognostic precision and addresses prior limitations in guiding individualized palliative decisions, offering clinicians a practical risk assessment framework.

The identification of chemotherapy as an independent protective factor (HR 0.420, 95% CI: 0.367–0.482) confirms findings from the ABC-02 trial and subsequent studies showing survival benefits with gemcitabine-cisplatin therapy in advanced biliary tract cancers (21,22). Notably, the discriminative ability of our nomogram decreased markedly in the non-chemotherapy subgroup (C-index 0.634), indicating unmeasured confounders such as performance status and comorbidity burden that influence both chemotherapy eligibility and survival outcomes in patients receiving best supportive care alone (23-25). Patients managed solely with best supportive care constitute a markedly heterogeneous cohort, ranging from frail elderly individuals to those with rapidly progressive disease too debilitated to tolerate active treatment. Such broad clinical variability is inadequately captured by routine registry data. These findings further support our interpretation that treatment status can serve as a robust prognostic surrogate for overall patient physical fitness. Surgical intervention, whether partial or radical, showed the strongest protective effects (HR 0.263, 95% CI: 0.210–0.329; and HR 0.278, 95% CI: 0.218–0.353, respectively), highlighting the importance of curative resection or cytoreduction in selected patients. However, given that only 33.1% of patients underwent surgery, with most presenting with advanced disease, accurate prognostication to guide palliative systemic therapy remains essential. The strong protective effects observed for surgery and chemotherapy likely reflect a combination of genuine treatment benefit and substantial selection bias. Patients who underwent surgical resection or received systemic chemotherapy were presumably younger, healthier, and had better performance status than those managed conservatively-factors that are independently prognostic but are not recorded in the SEER database. Consequently, these hazard ratios should not be interpreted as unbiased estimates of treatment efficacy in the causal sense. Instead, treatment status functions as a composite proxy for overall fitness, disease resectability, and access to multidisciplinary care, which are themselves powerful prognostic determinants. This is a fundamental and unavoidable limitation of any observational prognostic model that incorporates treatment variables.

An unexpected finding was the apparent paradox regarding N stage: while N1 and N2 disease showed higher mortality risk than N0 (HR 1.423, 95% CI: 1.146–1.766; and HR 2.580, 95% CI: 1.902–3.500, respectively), Kaplan-Meier curves indicated superior survival for node-positive patients. This counterintuitive observation likely stems from selection bias inherent in retrospective registry data. Patients with N0 disease may represent a heterogeneous group including those with inadequate lymph node dissection (<12 nodes) leading to understaging, or early-stage cases managed less aggressively due to perceived favorable biology. Conversely, patients with documented N1/N2 disease typically underwent intensive multimodal treatment, potentially offsetting the adverse impact of nodal metastasis. Furthermore, the NX category (unknown nodal status) showed the poorest survival, likely reflecting advanced disease with incomplete staging or rapid deterioration precluding comprehensive evaluation (26,27). Notably, the high proportion of TX cases (42.1% in the final cohort) substantially undermined the prognostic performance of T category and directly accounted for its exclusion from the AIC-driven final model. Although T stage reached statistical significance in univariate analysis, this effect was largely driven by the pronounced survival difference between the small subgroup with defined T1–T4 disease and the large TX group, rather than by meaningful prognostic stratification among the T1–T4 subgroups themselves. When incorporated into multivariate modeling together with N stage, M stage, and treatment-related variables, T stage offered no additional prognostic value beyond that already captured by these more complete and clinically robust factors. Several key observations support this interpretation: (I) the T1–T4 subgroups were severely imbalanced (T1: 5.7%, T2: 8.3%, T3: 30.3%, T4: 13.6%), with overrepresentation of T3 disease and limited statistical precision for T1 and T2 categories; (II) despite being classified as unstaged, TX patients conveyed strong prognostic information and experienced the poorest outcomes in Kaplan-Meier analysis, likely reflecting advanced, unresectable, or incompletely characterized disease; (III) N stage and M stage, with substantially lower missing rates, enabled more reliable assessment of anatomic disease extent than the T category, which was heavily distorted by the high proportion of TX cases; (IV) treatment variables including surgery and chemotherapy likely captured much of the prognostic signal ordinarily conveyed by T stage, given that resectability—closely tied to T category—is implicitly reflected in receipt of surgical intervention. Accordingly, T stage was eliminated from the model not because it lacks biological relevance, but because population-based T staging in ECC is inherently limited by the infiltrative growth pattern of the disease. This represents a data quality constraint rather than a biologically negative result. A substantial proportion of patients were excluded due to unknown T stage. This high exclusion rate raises concerns about selection bias. Patients with TX likely represent a heterogeneous group including: (I) patients with advanced, unresectable disease where the primary tumor could not be adequately assessed due to infiltrative growth along the bile duct; (II) patients who were too frail for comprehensive staging workup; (III) cases with incomplete pathology reporting. The TX group showed the poorest survival in our Kaplan-Meier analysis, suggesting that these patients were indeed more advanced or had poorer overall health status. By excluding them, our model may underestimate mortality risk for the most advanced cases and overestimate performance in the general ECC population. Users should consider this issue when applying the model to patients with incomplete staging information.

Risk stratification identified three prognostic groups: high-risk (score ≥70, expected survival <3 months), intermediate-risk (score 59–69, 3–12 months), and low-risk (score ≤58, >1 year). These categories can guide palliative biliary drainage decisions. For high-risk patients, less invasive approaches are preferable. Given the risks of post-ERCP pancreatitis (3–10%), hemorrhage (1–2%), and perforation (0.3–0.6%), PTBD offers a safer alternative for those with limited life expectancy (28). PTBD avoids endoscopic complications, can be performed rapidly under local anesthesia, and is suitable for patients with poor performance status, providing adequate decompression without excessive procedural risk (29-31). For intermediate-risk and low-risk patients (expected survival >3 months), endoscopic biliary stenting via ERCP is preferable, with self-expanding metal stents (SEMS) favored over plastic stents given longer patency (median 8–12 versus 3–4 months). ERCP avoids the external drainage and electrolyte disturbances associated with percutaneous approaches, while SEMS reduce the need for reintervention due to stent occlusion (32-37). This risk stratification framework offers clinicians practical criteria for selecting biliary drainage modalities based on anticipated survival, potentially improving resource allocation and quality of life. While accurate survival prognostication is a prerequisite for rational palliative care planning, we acknowledge that the SEER database provides no information on biliary drainage techniques, stent types, or procedural outcomes. Therefore, our risk stratification should not be interpreted as a directive for choosing between ERCP and PTBD, or between plastic and metal stents. Instead, it offers a prognostic framework that may contribute to multidisciplinary discussions regarding overall care intensity and goals-of-care conversations. Decisions regarding biliary drainage modality, systemic therapy selection, or best supportive care require integration of multiple clinical factors including performance status, comorbidities, patient preferences, and local expertise. Therefore, the role of this nomogram is to inform prognostic expectations, not to direct therapeutic strategies. The selection of therapeutic strategies we discussed merely reflects the clinical experience of our medical center and is intended for reference.

The period 2018–2021 reflects the mature chemotherapy era for advanced biliary tract cancers, during which gemcitabine plus cisplatin—established as the global standard of care by the ABC-02 trial in 2010—remained the first-line systemic therapy (21). Our prognostic model was developed and validated within this therapeutically homogeneous context and therefore reliably reflects real-world outcomes in patients treated under consistent standard-of-care conditions. In contrast, 2022 marks the early transition period toward immunotherapy-based combinations (25,30). Although durvalumab plus gemcitabine-cisplatin was approved in September 2022, and pembrolizumab-containing regimens gained increasing clinical relevance in late 2022, the adoption of these novel regimens into routine community oncology practice—especially within geographic areas covered by the SEER registry—was likely minimal during our study interval. As a result, only a small subset of patients included in the 2022 cohort may have received immunotherapy. Furthermore, SEER coding does not reliably differentiate immunotherapy from conventional chemotherapy in the fields capturing systemic treatment, limiting the ability to accurately characterize receipt of immunotherapy in this dataset. The C-index of 0.770–0.789 observed in our cohort reflects prognostic discrimination in a population predominantly treated with chemotherapy ± surgery. If applied to a contemporary cohort where 50–70% of patients receive first-line immunochemotherapy, the model may exhibit reduced calibration (systematic overestimation of mortality risk) because the baseline survival of modern patients is likely superior to that predicted by our chemotherapy-era model. However, the C-index may remain acceptable if the relative prognostic importance of age, stage, and treatment responsiveness is preserved across eras. Nevertheless, users should interpret predicted survival probabilities as conservative estimates for patients receiving contemporary immunotherapy or targeted therapy. To maintain clinical relevance, this nomogram requires prospective validation and potential recalibration in cohorts treated with contemporary standards of care. External validation using 2023–2025 SEER or prospective registry data to evaluate calibration drift in the immunotherapy era. Development of an immunotherapy‑adjusted prognostic model that integrates programmed death-ligand 1 (PD‑L1) expression, microsatellite instability status, and first‑line treatment regimen, once these variables become systematically available in population‑based registries. Until such validation is completed, the model should be regarded as applicable primarily to patients treated with conventional chemotherapy and surgery, with cautious extrapolation to those receiving novel systemic agents.

Several limitations should be acknowledged. The retrospective nature of SEER data introduces selection bias regarding treatment assignment—patients receiving chemotherapy or surgery likely had better performance status than those managed conservatively, a variable not captured in the registry. Additionally, the database lacks critical clinicopathological variables including tumor grade, carbohydrate antigen 19-9 (CA 19-9) levels, and R0/R1 resection status. Although tumor size data are available in SEER, the high proportion of missing values precluded inclusion; however, size is often difficult to quantify in ECC due to its infiltrative growth pattern, potentially limiting its prognostic value compared with pancreatic adenocarcinoma (38,39). Furthermore, detailed treatment specifications are unavailable, including specific chemotherapeutic regimens, cycle number, treatment intent (neoadjuvant, adjuvant, or palliative), radiation dosing, and surgical margin status. Moreover, a critical limitation is that treatment variables in the SEER database are non-randomly assigned and confounded by unmeasured indications. Performance status, Charlson Comorbidity Index, and treatment intent (neoadjuvant, adjuvant, or palliative) are not captured. Therefore, the model cannot disentangle the true prognostic effect of treatment from treatment selection bias. This limits the use of the model for evaluating treatment efficacy but does not invalidate its utility for OS prediction, provided that users interpret treatment variables as prognostic markers rather than causal interventions. Another important limitation is that our study period (for the period from 2018 to 2022) overlaps with the coronavirus disease 2019 (COVID-19) pandemic. From March 2020 onward, healthcare systems worldwide experienced severe disruptions, including deferred elective surgeries, delayed biliary interventions, chemotherapy interruptions, reduced emergency department access, and disrupted follow-up care. For patients with ECC—an aggressive malignancy requiring timely multidisciplinary management—such delays could have resulted in disease progression, irresectability, missed chemotherapy cycles, and increased non-cancer mortality. Consequently, survival outcomes in our cohort, particularly for patients diagnosed in 2020–2022, may reflect not only biological disease behavior and standard treatment efficacy, but also healthcare system strain during a global crisis. This limits the generalizability of our model to patient populations treated under usual, non-pandemic clinical circumstances. The reported C-indices of 0.770 (training cohort) and 0.789 (validation cohort) are acceptable for a prognostic model in an aggressive malignancy with substantial unmeasured heterogeneity, but they are not outstanding. For context, high-performing cancer prognostic models routinely achieve C-indices >0.85. Our more modest performance reflects: (I) the inherent biological heterogeneity of ECC; (II) the absence of key prognostic variables in SEER (performance status, tumor grade, molecular markers); (III) the limited record of treatment data (binary yes/no for surgery and chemotherapy, without regimen, intent, or response). Users should interpret the discrimination of the model as moderate-sufficient for broad risk stratification but not for precise individual prognostication. Calibration slopes of 1.210 (6-month) and 1.127 (1-year) in the training cohort, and 1.283 (6-month) and 1.227 (1-year) in the validation cohort, consistently exceed the ideal value of 1.0. This indicates that the model is slightly overfitted. Several factors mitigate the practical impact of this overfitting: (I) the magnitude of slope deviation from 1.0 is modest (1.12–1.28), not extreme (>1.5), suggesting that overfitting is mild rather than severe; (II) the Brier scores remain low (<0.18), indicating acceptable overall prediction error; (III) DCA demonstrates positive net benefit across clinically relevant thresholds, suggesting that the model remains clinically useful despite imperfect calibration. Nevertheless, we acknowledge that calibration should be monitored closely in external validation, and that recalibration may be necessary before clinical implementation. Our nomogram has undergone only internal validation via random split-sample allocation from the same database. This is the weakest form of validation in the prognostic model hierarchy. Thus, external validation is quite necessary. Finally, while our traditional regression-based approach prioritizes interpretability, broad population representativeness, and accessibility, we acknowledge that machine learning methods—particularly ensemble methods and deep learning—may capture complex nonlinear interactions and achieve superior discrimination in well-characterized cohorts with complete biomarker, genomic, and cross-sectional imaging data. Future research should directly compare the performance of our nomogram against ML models in independent, prospectively collected datasets that include the detailed clinical variables necessary for optimal ML performance.

Conclusions

We developed and internally validated a prognostic nomogram incorporating clinical and treatment variables to predict survival in ECC. This model provides individualized survival estimates and objective risk stratification for prognostic counseling and research purposes. Specific treatment decisions, including biliary drainage modality, require integration of clinical factors not captured in the SEER database and should be individualized based on patient condition, anatomy, and institutional expertise.

Acknowledgments

We acknowledge the National Cancer Institute and the SEER Program tumor registries for their invaluable contribution to the creation and maintenance of the SEER database.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jgo.amegroups.com/article/view/10.21037/jgo-2026-0258/rc

Peer Review File: Available at https://jgo.amegroups.com/article/view/10.21037/jgo-2026-0258/prf

Funding: This study was supported by the Military Medical Research Program (grant No. 24BJZ18).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jgo.amegroups.com/article/view/10.21037/jgo-2026-0258/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Razumilava N, Gores GJ. Cholangiocarcinoma. Lancet 2014;383:2168-79. [Crossref] [PubMed]
2. Kam AE, Masood A, Shroff RT. Current and emerging therapies for advanced biliary tract cancers. Lancet Gastroenterol Hepatol 2021;6:956-69. [Crossref] [PubMed]
3. Halder R, Amaraneni A, Shroff RT. Cholangiocarcinoma: a review of the literature and future directions in therapy. Hepatobiliary Surg Nutr 2022;11:555-66. [Crossref] [PubMed]
4. Brindley PJ, Bachini M, Ilyas SI, et al. Cholangiocarcinoma. Nat Rev Dis Primers 2021;7:65. [Crossref] [PubMed]
5. Al Nakshabandi A, Ali FS, Albustami I, et al. Biliary drainage in hilar and perihilar cholangiocarcinoma: 25-year experience at a tertiary cancer center. Gastrointest Endosc 2024;99:938-949.e15. [Crossref] [PubMed]
6. Wheless M, Agarwal R, Goff L, et al. Current Standards, Multidisciplinary Approaches, and Future Directions in the Management of Extrahepatic Cholangiocarcinoma. Curr Treat Options Oncol 2024;25:127-60. [Crossref] [PubMed]
7. Rushbrook SM, Kendall TJ, Zen Y, et al. British Society of Gastroenterology guidelines for the diagnosis and management of cholangiocarcinoma. Gut 2023;73:16-46. [Crossref] [PubMed]
8. Banales JM, Marin JJG, Lamarca A, et al. Cholangiocarcinoma 2020: the next horizon in mechanisms and management. Nat Rev Gastroenterol Hepatol 2020;17:557-88. [Crossref] [PubMed]
9. Dar FS, Abbas Z, Ahmed I, et al. National guidelines for the diagnosis and treatment of hilar cholangiocarcinoma. World J Gastroenterol 2024;30:1018-42. [Crossref] [PubMed]
10. Zori AG, Yang D, Draganov PV, et al. Advances in the management of cholangiocarcinoma. World J Hepatol 2021;13:1003-18. [Crossref] [PubMed]
11. Byrling J, Andersson B. Prediction of survival after pancreatoduodenectomy for distal cholangiocarcinoma: independent external validation of a prognostic model for 3-year overall survival in Sweden. Scand J Gastroenterol 2025;60:158-64. [Crossref] [PubMed]
12. Ma Y, Li Q, Tang Z, et al. Development and validation of a machine learning-based nomogram for survival prediction of patients with hilar cholangiocarcinoma after curative-intent resection. Sci Rep 2025;15:25220. [Crossref] [PubMed]
13. Zhu Z, Hu K, Zhao F, et al. Machine learning-based nomogram for 30-day mortality prediction for patients with unresectable malignant biliary obstruction after ERCP with metal stent: a retrospective observational cohort study. BMC Surg 2023;23:260. [Crossref] [PubMed]
14. Zhao J, Zhang W, Zhang J, et al. External validation study of the 8th edition of the American Joint Committee on Cancer staging system for perihilar cholangiocarcinoma: a single-center experience in China and proposal for simplification. J Gastrointest Oncol 2021;12:806-18.
15. Qurashi M, Vithayathil M, Khan SA. Epidemiology of cholangiocarcinoma. Eur J Surg Oncol 2025;51:107064. [Crossref] [PubMed]
16. Hu J, Shi Y, Jin L, et al. Predicting survival rates: the power of prognostic nomograms in distal cholangiocarcinoma. Front Oncol 2025;15:1478836. [Crossref] [PubMed]
17. Zhou SN, Jv DW, Meng XF, et al. Feasibility of machine learning-based modeling and prediction using multiple centers data to assess intrahepatic cholangiocarcinoma outcomes. Ann Med 2023;55:215-23. [Crossref] [PubMed]
18. Zhang Y, Qiao C, Zhao P, et al. Prognostic model for oversurvival and tumor-specific survival prediction in patients with advanced extrahepatic cholangiocarcinoma: a population-based analysis. BMC Gastroenterol 2023;23:422. [Crossref] [PubMed]
19. Zeng Y, Wang X, Wu J, et al. Survival analysis of patients with extrahepatic cholangiocarcinoma: a nomogram for clinical and MRI features. BMC Med Imaging 2024;24:7. [Crossref] [PubMed]
20. Saragih P, Makmun D, Kurniawan J, et al. One Year Survival of Extrahepatic Cholangiocarcinoma Patients Who Did Not Undergo Curative Resection and Paliative Chemotherapy and Its Associated Factors. Acta Med Indones 2022;54:35-41.
21. Valle J, Wasan H, Palmer DH, et al. Cisplatin plus gemcitabine versus gemcitabine for biliary tract cancer. N Engl J Med 2010;362:1273-81. [Crossref] [PubMed]
22. Lamarca A, Ross P, Wasan HS, et al. Advanced Intrahepatic Cholangiocarcinoma: Post Hoc Analysis of the ABC-01, -02, and -03 Clinical Trials. J Natl Cancer Inst 2020;112:200-10. [Crossref] [PubMed]
23. Zhang Q, Liu Z, Liu S, et al. A novel nomogram for adult primary perihilar cholangiocarcinoma and considerations concerning lymph node dissection. Front Surg 2022;9:965401. [Crossref] [PubMed]
24. Li Q, Zhang J, Chen C, et al. A Nomogram Model to Predict Early Recurrence of Patients With Intrahepatic Cholangiocarcinoma for Adjuvant Chemotherapy Guidance: A Multi-Institutional Analysis. Front Oncol 2022;12:896764. [Crossref] [PubMed]
25. Kelley RK, Ueno M, Yoo C, et al. Pembrolizumab in combination with gemcitabine and cisplatin compared with gemcitabine and cisplatin alone for patients with advanced biliary tract cancer (KEYNOTE-966): a randomised, double-blind, placebo-controlled, phase 3 trial. Lancet 2023;401:1853-65. [Crossref] [PubMed]
26. Blechacz B, Komuta M, Roskams T, et al. Clinical diagnosis and staging of cholangiocarcinoma. Nat Rev Gastroenterol Hepatol 2011;8:512-22. [Crossref] [PubMed]
27. Bajracharya AD, Shrestha S, Kim HS, et al. Retrospective analysis of 8th edition American Joint Cancer Classification: Distal cholangiocarcinoma. Ann Hepatobiliary Pancreat Surg 2023;27:251-7.
28. Zhang G, Zhang G, Meng K, et al. A risk score model for post-endoscopic retrograde cholangiopancreatography complications in elderly patients with choledocholithiasis. Saudi J Gastroenterol 2025;31:347-55. [Crossref] [PubMed]
29. Moll CF, de Moura DTH, Ribeiro IB, et al. Endoscopic Biliary Darinage (EBD) versus Percutaneous Transhepatic Biliary Drainage (PTBD) for biliary drainage in patients with Perihilar Cholangiocarcinoma (PCCA): A systematic review and meta-analysis. Clinics (Sao Paulo) 2023;78:100163. [Crossref] [PubMed]
30. Izquierdo-Sanchez L, Lamarca A, La Casta A, et al. Cholangiocarcinoma landscape in Europe: Diagnostic, prognostic and therapeutic insights from the ENSCCA Registry. J Hepatol 2022;76:1109-21. [Crossref] [PubMed]
31. Bang JY, Hawes R, Varadarajulu S. Endoscopic biliary drainage for malignant distal biliary obstruction: Which is better - endoscopic retrograde cholangiopancreatography or endoscopic ultrasound? Dig Endosc 2022;34:317-24. [Crossref] [PubMed]
32. de Jong DM, Gilbert TM, Nooijen LE, et al. Preoperative endoscopic biliary drainage by metal versus plastic stents for resectable perihilar cholangiocarcinoma. Gastrointest Endosc 2024;99:566-576.e8. [Crossref] [PubMed]
33. Kawashima H, Ohno E, Ishikawa T, et al. Endoscopic management of perihilar cholangiocarcinoma. Dig Endosc 2022;34:1147-56. [Crossref] [PubMed]
34. Lee TH, Moon JH, Stuart S. Role of ERCP in Malignant Hilar Biliary Obstruction. Gastrointest Endosc Clin N Am 2022;32:427-53. [Crossref] [PubMed]
35. Zuniga Cisneros J, Mc-Kenzie A, Ramos I, et al. Endoscopic Biliary Drainage in Malignant Biliary Strictures: A Prospective Analysis Based on Real-World Data. Gastroenterology Res 2025;18:192-201. [Crossref] [PubMed]
36. Marzioni M, Crinò SF, Lisotti A, et al. Biliary drainage in patients with malignant distal biliary obstruction: results of an Italian consensus conference. Surg Endosc 2024;38:6207-26. [Crossref] [PubMed]
37. Mocan T, Horhat A, Mois E, et al. Endoscopic or percutaneous biliary drainage in hilar cholangiocarcinoma: When and how? World J Gastrointest Oncol 2021;13:2050-63. [Crossref] [PubMed]
38. Ishiwatari H, Sato J, Sakamoto H, et al. Current status of preoperative endoscopic biliary drainage for distal and hilar biliary obstruction. Dig Endosc 2024;36:969-80. [Crossref] [PubMed]
39. Vithayathil M, Khan SA. Current epidemiology of cholangiocarcinoma in Western countries. J Hepatol 2022;77:1690-8. [Crossref] [PubMed]