This is a retrospective two-center clinical cohort study based on a dataset of patients with histologically confirmed ICC who underwent curative hepatectomy in the First Affiliated Hospital of Sun Yat-sen University (FAH-SYSU) from January 2012 to January 2022 and Sun Yat-sen University Cancer Center (SYSUCC) from January 2018 to January 2022. Patients were excluded if they had positive surgical margin (R1/R2), positive lymph node, received other preoperative anticancer therapy, were presence of distant metastases, combination of other malignancies, or perioperative mortality (within 30 postoperative days). All patients included in the two centers underwent curative-intent hepatectomy. The operative strategies were determined based on the multidisciplinary tumor meeting discussions and performed according to standard clinical practice. Specifically, the extent of surgical resection was considered based on the size, location, and number of the tumor and the presence of cirrhosis. R0 resection was achieved in the presence of sufficient volume of future liver remnant. The extent of lymph node dissection was based on preoperative imaging suspicion of lymph node metastasis or intraoperative evaluation. Remove direct invasion of adjacent structures and local extrahepatic metastases identified intraoperatively. The ethics committees of both participating institutions approved the study. A waiver for informed consent was obtained and patient identifying information was removed.

The prognosis model was developed in the derivation cohort of FAH-SYSU and tested in the external validation cohort of SYSUCC. All histopathologically confirmed ICC were positive for CK7, CK19 and MUC1, and negative for CK20, HepPar1 and glypican-3 (18). The definition of ESICC refers to ICC patients with R0 and lymph node-negative. Considering that compared with R1 and lymph node-positive, ICC patients with R0 and lymph node-negative tend to have a better prognosis, and guidelines do not mandatorily recommend postoperative chemotherapy for such patients (3, 5, 9, 19).

Information was collected according to a standardized datasheet. Demographic and clinicopathological variables included age, gender, medical history (i.e., diabetes, cholelithiasis, hepatitis and cirrhosis), preoperative laboratory values (i.e., neutrophil–lymphocyte ratio [NLR], platelet–lymphocyte ratio [PLR], albumin [ALB], total bilirubin [TBIL], alanine aminotransferase [ALT], aspartate aminotransferase [AST], γ-glutamyl transferase [γ-GGT], alkaline phosphatase [ALP], carbohydrate antigen 19-9 [CA19-9], carcinoembryonic antigen [CEA]), and tumor-related characteristics (ie tumor number, tumor size, lymph node status, surgical margin width, margin status, tumor differentiation grade, vascular and adjacent organs invasion). Tumor stage was categorized according to the 8th edition of the American Joint Committee on Cancer (AJCC) staging manual (20).

After liver resection, patients were regularly followed and monitored for recurrence by detection of CA19-9, CEA, and imaging scanning including ultrasonography, computed tomography, or/and magnetic resonance imaging, once every 3–6 months for the first two years and once every 6–12 months thereafter. ICC recurrence was defined as suspicious or positive findings on two types of surveillance imaging or histologically confirmed disease. The date of the recurrence was confirmed by reviewing imaging reports, pathology reports, and hospitalization records from the electronic medical records. OS was defined from the date of surgery to the date of death or last follow-up. RFS was measured from the date of surgery to the date of first recurrence or death or last follow-up.

Categorical variables were presented as frequencies with percentages. Continuous variables were presented as means with standard deviations (SD) or medians with interquartile ranges (IQR). Differences between continuous variables were assessed with the Student's t tests or Mann–Whitney tests, and differences between categorical variables were assessed using the Fisher's exact tests. In the derivation cohort, the association of clinicopathological variables with the prognosis following ESICC resection was assessed by Cox proportional hazards regression analysis. Variables significant on univariate analysis (P < 0.1) were subsequently included into the multivariable Cox regression model, and a step-wise backward selection approach was used to remove nonsignificant variables using a P value < 0.05. The risk scores to predict ESICC prognosis were developed based on the β-coefficients of the clinicopathological variables significant in the final step of the multivariable Cox regression analysis. Relative risk was expressed as hazard ratios (HR) with 95% confidence interval (CI). The Harrell's concordance index (C-index) was calculated to evaluate the discriminatory performance of the prognosis model. The prognostic discrimination of the model was subsequently estimated in the external validation cohort. Nomogram for 1-, 2-, and 3-year OS probabilities was constructed based on the prognosis model. Moreover, we evaluated the calibration of the nomogram to compare the risk predicted by the nomogram with the observed risk. Receiver operating characteristic curves (ROC) were used to compare the predictive performance of the prognosis model with other prognostic factors for 3-year OS. The median of risk scores was used to classify patients into high- and low-risk groups. Kaplan-Meier curves and log-rank tests were performed to assess the distinctions in OS and RFS between different risk stratifications, and to compare the prognosis model with the preexisting staging system. In addition, the clinical effectiveness and utility were assessed by decision curve analysis (DCA) and net reclassification index (NRI), respectively. All P values were two-sided, and P < 0.05 was considered statistically significant. All statistical analyses were constructed with Stata/MP version 14.0, and R version 3.6.2.

A total of 263 patients matched the inclusion and exclusion criteria and were enrolled in the final analytic cohorts, comprising 180 patients in the derivation cohort and 83 patients in the external validation cohort. The demographic and clinicopathological variables are detailed in Table 1. The mean age of patients was 58.0 years (SD, 10.5 years), and 153 patients (58.2%) were male. Diabetes, cholelithiasis, hepatitis, and cirrhosis were present in 41 (15.6%), 50 (19.0%), 108 (41.1%), and 24 (9.1%) patients, respectively. A subset of patients (n = 53, 20.2%) were suspicious or positive regional lymph node metastasis on preoperative imaging. Tumor size was universally large, with a median diameter of 5 cm (IQR, 3.5–7.0 cm), and single nodule was predominant (n = 205, 77.9%). Most patients underwent wide margin hepatectomy (n = 187, 71.1%). Nearly half of the patients held poorly differentiated tumors (n = 124, 47.1%), only a few patients had vascular invasion (n = 36, 13.7%), and the majority of patients were in the AJCC T1 stage (n = 166, 63.1%).

After a median follow-up of 42.4 months (95% CI, 32.4–46.6 months) in the derivation cohort, 77 (42.8%) patients recurred within one year after surgery and median OS was 33.8 months (95% CI 27.9–45.6 months). The 3-year OS was 49.0% (95% CI, 40.0–57.4%). Differential coefficients of 13 clinicopathological covariates were associated with outcomes in univariate Cox regression analysis of the derivation cohort (Table 2, Supplementary Figure S1). Among them, five factors showed significant in a further multivariable analysis, which demonstrated that higher count of CA19-9 (HR, 1.01; 95% CI, 1.00–1.01) and CEA (HR, 1.01; 95% CI, 1.00–1.01), larger tumor size (HR, 1.15; 95% CI, 1.07–1.23), and T2–T4 stage disease (HR, 1.78; 95% CI, 1.15–2.78) were respectively associated with lower odds of survival, whereas well or moderately differentiated tumors (HR, 0.54; 95% CI, 0.34–0.85) were again associated with greater odds of survival (Table 2, Supplementary Figure S2). None of the remaining variables noted to be independent predictors of survival.

Risk scores were developed based on the aforementioned five independent prognostic factors and a nomogram was constructed to estimate 1-, 2-, and 3-year survival probability (Figure 1). The discriminatory performance of the prognosis model was satisfying in the derivation cohort (C-index, 0.71; 95% CI, 0.65–0.77), and superior to the discriminative accuracy of the 8th edition AJCC TNM staging system, which had a C-index of 0.59 (95% CI, 0.53–0.65). Meanwhile, the prognosis model outperformed individual prognostic parameters, including CA19-9 (C-index, 0.64; 95% CI, 0.56–0.71), CEA (C-index, 0.59; 95% CI, 0.52–0.65), tumor size (C-index, 0.66; 95% CI, 0.60–0.73), tumor differentiation (C-index, 0.57; 95% CI, 0.52–0.63) and T stage (C-index, 0.57; 95% CI, 0.51–0.63) (Supplementary Figure S3A). Furthermore, the calibration cures indicated overall good agreement between the predicted and the observed probabilities of OS in the prognosis model (Supplementary Figure S3B). The NRI further demonstrated that the model had higher predictive power than the TNM staging system and other prognostic markers (Supplementary Table S1). In addition, the DCA showed that using this prognosis model to predict survival provided more clinical benefit than TNM system (Supplementary Figure S3C).Based on the risk scores, patients were noticeably stratified into high — and low-risk groups. The survival of low-risk patients was significantly better than that of high-risk patients, with 3-year OS of 70.9% and 24.9% (P < 0.001), respectively (Figure 2A). Moreover, the model accurately distinguished RFS between high- and low-risk patients, with the low-risk subgroup having a longer RFS compared to the high-risk subgroup (median RFS: low risk, 24.3 months vs. high risk, 6.4 months; P < 0.001) (Figure 2B).

The clinicopathological characteristics of the patients in the external validation cohort are shown in Table 1. In this cohort, the median OS was 49.5 months (95% CI, 31.5-NA months), the number of recurrences within one year was 26 (31.3%), and the 3-year OS was 69.6% (95% CI, 52.8–81.4%). Validation of the proposed model was performed in this independent cohort from another clinical center. In the assessment of discriminative capacity, the prognosis model exhibited remarkable accuracy (C-index, 0.78; 95% CI, 0.67–0.88) overperforming the 8th edition AJCC staging system (C-index, 0.54; 95% CI, 0.41–0.67) and the individual prognostic parameters (CA19-9 [C-index, 0.65; 95% CI, 0.47–0.82), CEA [C-index, 0.62; 95% CI, 0.46–0.78], tumor size [C-index, 0.60; 95% CI, 0.46–0.74], tumor differentiation [C-index, 0.57; 95% CI, 0.46–0.69] and T stage [C-index, 0.54; 95% CI, 0.41–0.67]) (Supplementary Figure S4A), with a good calibration (Supplementary Figure S4B). Compared to the TNM staging system and other prognostic variables, the model had higher NRI values (Supplementary Table S1) and net clinical benefit (Supplementary Figure S4C).

Our intent of categorization was achieved excellently in the validation cohort that risk stratifications based on the prognosis model showed homogeny to the derivation cohort (Figures 3A,B), and superior risk-dependent OS and RFS observed in the prognosis model compared to the TNM staging system (Figures 3C,D). Specifically, patients deemed high risk had a worse OS and RFS compared with patients who were low risk. The detailed 3-year OS, RFS at 6 months and 1 year, and hazard ratios for both high- and low-risk groups are shown in Table 3.