Preoperative Prediction Model for Early Recurrence of Intrahepatic Cholangiocarcinoma after Surgical Resection: Development and External Validation Study

1. Study design and population

This multicenter, retrospective study was approved by the institutional review boards of all participating centers, which waived the requirements for informed consent. The study included consecutive patients who underwent curative-intent hepatic resection for ICCA at one of six institutions between January 2009 and December 2016. Patients were included if they (1) were 18 years of age or older and (2) underwent multiphasic CT prior to surgery. Patients were excluded if they met any of the following criteria: (1) a history of other malignancies, (2) synchronous hepatocellular carcinoma (HCC), (3) any prior treatment for ICCA, (4) no contrast-enhanced CT within 3 months before surgery, (5) other types of ICCA (e.g., intraductal papillary neoplasm of the bile duct [IPNB]), (6) palliative hepatic resection, or (7) death or loss to follow-up without any evidence of recurrence within 1 year after resection. Data from Asan Medical Center were used to develop the scoring system, while data from the other five institutions (Severance Hospital, Gangnam Severance Hospital, Seoul National University Bundang Hospital, Korea University Ansan Hospital, and National Health Insurance Service Ilsan Hospital) were utilized for external validation (Fig. 1). Of the 365 patients finally enrolled in this study, 314 had been previously reported [19]. While the prior study developed a prognostic model using magnetic resonance imaging (MRI) features to predict OS in patients with ICCA after resection, this study aimed to develop a risk scoring system based on CT and clinical features to predict ER of ICCA.

Fig. 1.

Flow diagram of the study population. CT, computed tomography; HCC, hepatocellular carcinoma; ICCA, intrahepatic cholangiocarcinoma.

2. Clinical data

Clinical and pathological data, including demographics, serum tumor markers (carbohydrate antigen 19-9 [CA19-9] and carcinoembryonic antigen [CEA]), viral markers (hepatitis B surface antigen [HBsAg] and hepatitis C antibody), tumor differentiation, the extent of hepatectomy, adjuvant treatment status, and postoperative outcomes, were retrospectively collected from the electronic medical records of each institution. Major hepatectomy was defined as the resection of three or more Couinaud segments, while minor hepatectomy was defined as the resection of two or fewer segments or a wedge resection. Chronic biliary disease was defined as a clinical history of any chronic biliary condition, such as intrahepatic bile duct stones, liver fluke infection, primary sclerosing cholangitis, or primary biliary cirrhosis.

3. CT acquisition and analysis

CT examinations were performed using 16-, 64-, 128-, or 256-slice CT scanners from various vendors (GE HealthCare, Philips Healthcare, and Siemens Healthineers) (Supplementary Methods). CT images were acquired both before and after the intravenous injection of 120-150 mL of non-ionic iodinated contrast at a rate of 3-4 mL/sec, followed by a 20 mL saline flush administered via an automatic power injector. Arterial-phase (AP) images were obtained using a bolus-tracking technique, with a delay of 10-25 seconds after reaching a peak abdominal aortic enhancement of 100 Hounsfield units. Portal venous phase and delayed-phase images were captured at 70-80 seconds and 3 minutes after contrast injection, respectively. The scanning parameters were 70-130 kVp, 120-240 mAs utilizing automatic exposure control, and a slice thickness of 2-5 mm with no gap. Coronal reformatted images for the portal venous phase were available for all patients.

Three board-certified abdominal radiologists, each with over 13 years of experience in hepatic imaging, independently reviewed the CT images from both the development and validation cohorts. To minimize the inter-reader variability across the institutions, prior to the independent review, the three reviewers discussed and defined each imaging finding using 50 representative ICCA cases that were not included in this study. The reviewers were blinded to clinical, pathological, and prognostic information during the image analysis. All images were provided as anonymized Digital Imaging and Communications in Medicine (DICOM) files and were analyzed using a commercial DICOM viewer (RadiAnt DICOM Viewer, Medixant). The CT findings assessed included tumor size, tumor contour, tumor location, AP enhancement pattern, AP peritumoral enhancement, delayed central enhancement, tumor multiplicity, vascular invasion, tumor-in-vein, extrahepatic organ invasion, suspicious lymph node metastasis, and bile duct invasion. The definitions of each CT feature are detailed in S1 Table. In cases where discrepancies arose among the reviewers’ interpretations, they re-evaluated the images collaboratively to reach a consensus.

4. Pathologic analysis and determination of American Joint Committee on Cancer TNM stage

ICCA was pathologically diagnosed using the surgical specimen, in accordance with the 2019 World Health Organization classification. Tumor size, number, microvascular or macrovascular invasion, visceral peritoneal invasion, and extrahepatic organ invasion were evaluated, and the pathologic T category was determined according to the eighth edition of the American Joint Committee on Cancer (AJCC) system (S2 Table) [20]. The presence of lymph node metastasis was also assessed (pN0 or pN1 category). In patients who did not undergo lymphadenectomy (Nx category), the following composite reference standards were used to determine the N category: the development of suspicious regional lymph nodes on follow-up imaging tests, such as CT, MRI, or positron emission tomography (PET), within 3 months after surgery was classified as N1; conversely, the absence of suspicious regional lymph nodes on follow-up imaging tests for at least 1 year after surgery was classified as N0.

5. Follow-up

After surgical resection, patients were regularly monitored through serum tumor markers and imaging studies, including contrast-enhanced CT or MRI, at intervals of 3 to 6 months, in accordance with each institution’s protocol. For patients with rising tumor marker levels and inconclusive imaging findings, additional PET-CT was performed. Recurrence was defined as any sign of recurrent ICCA, either pathologically-proven or suspected on CT/MRI (documented progression on serial imaging or confirmed hypermetabolism on additional PET/CT) with or without elevated CA 19-9 [21].

The primary endpoint of our study was the incidence of ER, defined as the occurrence of recurrence within 1 year after resection. While the definition of ER varied across studies, the majority of studies used 12 months as the cutoff for ER [15,16,22,23], and we adopted this definition of ER. Very early recurrence (VER) was defined as the occurrence of recurrence within 6 months after resection. Recurrence-free survival (RFS) was defined as the interval between surgical resection and either tumor recurrence or all-cause death, while OS was defined as the interval between resection and all-cause death.

6. Statistical analysis

Categorical variables were compared using chi-squared or Fisher’s exact tests and are summarized as counts with percentages. Continuous variables were compared using two-sample t-tests or Wilcoxon rank-sum tests and are expressed as means with standard deviations or medians with interquartile ranges (IQRs). Inter-reader agreements for imaging characteristics were evaluated using Fleiss’ κ statistics for categorical variables and the intraclass correlation coefficient for tumor size. A κ statistic of 0.8-1.0, 0.6-0.79, 0.40-0.59, 0.2-0.39, and 0-0.19 was considered to indicate excellent, good, moderate, fair, and poor agreement, respectively.

In the development cohort, univariable and multivariable logistic regression analyses were performed to identify preoperative factors independently associated with ER following surgical resection. For continuous variables, the optimal cutoff points were determined using the maximum Youden index from the receiver operating characteristic (ROC) curve for ER. Variables with a p-value < 0.05 in the univariable analysis were retained for the multivariable analysis. To assess the additional value of imaging features in predicting ER, we tested two models: Model 1, which utilized only preoperative clinical features, and Model 2, which incorporated both preoperative clinical and CT features. These risk models were developed based on the final step of the multivariable logistic regression analysis. Specifically, the β coefficients of the risk factors for ER identified in the final step of the respective multivariable logistic regression models were employed. The discrimination performance of the two models was quantified using the area under the ROC curve (AUC) and compared through DeLong’s test. Among the two risk models (model 1 vs. model 2), the preoperative model demonstrating superior predictive performance was selected, and the risk score was calculated by dividing each β by the smallest β. The performance of the preoperative risk scoring system was compared with that of the AJCC TNM staging system, which was based on postoperative pathological results. The calibration abilities were evaluated using a calibration plot of the predicted probabilities against the observed outcomes. Additionally, the optimal cutoffs for the preoperative risk scoring system were determined to stratify patients into low, intermediate, or high-risk categories for ER, utilizing the X-tile program [24]. Differences in RFS among these three patient groups were assessed using the Kaplan-Meier method with the log-rank test. In the external validation cohort, we compared the predictive performance of the preoperative risk scoring system with the AJCC TNM staging system, and we assessed the differences in RFS among the three risk groups: low, intermediate, and high risk. Additionally, the performance of preoperative risk scoring system for predicting VER was evaluated in both the development and validation cohorts.

Statistical analyses were performed using SPSS ver. 29.0 (IBM Corp.) or R ver. 4.3.2 (R Foundation for Statistical Computing), with p < 0.05 indicating statistical significance.

1. Patient characteristics

Of the 494 eligible patients, 129 were excluded for the following reasons: five with a history of other malignancies, seven with synchronous HCC, seven who had received prior ICCA treatment, 42 without preoperative contrast-enhanced CT, five with IPNB, 31 who underwent palliative resection, and 32 who either died or were lost to follow-up within a year without recurrence (Fig. 1). Consequently, the study included 365 patients (mean age, 62±10 years, 264 in the development cohort and 101 in the validation cohort). Clinical, pathological, and CT imaging characteristics of the development and validation cohorts are summarized in Table 1 and S3 Table. Compared to the validation cohort, the development cohort exhibited a higher proportion of patients who were positive for HBsAg (23.5% vs. 9.7%, p=0.010) and a lower proportion of those with centrally located tumors (22.0% vs. 32.7%, p=0.034). Also, there were significant differences in microvascular invasion (MVI; p < 0.001), lymph node metastasis (p=0.045), visceral peritoneum perforation (p=0.008), and TNM stage (p=0.015) between the development and validation cohorts. No significant differences were observed in other clinical, pathological, or CT imaging characteristics between the two cohorts.

Table 1.

Clinical and imaging characteristics of the developmental and validation cohorts

During a median follow-up of 36.8 months, 365 patients received follow-up CT (median exam number, 13; IQR, 7 to 24) and 118 received MRI (median exam number, 2; IQR, 1 to 4). Among the 365 patients, 240 experience recurrence: 96 (40.0%) were diagnosed with recurrence by confirming hypermetabolism at the suspected recurrence site on PET/CT in addition to CT/MRI examinations, 45 (18.8%) by pathological diagnosis, and 99 (41.2%) by confirming progression on serial imaging with or without elevated CA 19-9. Of the 240 patients with recurrence, 153 patients experienced ER, and 102 experienced VER. Among the 153 patients with ER, 44 (28.8%) had intrahepatic recurrence only, 68 (44.4%) had extrahepatic recurrence only, and 41 (26.8%) had both intrahepatic and extrahepatic recurrences. Patients without ER demonstrated significantly longer OS compared to those with ER (median OS, 63.0 vs. 15.7 months; p < 0.001) (S4 Fig.).

2. Development of a preoperative risk scoring system to predict ER

According to the maximum Youden index, the optimal cutoff values for serum CA 19-9 and CEA were 330 U/mL and 7.5 ng/mL, respectively. In the multivariable analysis using only clinical features (model 1), CA 19-9 (≥ 330 U/mL; odds ratio [OR], 3.44; 95% confidence interval [CI], 1.74 to 6.78; p < 0.001) and tumor differentiation (poorly differentiated; OR, 2.31; 95% CI, 1.26 to 4.21; p=0.006) were identified as independent predictors of ER after surgery (Table 2). In the multivariable analysis incorporating both clinical and CT features (model 2), CA 19-9 (OR, 2.88; 95% CI, 1.41 to 5.87; p=0.004), tumor differentiation (OR, 2.42; 95% CI, 1.26 to 4.64; p=0.008), central location (OR, 2.26; 95% CI, 1.12 to 4.53; p=0.022), AP peritumoral enhancement (OR, 2.28; 95% CI, 1.26 to 4.12; p=0.007), tumor multiplicity (OR, 5.52; 95% CI, 2.62 to 11.65; p < 0.001), and extrahepatic organ invasion (OR, 5.13; 95% CI, 1.21 to 21.74; p=0.026) were identified as independent predictors of ER.

Table 2.

Univariable and multivariable analyses of predictors of early recurrence of ICCA in the development cohort

The inter-reader agreement for the four significant CT imaging features among the three reviewers ranged from fair to good (S5 Table). Among these features, the agreement for tumor location (κ=0.78), AP peritumoral enhancement (κ=0.63), and tumor multiplicity (κ=0.59) were moderate to good, whereas the agreement for extrahepatic organ invasion (κ=0.39) was fair. The overall percentage of agreement ranged from 74.2% to 87.5%.

Because Model 2 demonstrated significantly higher discrimination performance than Model 1 (AUC, 0.79 [95% CI, 0.73 to 0.84] vs. 0.68 [95% CI, 0.62 to 0.74]; p < 0.001), Model 2 was selected as the final predictive model and was subsequently translated into a risk scoring system. The risk score was calculated as follows: score sum=2 (CA 19-9 ≥ 330 U/mL)+1.5 (poorly differentiated)+1.5 (central location)+1.5 (AP peritumoral enhancement)+3 (tumor multiplicity)+3 (extrahepatic organ invasion) (Table 3).

Table 3.

Final preoperative risk scoring system for early recurrence of ICCA

3. Performance of the preoperative risk scoring system to predict ER

In the development cohort, the risk scoring system showed good discriminatory performance with an AUC of 0.78 (95% CI, 0.73 to 0.84), which was significantly higher than that of the TNM staging system (AUC, 0.68 [95% CI, 0.63 to 0.74]) (p=0.002). Calibration plots indicated that the predicted risk was well-aligned with the observed risk (S6 Fig.). Table 4 shows the predicted and observed risks for ER stratified by risk groups. Based on the risk score, patients were categorized into low-risk (risk score < 1.5), intermediate-risk (risk score 1.5-5), or high-risk (risk score ≥ 5) groups. The 1-year RFS rates in the low-, intermediate-, and high-risk groups were 85.5%, 56.6%, and 15.6%, respectively (p < 0.001) (Fig. 2A).

Table 4.

Predicted and observed risks for early recurrence of ICCA based on the risk scores

Fig. 2.

Kaplan-Meier curves of 1-year recurrence-free survival across risk groups stratified based on the preoperative risk scoring system in the development (A) and external validation (B) cohorts.

In the validation cohort, the AUC was 0.69 (95% CI, 0.58 to 0.79), which was slightly higher than that of the TNM staging system (AUC, 0.66 [95% CI, 0.59 to 0.73]) (p=0.641). The 1-year RFS rates in the low-, intermediate-, and highrisk groups were 87.5%, 58.5%, and 25.0%, respectively (p < 0.001) (Fig. 2B).

For predicting VER, the AUC of the risk scoring system was 0.77 (95% CI, 0.70 to 0.83) in the development cohort and 0.74 (95% CI, 0.61 to 0.86) in the validation cohort.