Proposed Modification of Staging for Distal Cholangiocarcinoma Based on the Lymph Node Ratio Using Korean Multicenter Database

The 8th American Joint Committee on Cancer (AJCC) staging system for distal cholangiocarcinoma (DCC) included a positive lymph node count (PLNC), but a comparison of the prognostic predictive power of PLNC and lymph node ratio (LNR) is still under debate. This study aimed to compare various staging models made by combining the abovementioned factors, identify the model with the best predictive power, and propose a modified staging system. We retrospectively reviewed 251 patients who underwent surgery for DCC at four centers. To determine the superiority of various staging models for predicting overall OSR, Akaike information criterion (AIC), Bayesian information criterion (BIC), AIC correction (AICc), and Harrell’s C-statistic were calculated. In multivariate analysis, age (p = 0.003), total lymph node count (p = 0.033), and revised T(LNR)M staging (p < 0.001) were identified as independent factors for overall survival rate. The predictive performance of revised T (LNR) M staging (AIC: 1288.925, BIC: 1303.377, AICc: 1291.52, and Harrell’s C statics: 0.667) was superior to other staging system. A modified staging system consisting of revised T category and LNR predicted better overall survival of DCC than AJCC 7th and AJCC 8th editions. In the future, external validation of the proposed new system using a larger cohort will be required.


Introduction
Cholangiocarcinoma (CCC) is a malignant disease that occurs along the biliary tract and is known to constitute 3% of all gastrointestinal cancers [1]. Distal cholangiocarcinoma (DCC) refers to CCC that occurs from the point where the cystic duct joins the common hepatic duct to the ampulla of Vater [2]. Previous studies have reported poor prognosis after surgical treatment of DCC [3,4]. The American Joint Committee on Cancer (AJCC) TNM classification is widely used as a staging system to predict this dismal prognosis of the disease. Recently, the 8th edition of the AJCC [5] has undergone several changes compared to the previous (7th) edition. First, the obscure anatomical landmark called bile duct wall invasion, which was the boundary separating T2 and T1, was discarded. Instead, the T category was categorized, using objective figures expressed as the depth of invasion (DOI) [6], which was defined as the distance from the basal lamina of the adjacent normal epithelium to the deepest infiltrating tumor cells. Applying this definition, the T category is subdivided as follows: T1 (DOI < 5 mm), T2 (DOI 5-12 mm), T3 (DOI > 12 mm), and T4 (tumor involves the celiac axis or the superior mesenteric artery). Second, the group with positive regional lymph nodes that were N1 in AJCC 7th edition was subdivided into two groups, according to the positive lymph node count (PLNC) in the 8th edition. As a result, metastases in 1-3 regional lymph nodes were classified as N1, and patients with four or more metastases as N2. Validation of this new AJCC 8th staging system was performed through a single institution retrospective study [7]. It was shown that the 8th edition can predict a patient's prognosis more accurately than the 7th edition.
At this point, however, there are some points to reconsider for the current staging system. First, a multicenter study of nearly 300 cohorts suggested cutoff values (≤ 3 mm, 3-10 mm, > 10 mm) for DOI different from the current system [8]. This cohort number is more than studies supporting the current cutoff value of the T stage [9,10]. As to N category, recent studies have reported that lymph node ratio (LNR), which is defined as the ratio of PLNC to total lymph node count (TLNC), is relatively superior to PLNC as an indicator for predicting prognosis of DCC [11][12][13]. Considering the debates about the adequacy of the T category and N category that make up the current staging system, we attempted to validate the prognostic predictive power of the AJCC 7th and 8th TNM classification in this study. In addition, the optimal cutoff value for overall survival rate (OSR) of T category, TLNC, PLNC, and LNR was also identified. Ultimately, we constructed a new TNM staging model consisting of each of the factors mentioned above and compared the prognostic predictive powers of the existing models to identify the best staging system.

Optimal Cutoff Value for the T Category
We compared the χ 2 score by moving from 10 to 15 mm, to obtain the cutoff value of the tumor invasion depth that best distinguishes T2 and T3. As a result, when T2 and T3 were separated based on 10 mm (T category-a), the χ 2 score was the highest ( Table 2). The new T category using this optimal cutoff value was defined as the 'Revised T category', which stratified T1, T2, and T3 as follows: T1 (<5 mm), T2 (5-10 mm), and T3 (>10 mm). In the 7th and 8th AJCC classifications, the survival curves between the T2 and T3 were not significantly separated (52.0% vs. 43.0%; p = 0.483 and 38.5% vs. 30.4%; p = 0.331) (Figure 1a,b). In the Revised T category, the five-year cumulative OSR of T2 and T3 showed marginally significant differences (40.6% vs. 26.9%; p = 0.087) (Figure 1c).

Optimal Cutoff Value for the TLNC, PLNC, and LNR
A stepwise algorithm was used to obtain optimal cutoff values for TLNC, PLNC, and LNR (Table 3). In the univariate analysis, the optimal value of TLNC (optimal TLNC) was found to have the highest χ 2 when the cutoff value was set to 13. A univariate cox proportional regression analysis was also performed on the other variables, to determine if the TLNC could affect the prognosis even when adjusted with other factors affecting OSR. In multivariate analysis with age, LVI, tumor differentiation, and revised T category, which were found to affect OSR in univariate Cox proportional regression analysis, TLNC ≥ 13 was significant as optimal cutoff (HR 0.597, 95% CI 0.411-0.866; p = 0.007).

Optimal Cutoff Value for the TLNC, PLNC, and LNR
A stepwise algorithm was used to obtain optimal cutoff values for TLNC, PLNC, and LNR (Table 3). In the univariate analysis, the optimal value of TLNC (optimal TLNC) was found to have the highest χ2 when the cutoff value was set to 13. A univariate cox proportional regression analysis was also performed on the other variables, to determine if the TLNC could affect the prognosis even when adjusted with other factors affecting OSR. In multivariate analysis with age, LVI, tumor differentiation, and revised T category, which were found to affect OSR in univariate Cox proportional regression analysis, TLNC ≥ 13 was significant as optimal cutoff (HR 0.597, 95% CI 0.411-0.866; p = 0.007).

Multivariable Analysis for OSR
We constructed a multivariate model that includes the factors described below, to identify the independent factors affecting OSR: age, sex, preoperative bile drainage, operation type, operating time, tumor size, LVI, PNI, tumor differentiation, margin status, AJCC 7th T category, AJCC 7th N category, AJCC 8th T category, AJCC 8th N category, revised T category, revised N category, optimal TLNC, LNR-c, AJCC 7th TNM staging, AJCC 8th TNM staging, revised TNM staging, and revised T(LNR-c)M staging. Multivariate analysis using this model showed that age (   AIC, BIC, AICc, and C-statistics values for the AJCC 7th TNM staging, AJCC 8th TNM staging, revised TNM staging, and revised T (LNR-c) M staging were summarized in Table 5. Among these, the revised T(LNR-c)M-staging model had the lowest AIC (1288.925), BIC (1303.377), and AICc (1291.592) values. In contrast, the model showed the highest C-statistic (0.667).

Discussion
To determine the best treatment modality, an accurate assessment of the patient's prognosis after surgery is essential. TNM staging is the most widely accepted and clinically used system in determining the prognosis of cancers occurring in almost all gastrointestinal tracts. Recently, the eighth edition of the AJCC TNM classification was published, and a retrospective single center study was conducted to validate the changed staging system for DCC [7]. In the study, the application of T category defined by DOI and N category with three-tier classification showed better separation of survival curves of each prognostic group than the conventional system. The concept of DOI stems from consideration of the unique anatomical structure and arrangement of components of the extrahepatic bile duct, which is different from other GI tracts [14]. Indeed, several studies have shown that this concept is clinically superior to existing T-stage criteria based on the obscure anatomical layer of extrahepatic bile duct [8][9][10]. The cutoff value that distinguishes T2 and T3 in this study was 10 mm, which is different from 12 mm in the 8th edition of AJCC. Nevertheless, this study reinforces the evidence that the concept of DOI can provide better prognostic predictive power in the T stage.
On the other hand, the debate over which of the two factors, PLNC or LNR, can predict the prognosis of DCC more accurately is still ongoing. PLNC simply refers to the number of locally metastatic lymph nodes, whereas LNR is PLNC divided by TLNC. Theoretically, the PLNC itself is more exposed to bias than LNR in that it can be influenced by the skill of the surgeon. In contrast, the findings supporting that LNR is not affected by TLNC have been reported through studies on other GI cancers [15][16][17]. For DCC, there are three single center retrospective studies that have shown that LNR actually affects prognosis independently. Kawai et al. demonstrated LNR > 0.2 affects poor prognosis of extrahepatic CCC after surgery. In the study, the enrolled cohort comprised 62 extrahepatic CCC, except for hilar CCC [18]. A cutoff value of LNR > 0.2 was also reported in another study by Oshiro et al. as an independent factor predicting poor prognosis in 60 patients with extrahepatic CCC [11]. Zhang et al. identified a four-tier classification of LNR (0 vs. 0-0.2 vs. 0.2-0.5 vs. > 0.5), which further refines the group with LNR > 0.2, as a significant risk factor in multivariate analysis [13]. In the current study, however, the three-tier classification of LNR (0 vs. > 0 to < 0.1 vs. ≥ 0.1), which was based on cutoff value = 0.1, was found to best divide the survival curve. In fact, the cutoff values of the three studies mentioned above omit the statistical effort to find the highest χ 2 score by using the stepwise algorithm, as in current study. In addition, their LNR > 0.2 values used the same values previously suggested in studies of pancreatic cancer and ampulla of Vater cancer [19,20]. Given these points, it is questionable whether LNR > 0.2 is an appropriate cutoff value for LNR in predicting the prognosis of DCC. Further research will be needed to find the appropriate cutoff value of LNR, using refined large-scale cohorts and careful statistical analysis.
One noteworthy aspect of this study is that TLNC is found to be an independent factor in predicting survival for DCC. There has been one study showing that TLNC affects DCC survival in univariate analysis [18], but there have been no reports of significant impacts on multivariate analysis. In this study, the optimal cutoff value was TLNC ≥ 13, which is a comparable result compared to TLNC ≥ 10-11 suggested in other studies [18,21,22]. Higher TLNCs can escape the risk of under-staging by increasing the accuracy of staging for N categories. It can also be expected to ensure a better prognosis by reducing the likelihood of metastatic lymph nodes that may remain. Despite the criticism that TLNC may be influenced by the surgical policy of each institution and specimen handling by pathologists, achieving adequate TLNC is expected to enable accurate cancer staging and provide clinical benefits for the patient's postoperative outcomes.
The staging model proposed in this study is meaningful in that it introduces LNR instead of the existing N category. As far as we know, this is the first study to compare the predictive power of the AJCC TNM staging system for DCC and the new staging system including LNR. We assumed that the LNR model predicted the prognosis better than the existing staging model composed of PLNC. Using sophisticated statistical methods, we actually demonstrated that revised T (LNR-c) M staging using LNR can predict the prognosis of DCC more accurately than other staging models. The value of 0.667, Harrell's C-statistics of the newly revised T (LNR-c) M-staging model, was acceptable.
This study has several limitations. First, there is a flaw in the research design itself, which is analyzed by retrospective data collected from multicenters for about 10 years. Changes in the operator's proficiency over time could not rule out a bias in the study. In addition, although the specimens were reviewed to measure DOI of tumor according to the AJCC 8th edition, there may have been some differences in specimen handling in each hospital. Second, despite the multicenter research design, there is a limitation that almost all of the data are from a single tertiary institution. Thus, validation of the new staging system must be performed in the future, using large data from other tertiary institutions.

Patients and Data Collection
Using the multicenter database, we identified 251 patients who underwent surgery for DCC during 2002-2012. Four Korean hospitals provided data on DCC based on medical records. Data from the following institute was retrospectively analyzed: Samsung Medical Center (n = 172), Seoul National University Boramae Hospital (n = 40), Inje University Ilsan Paik Hospital (n = 26), and Dongguk University Ilsan Hospital (n = 13). Of these patients, we excluded 16 (6.4%) patients, including 10 with no follow up data, 4 with stage of Tis, 1 with distant metastasis, and 1 with no records of total lymph node count (TLNC). The remaining 235 patients were enrolled for analysis in this study.
The analyzed variables were age, sex, preoperative bile drainage, operation type, operating time, tumor size, lymphovascular invasion (LVI), perineural invasion (PNI), tumor differentiation, margin status, tumor invasion depth, TLNC, positive lymph node count (PLNC), lymph node ratio (LNR), and tumor stage by the AJCC 7th [23]. We performed restaging to validate the 8th AJCC classification. For this purpose, the resected specimens were reviewed, and DOI was numerically analyzed. In the case of the lymph node, it was restaged by using PLNC recorded in the pathologic report. Regional lymph nodes were defined based on the Japanese system [24]. Accordingly, the range of lymph node resections performed on the same criteria in four centers was as follows: in the hepatoduodenal ligament (no. 12); anterior/posterior aspect of the pancreas and duodenum (no. 13 and 17); around the stomach (no. 1-6); around the common hepatic artery (no. 8); and along the superior mesenteric artery (station 14). PLNC was defined as the number of lymph nodes with metastases in the obtained lymph nodes, and TLNC was defined as the number of total lymph nodes found by pathologic examination in specimens. LNR was defined as PLNC divided by TLNC. We defined the overall survival time as the period from the first diagnosis to death or follow-up loss. All subjects gave their informed consent for inclusion before they participated in the study. The study was conducted in accordance with the Declaration of Helsinki, and the protocol was approved by the Ethics Committee of Institutional Review Board of Samsung Medical Center, Seoul, Republic of Korea (2020-01-126).

Statistical Analysis
The Kaplan-Meier method and log rank test were used to compare the overall survival rate (OSR) for each category of the AJCC 7th and AJCC 8th classifications. A stepwise algorithm is commonly used to obtain optimal cutoff values for T category, PLNC, TLNC, and LNR. In detail, the optimal cut off value for the T category was determined by the highest χ 2 score in the univariate Cox proportional regression methods. In the case of TLNC, PLNC, and LNR, the optimal cutoff value was considered to be statistically significant in multivariate analysis with the highest χ 2 score in the univariate cox progression model. The backward selection method of Cox proportional regression analysis was performed to identify independent factors affecting OSR.
We constructed new staging models by using this optimal cutoff value of T category, PLNC, and LNR. The predictive powers of these models and the existing staging modes (AJCC 7th and AJCC 8th editions) were compared, using the following statistical methods: Akaike information criterion [AIC = −2 log maximum likelihood + 2× degrees of freedom (df)] [25], Bayesian information criterion (BIC = −2 log maximum likelihood + log (sample size) × df) [26], corrected Akaike information criterion (AICc = −2 log maximum likelihood + [2 × n × (df + 1)/(n − df − 2)]) [27], and Harrell's C-statistic [28]. In this study, higher values of AIC, BIC, and AICc and lower values of C-statistics were interpreted as better prediction models. We considered values to be statistically significant when the p-value was less than 0.05. The PASW Statistics version 23.0 (SPSS, IBM corp., Armonk, NY, USA) was used for all the abovementioned statistical analyses.

Conclusions
Modified staging system consisting of revised T category (T1: < 5 mm, T2: 5-10 mm, and T3: > 10 mm) and LNR ≥ 0.1 predicted better overall survival of DCC than the AJCC 7th and AJCC 8th editions. In the future, external validation of the proposed new system, using a larger cohort, will be required.