Value of the log odds of positive lymph nodes for prognostic assessment of colon mucinous adenocarcinoma: Analysis and external validation

Abstract Purpose To evaluate the impact of the log odds of positive lymph nodes (LODDS) on cancer‐specific survival (CSS) in colon mucinous adenocarcinoma (MAC) patients, compared with pN stage and the lymph nodes ratio (LNR). Methods A total of 10,182 colon MAC patients from the Surveillance, Epidemiology, and End Results database were divided into the training group. The external validation group included 153 patients from Fujian Medical University Union Hospital. The Cox regression method was used to identify prognostic risk factors. Nomograms were evaluated by Harrell's concordance index (C‐index) and calibration curves. Recursive partitioning analysis (RPA) was used to develop a novel staging system. Results Time‐dependent receiver operating characteristic curves (ROC) to predict CSS showed the areas under the ROC curve of LODDS were always higher than pN stage and LNR. LNR and LODDS classifications can well distinguish the prognosis of patients with the same pN stage. Cox analyses indicated that age, tumor size, pT stage, pN stage, LNR, and LODDS were independent predictors of CSS (p < 0.05). Based on three lymph nodes classifications, we constructed three prognostic nomograms models for CSS. The C‐index of the pN, LNR, and LODDS classification nomograms were 0.746 (95% confidence interval [95% CI]: 0.736–0.756), 0.750 (95% CI: 0.740–0.760), and 0.758 (95% CI: 0.748–0.768), respectively. In external validation, we observed the C‐index of LODDS classification nomograms was 0.787 (95% CI: 0.648–0.926). RPA stage, including four stages, was constructed successfully based on pT stage and LNR or LODDS, respectively. The 3‐, 5‐, and 8‐year areas under the ROC curve of LNR‐RPA stage and LODDS‐RPA stage were superior to tumor‐node‐metastasis stage. Conclusion LODDS to be a better prognostic factor of CSS for colon MAC patients than pN stage and LNR. A nomogram and RPA stage base on LODDS can provide accurate information for personalized cancer treatment.


| INTRODUCTION
Worldwide, colorectal cancer remains a common malignant tumor and is a dominant cause of cancer-specific mortality. 1,2 In the context of personalized medicine, patient management based on histological types is required. Colon mucinous adenocarcinoma (MAC), a particular histological subtype of colon cancer, is characterized by mucinous components that exceed 50% of tumor tissue components. 3,4 MAC accounts for about 10%-15% of total colorectal cancer cases. 5 At diagnosis, many MAC patients present with advanced stage and regional lymph node metastasis. [6][7][8] However, the prognostic significance of MAC still is controversial. Multiple studies indicate that MAC subtype patients have a dismal survival. [9][10][11] In contrast, a few studies have demonstrated that MAC does not affect survival outcomes compared to the adenocarcinoma subtype. [12][13][14] Hence, given this scenario, a more suitable and accurate prognostic model is explicitly warranted for use in colon MAC patients.
Lymph node metastasis is an essential driver of clinical outcomes. The American Joint Committee on Cancer (AJCC) tumor-node-metastasis (TNM) classification is based on the number of positive regional lymph nodes (PLN). Nevertheless, patients with the same TNM stage remain heterogeneous. Recently, the lymph nodes ratio (LNR) and the log odds of positive lymph nodes (LODDS) are applied to clinical management and survival prediction in multiple tumors, which demonstrate distinct superiority. [15][16][17][18] Although some studies have reported the prognostic value of LNR and LODDS in colon patients. [19][20][21] However, these studies are often directed against all histological types in colon cancer. Meanwhile, some of these lacked external validations. Therefore, the role of LNR and LODDS in MAC patients is still unclear.
In the present study, we enrolled colon MAC patients from the Surveillance, Epidemiology, and End Results (SEER) database, compared the ability of AJCC (pN stage) classification, LNR classification, and LODDS classification to predictive survival. Nomogram prognostic model was constructed base on independent prognostic factors. Furthermore, we also validated this model using patients from an external cohort.

| Study cohort
Colon MAC patients retrieved from the SEER database from 2004 to 2015 were assigned to the training group using SEER*Stat software (www.seer.cancer.gov/seerstat).
The external validation group consisted of colon MAC patients from Fujian Medical University Union Hospital (FJMUUH) between January 2008 to September 2017. All participants meet the following inclusion and exclusion criteria. Specifically, details of inclusion criteria were: (1) pathologically confirmed MAC based on International Classification of Diseases of oncology (ICD-O-3); (2) underwent colon resection and regional lymph node dissection; (3) AJCC TNM stage I-III. Details of exclusion criteria were: (1) distant metastasis at the time of diagnosis or during surgery; (2) complicated with another primary tumor; (3) incomplete follow-up data; or (4) incomplete tumor clinicopathological features data (including pT status, pN status, tumor size). Ultimately, 10,182 patients in the SEER cohort and 153 patients in the FJMUUH cohort were included. The study was approved by the Hospital Ethics Committee. Given that the SEER database is publicly accessible, informed consent in training does not require. In the external validation group, informed consent was exempted considering this retrospective design.
We retrieved clinicopathological features information from the SEER database, including age at diagnosis, sex, race, tumor site, tumor size, pT stage, pN stage, the number of total regional lymph nodes (TLN) examined, and the number of positive regional lymph nodes. Cecum, ascending colon, hepatic flexure of colon, and transverse colon were classed as proximal colon, while splenic flexure of colon, descending colon, and sigmoid colon were categorized as distal colon. And the pN status is reclassified according to the 8th edition of the AJCC staging system (pN status: pN0, pN1a, pN1b, pN2a, pN2b). Colon cancerspecific survival (CSS) was used as the primary endpoint.

| LNR classification
The ratio of PLN/the number of TLN was calculated as LNR. Aside from LNR = 0 and LNR = 1, the value of PLN/TLN was partitioned into 10 groups by 0.1 intervals. Kaplan-Meier (KM) analysis was performed to compare the survival differences between two neighboring groups. Then, 12 LNR subgroups were regrouped according to similar CSS.

| LODDS classification
Log[(PLN + 0.5)/(TLN -PLN + 0.5)] was counted as LODDS. The addition of a value of 0.5 to the numerator and denominator in the formula is to avoid singularity. 22

K E Y W O R D S
colon mucinous adenocarcinoma, LODDS, nomograms, prognosis, RPA stage The classification of LODDS subgroups was similar in design to the LNR subgroups with 0.5 intervals.

| Comparison among three lymph nodes classifications and prognostic model
A time-dependent receiver operating characteristic (ROC) plot was performed, and the areas under the ROC curve (AUC) were counted to compare the predictive ability of three lymph nodes classifications. A univariate Cox proportional hazards model was performed to determine the prognostic values factor for CSS. The median tumor size was set as the cutoff point. The pN stage, LNR, LODDS were included in multivariate analysis, respectively. Finally, three nomogram models were constructed base on independent prognostic factors. The Harrell's concordance index (C-index) and calibration curve were applied to evaluate the predictive ability of the three models. FJMUUH cohort was used for external validation of the nomogram model.

| Statistical analysis
All the statistical analyses were carried out on SPSS (versions 22.0) and R software (versions 3.6.3). Difference testing between SEER cohort and FJMUUH cohort was assessed by student t-test or Chi-squared test, when appropriate. The cox regression was used for univariate or multivariate survival analyses. The "survival ROC" package and "rms" package were used to generate time-ROC plots and nomograms, respectively. A novel tumor stage was reclassified using recursive partitioning analysis, which can be accessed from the online website (http://rpa.renlab.org). 23 p < 0.05 was considered statistically significant.

| Clinicopathological features of patients
There were 10,182 and 153 patients in the SEER cohort and FJMUUH cohort, respectively. Table 1 illustrates two cohort baselines. Of the SEER cohort, the mean age was 68.7 (±14.5) years. Most of them (84.0%) had advanced pT stage (pT3 or pT4), and 4147 (40.7%) patients had positive lymph nodes. A majority of tumors (76.3%) were located in the proximal colon. Of the FJMUUH, the mean age was 59.0 (±13.9) years. Similar to the SEER cohort, there were 145 (94.8%) and 70 (45.8%) patients who had advanced pT stage and positive lymph nodes. The tumor size in FJMUUH cohort was larger than that in SEER cohort significantly (p = 0.012).

| LODDS classification
Next, we also grouped LODDS subgroups according to similar CSS. Detailed data are available in Table 3. Concretely, for LODDS classification: LODDS1  Figure 1B).

| Comparison among three lymph nodes classifications
To further demonstrate predictive capabilities of pN stage, LNR classifications, and LODDS classifications, time-ROC were drawn, as shown in Figure 2. AUC of LODDS classifications was consistently outperformed pN stage and LNR classifications, which demonstrates the dominance of LODDS classifications in predicting CSS. Moreover, we contrasted CSS of patients based on three lymph nodes classifications. As is evident from Tables 4 and 5 and Figure 3, LNR and LODDS classifications can significantly distinguish the survival outcomes of patients with the same pN stage. However, this approach is not applicable for LNR classification in pN0 stage or LNR = 1. Patients with the same LNR and LODDS classifications presented a similar prognosis. Further, LODDS classification showed more remarkable prediction ability for the patient with LNR 1, 4, 5. As shown here (Figure 4), significant differences in CSS between LNR subgroups and LODDS subgroup were observed in TNM pIII stage, while

| Prognostic factors of CSS
Univariate and multivariate Cox regression, as shown in Table 6, were performed to determine prognostic factors of CSS. On univariate analyses, colon CSS was significantly associated with age, tumor size, pT stage, pN stage, LNR, and LODDS (all p < 0.05). Next, pT stage, LNR, LODDS were individually included in three multivariate analyses.
On multivariate analyses base on three independent lymph nodes classifications, there were significant correlations between CSS and pN stage (p < 0.001), LNR (p < 0.001), and LODDS (p < 0.001). Beyond this, age, tumor size, and pT stage were independent prognostic factors affecting CSS.

| Development and validation nomograms model
Based on the multivariate analysis results, we constructed three nomograms at the base of three lymph nodes classifications, respectively, as shown in Figure 5.  Figure 6). Furthermore, we validated LODDS classification nomograms using MAC patients from the FJMUUH cohort. The C-index was 0.787 (95% CI: 0.648-0.926). The calibration curves predicting the 5-year CSS, as shown in Figure 7, showed excellent agreement for nomograms.

| Novel staging system based on recursive partitioning analysis
We develop a novel staging system using recursive partitioning analysis based on LNR classification ( Figure 8A

| DISCUSSION
In this study, we compared the predictive power of three lymph node classifications using a large population from the SEER database. LODDS showed great potential to distinguish colon MAC patients with differing clinical prognoses. This classification can distinguish different prognoses for patients with the same pN stage. As shown by time-ROC, the predictive capability of LODDS classifications always outperforms pN stage and LNR classifications. Cox regression analysis revealed that three lymph node classifications (pN stage, LNR, LODDS) were significantly associated with CSS. Based upon the results, we constructed the nomograms prediction model, and have further confirmed the performance of this model using an external cohort. It can provide constructive information on prognosis to clinicians. The histological type of cancer potentially correlates with biological properties and clinical outcomes. MAC is a relatively rare subtype that commonly manifests as an advanced stage and lymph node dissemination. Indeed, the effect of MAC on the survival of colon MAC  Ref. Ref. Ref.
>5.5 Ref. (Continues) patients remains debatable. Several studies have demonstrated that the MAC subtype was associated with a worse prognosis. 6,[9][10][11][24][25][26] In a population-based survival study, MAC had a poor outcome because of advanced disease. 27 This association has been confirmed in a large metaanalysis. 28 This study found that MAC portends worse survival. There was, however, a different result in other studies. They found that no differences between MAC and CA were observed in outcomes after correction for the AJCC stage. [12][13][14]29 Those findings reflected potential heterogeneity in the prognosis of MAC patients with the same pN stage. As one of the most dominant factors in affecting the survival outcomes, the lymph nodes status and lymph node staging can provide critical information. Thus, identifying and optimizing the prognosis risk factors and guidance personalized treatment and surveillance become important for colon MAC patients.
Lymph nodes ratio has proved to be the better approach to lymph nodes classification in a variety of cancers. [30][31][32] A number of studies demonstrated LNR had prognostic value in patients with colon cancer. [33][34][35][36] Our findings in colon MAC patients were similar to previous studies. Specifically, in the present study, we observed that LNR classification could identify relatively good-prognosis patients in the advanced pN stage (e.g., pN2a and pN2b stage) and poorprognosis patients in the early pN stage (e.g., pN1a stage).
The accurately discriminating prognosis was of significance in personalized treatment, and this approach enabled patients with advanced disease to have the confidence to active treatment. Whereas, unfortunately, the performance of LNR classification was limited for nodenegative patients. 37 This limited the application and promotion of LNR classification.
In recent years, LODDS was introduced into cancer prognostic research. Benefiting from the unique computational approach, compared with LNR, LODDS largely circumvents the limitation of the status of negative lymph nodes and improves prognostic accuracy. In LODDS classification, the prognosis of those patients was further stratified by the number of negative lymph nodes, particularly in distinct differences in the number of retrieved. However, the outcomes of patients with negative lymph nodes were thought to be similar in LNR classification. On the other hand, our results revealed, there are differential outcomes between LODDS subgroups in the same LNR classifications. It implied LNR classification might be unable to discriminate well the survival of positive lymph nodes patients with the same LNR value but different total lymph node dissections. LODDS was significantly associated with overall survival as published in the study by Wang et al. 38 However, their studies were limited to stage III colon cancer, while the role of LODDS in stage I/II colon cancer was not investigated. 19  were also suggested the value of LODDS was in prognosis prediction of non-metastatic colon cancer. Indeed, this is a generalized observation for all histology types of colon cancer. Those results may not satisfy the need for personalized cancer therapies. Meanwhile, the lack of external validation limited the generalizability of those studies. In In the present study, we regroup MAC patients based on the value of LODDS with 0.5 intervals. Lastly, six subgroups were generated with significant differences in CSS between adjacent groups. In either pN stage, differences in survival among LODDS subgroups were observed. This implies that MAC patients with the same pN stage were heterogeneous, and LODDS classification can help clinicians identify different prognoses and develop personalized treatment and follow-up strategies. Meanwhile, LODDS classification can aid patients who with lymph node metastases but the early LODDS subgroup enhances confidence in cancer therapy. However, for patients with pI stage, there were no differences in CSS in the LODDS subgroup, which was possibly due to early disease. Furthermore, Cox regression analysis LODDS was an independent risk factor of CSS. We further developed a nomograms model for three lymph nodes classifications, and LODDS classification nomograms exhibited the best performance toward prognostic stratification, compared with the AJCC stage and LNR. Meanwhile, we validated nomograms using an external validation cohort and demonstrated stable performance. Additionally, we constructed a new staging system based on pT stage and LODDS. In RPA stage, a better survival was observed in node-positive MAC patients with early LODDS stage, which was clinically significant.
There are limitations in the present study. First, despite as a population-based database, SEER lacked some tumorrelated (e.g., vascular invasion) and treatment-related information (e.g., the details of radiotherapy and chemotherapy), and we cannot adjust potential confounders. Second, the inability to subdivide pT4a and pT4b status in the pT4 stage in patients diagnosed before 2010 from the SEER database, which made more accurate analyses difficult. Third, the number of external validation cohorts were relatively limited. As a consequence, subsequent studies should recruit a large sample and perform multicenter studies to further confirm and generalize our results.

| CONCLUSION
In summary, we proved the superiority of LODDS in prognostic stratification for colon MAC patients compared with pN stage and LNR. LODDS classification nomograms and RPA stage can provide stable assessments of patient clinical outcomes and contributed to personalized cancer treatment.