Log Odds of Positive Lymph Node- (LODDS-) Based Competing- Risk Nomogram for Predicting Prognosis of Resected Rectal Cancer: A Development and Validation Study

Background and Aims. Cancer-specific survival (CSS) of rectal cancer (RC) is associated with several factors. We aimed to build an efficient competing-risk nomogram based on log odds of positive lymph nodes (LODDS) to predict RC survival.Methods. Medical records of 8754 patients were collected from the Surveillance, Epidemiology, and End Results (SEER) database, of 4895 patients from SEER during 2011–2014 and of 478 patients from an Eastern center as a development cohort, validation cohort, and test cohort, respectively. Univariate and multivariate competing-risk analyses were performed to build competing-risk nomogram for predicting the CSS of RC patients. Prediction efficacy was evaluated and compared with reference to the 8 TNM classification using the factor areas under the receiver operating characteristic curve (AUC) and Brier score. Results. The competing-risk nomogram was based on 6 variables: size, M stage, LODDS, T stage, grade, and age. The competing-risk nomogram showed a higher AUC value in predicting the 5-year death rate due to RC than the 8 TNM stage in the development cohort (0.81 vs. 0.76), validation cohort (0.85 vs. 0.82), and test cohort (0.71 vs. 0.66). The competing-risk nomogram also showed a higher Brier score in predicting the 5-year death rate due to RC than the 8 TNM stage in the development cohort (0.120 vs. 0.127), validation cohort (0.123 vs. 0.128), and test cohort (0.202 vs. 0.226). Conclusion. We developed and validated a competing-risk nomogram for RC death, which could provide the probability of survival averting competing risk to facilitate clinical decision-making.


Introduction
Rectal cancer (RC) is one of the most commonly diagnosed cancers worldwide [1]. Past studies have suggested that the prognosis of RC may be influenced by several factors [2][3][4], such as gender, age, carcinoembryonic antigen (CEA), location, pT, pN, tumor size, lymph node (LN) metastasis, concomitant diseases, surgical procedure, residual cancer, and complications. The involvement of a vari-ety of factors has added to the complexity of the predicting model. The survival of RC may also be contended by several risks. The competing-risk nomogram model reduces the risk of influencing factors, which gives it a good predicting accuracy [5].
Currently, the American Joint Committee on Cancer (AJCC) 8 th N staging is the most widely used staging platform for predicting RC survival [6]. However, accurate N staging relies on sufficient histological examination of at least 12 LNs. A lesser number of examined LNs could lead to staging migration and the mistake of indulging in further treatment. Log odds of positive lymph nodes (LODDS) represent the logarithm of retrieving the LN ratio between the probability of positive and negative LNs, which signifies greater predicting ability relative to that by the conventional LN staging in gastrointestinal cancers [7][8][9].
In this study, we developed and validated a LODDSbased competing-risk nomogram using a large populationbased cohort and tested the nomogram in another independent cohort.

Study Design and Data Collection.
We performed this study based on the incidence and survival data obtained from the Surveillance, Epidemiology, and End Results (SEER 18) database and the Shanghai Tongren Hospital. We collected medical records of resected RC patients of age > 18 years with a known number of histologically examined LN and histologically positive LN. Patients with other tumors or those without histological confirmation were excluded from the study. We also excluded patients with missing data for development and validation of the model, inclusive of unknown gender, Patients whose data were collected from the SEER database during 2004-2010 were included in the development cohort, while those whose data were collected during 2011-2014 were included in the validation cohort. Patients diagnosed with RC during 2008-2013 were retrieved as the test cohort. We collected patients' clinicopathological variables including age, gender, tumor location, TNM staging, grade of tumor, histological grade, number of examined LNs, number of positive LNs, tumor size, and survival months. LODDS was defined as loge½ðpN + 0:5Þ/ðnN + 0:5Þ, where pN is the number of positive lymph nodes and nN is the number of negative nodes retrieved [10].

Statistical Analysis.
In the competitive-risk analysis, we defined RC-specific death as the primary outcome and other causes of death as competing events. Univariate and multivariate proportional subdistribution hazard models were applied to calculate the adjusted subdistribution hazard ratio (SHR) for each included variable. Variables with P < 0:1 in the univariate analysis and P < 0:05 in the initial multivariate analysis were finally included to develop the competing-risk nomogram. We presented the model visible as a nomogram based on Fine and Gray's model [11].
This model was internally developed, externally validated, and then tested in 3 independent cohorts. A calibration plot was employed to compare the predicted to observed probabilities of death at a certain point of time [12]. Pairs of predicted and observed probabilities lay on a 45-degree angle line if both the probabilities matched well and proved the model ideal. We then compared this model to the 8 th TNM staging by several assessment methods. The area under the curve (AUC) and higher Brier score were used to assess discrimination of the competing-risk nomogram [13][14][15]. All statistical analyses were performed by R version 3.3.3 software (Institute for Statistics and Mathematics, Vienna, Austria; http://www.r-project.org).
The RC cause-specific death predicting model of nomogram was established based on selected prognostic factors (Figure 1). The nomogram revealed that LODDS was the most important risk factor for RC prognosis, followed by T stage, M stage, age, size, and grade. Each subtype of the variables was assigned a specific score. A straight line to determine the estimated probability of survival was drawn at each time point on the total point scale, according to the total point.

Evaluation of the Competing-Risk
Nomogram. The calibration curves given in (Figure 2) show the dots close to a 45°diagonal line. The figure also suggests that the competing-risk nomogram showed good calibration for predicting 1-year, 3-year, and 5-year RC death in the development, validation, and test cohorts, respectively.
We used both AUC and Brier score to evaluate the diagnosis value and accuracy of the nomogram model ( Table 3). The AUC value was greater in the nomogram model than in the conventional TNM staging method for the development cohort

Discussion
As one of the most common cancer, RC has received much attention in both the Western and Eastern countries [1,16]. The prediction of prognosis is critically important for accurate clinical decision-making. To the best of our knowledge, this is the first study to develop a competing-risk nomogram for predicting RC death. In this study, we developed and validated a new competing-risk nomogram model based on a large Western cohort from SEER, which we then tested on an Eastern cohort. We found that this model demonstrated greater clinical predicting ability than the 8 th TNM staging approach performed conventionally in RC patients.
The new 8 th AJCC TNM staging stresses the personalization of the predicting classification [6]. Thus, the more effective factors of cancer are paid more attention [2,4]. In a prior study, gender, age, CEA, location, pT, pN, size, LN metastasis, concomitant diseases, surgical procedure, residual cancer, and complications were considered factors that could affect the prognosis of RC [2,4,17]. In fact, several factors are possibly associated with many other death causes, such as other cancers, concomitant diseases, complications, and accidents. As these causes compete with the RC cause-specific death, it adds bias to prognosis prediction. For accurately predicting the RC cause-specific deaths, we applied the competing-risk nomogram to reduce the potential bias [11]. The nomogram showed that LODDS was the most impactful contributing factor for prognosis, followed by T stage, M stage, age, size, Moreover, in the past, certain models had been developed and validated in the Western cohorts, which did not show clinical applicability in Eastern cohorts. In this study, we also tested this model in an Eastern cohort, which increased the scope of application. LN is an important prognostic factor for overall survival. The AJCC N classification is widely used in clinical practice [18], although its accuracy remains debatable in gastrointestinal cancers [7,8,19]. This staging approach is based on the absolute number of positive LNs, which in turn relies on sufficient numbers of examined LNs to fully assess the metastatic status. AJCC recommends at least 12 LNs, while some study recommends even more [20][21][22]. This staging approach involves complex functionalities for surgeons and pathologists and can hence result in migration. On the other hand, LODDS is a logarithm of the probability between being a positive or negative LN when it is retrieved [7]. It has been applied in gastrointestinal cancers to improve the accuracy of LN staging for predicting the prognosis [7][8][9][10]. In this study, we collected the LODDS of each patient and performed univariate and multivariate analyses to select factors for the development of a competing-risk nomogram. The resultant nomogram indicated that LODDS is the most impactful contributing factor of prognosis.
We identified some limitations in this study. First, this is a retrospective trial based on the SEER database and an Eastern cohort data, which may have caused selection bias. The missing data could also influence the analysis of prognosis and the development of the model. The large number of this trial may reduce this bias. Second, AJCC pays attention to the new therapy and biomarkers, which could influence the 7 Gastroenterology Research and Practice prognosis of RC. However, as the current database did not contain this information, we expect more innovations and trials of this field in the future.

Conclusion
In conclusion, we developed a new competing-risk nomogram based on LODDS to predict the prognosis of RC. This model reduced the bias of competing with other death causes and staging migration of the conventional N staging. The new staging system shows greater prediction accuracy and better ability than the conventional 8 th TNM staging approach for both the Eastern and Western cohorts.

Data Availability
The raw data used to support the findings of this study are included within the article.

Ethical Approval
This study was approved by the Ethics Committee of the Shanghai Tongren Hospital Affiliated to Shanghai Jiao Tong University School of Medicine.