Predicting the Survival Probability of Neuroendocrine Tumor Populations: Developing and Evaluating a New Predictive Nomogram

Purpose The purpose of this study was to develop and initially validate a nomogram model in order to predict the 3-year and 5-year survival rates of neuroendocrine tumor patients. Methods Accordingly, 348 neuroendocrine tumor patients were enrolled as study objects, of which 244 (70%) patients were included in the training set to establish the nomogram model, while 104 (30%) patients were included in the validation set to verify the robustness of the model. First, the variables related to the survival rate were determined by univariable analysis. In addition, variables that were sufficiently significant were selected for constructing the nomogram model. Furthermore, the concordance index (C-index), receiver operating characteristic (ROC), and calibration curve analysis were used to evaluate the performance of the proposed nomogram model. The survival analysis was then used to evaluate the return to survival probability as well as the indicators of constructing the nomogram model. Results According to the multivariable analysis, lymphatic metastasis, international normalized ratio (INR), prothrombin time (PT), tumor differentiation, and the number of tumor metastases were found to be independent predictors of survival rate. Moreover, the C-index results demonstrated that the model was robust in both the training set (0.891) and validation set (0.804). In addition, the ROC results further verified the robustness of the model either in the training set (AUC = 0.823) or training set (AUC = 0.768). Furthermore, the calibration curve results showed that the model can be used to predict the 3-year and 5-year survival probability of neuroendocrine tumor patients. Meaningfully, five variables were found: lymphatic metastasis (p = 0.0095), international standardized ratio (p = 0.024), prothrombin time (p = 0.0036), tumor differentiation (p = 0.0026), and the number of tumor metastases (p = 0.00096), which were all significantly related to the 3-year and 5-year survival probability of neuroendocrine tumor patients. Conclusion In summary, a nomogram model was constructed in this study based on five variables (lymphatic metastasis, international normalized ratio (INR), prothrombin time (PT), tumor differentiation, and number of tumor metastases), which was shown to predict the survival probability of patients with neuroendocrine tumors. Additionally, the proposed nomogram exhibited good ability in predicting survival probability, which may be easily adopted for clinical use.


Introduction
The endocrine system is composed of neuroendocrine cells that exist throughout the body. These cells can be found in isolation or form small aggregates from which neuroendocrine tumors arise [1]. From 1973 to 2004, the incidence of neuroendocrine tumors (NETs) reported in the United States increased significantly from 1.09/10 million to 5.25/10 million [2]. One of the important clinical features of neuroendocrine tumors is heterogeneity. Neuroendocrine tumors can be invasive or inert, and they may be asymptomatic or have symptoms that are completely different. This is because these tumors secrete different neuropeptides or biogenic amines. Additionally, neuroendocrine tumors may be nonfunctional or functional [3]. Due to their complex and heterogeneous biological behavior, the prognosis of patients with NETs remains unclear. Therefore, it is important to understand the biological and clinical characteristics of NETs and how they affect the prognosis of specific individuals.
Currently, the most common prediction systems for NETs are the American Joint Committee on Cancer (AJCC) and European Society of Neuroendocrine Tumors (ENETS) TNM staging system, which are based on the depth of tumor infiltration (T), nodules of metastatic lymph nodes (N), and distant metastasis (M) [1]. In addition, the World Health Organization (WHO) classification standard proposed in 2010 is commonly used, which divides neuroendocrine tumors into neuroendocrine tumors G1 (NET G1), neuroendocrine tumors G2 (NET G2), and neuroendocrine tumors G3 (NEC G3), as well as mixed adenocarcinoma and neuroendocrine carcinoma (MANEC). The main parameters used (Ki67 proliferation index and mitotic count) provide a possible prognostic treatment for this particular type of malignant tumor [4]. However, some limitations are present in these staging systems in predicting the prognosis of patients with NETs. For example, based on the ENETS staging system, the prognosis of patients with stage I is similar to that of patients with stage IIA [5]. In addition, some studies have shown that clinicopathological characteristics may have an impact on the prognosis of patients, such as gender, age, tumor location, and primary tumor diameter [6]. Therefore, it is important to determine other variables that predict the prognosis of NET patients and build a more robust prediction model.
Among the available decision-making tools, the nomogram is currently the best and most accurate method in predicting the prognosis of cancer patients. According to its statistical definition, the nomogram is an algorithm or graphical calculation encompassing the use of multiple continuous variables in order to predict specific endpoints, such as logistic or COX proportional hazard (PHs) regression models. The impact of these continuous variables on specific results is represented on an axis, and risk points are determined according to the prognostic importance of the variables of interest [7].
Previous studies have successfully quantified the risk of certain malignancy tumors by combining prognostic factors. In addition, various studies have shown that the nomogram can predict the prognosis of patients with gastrointestinal NETs [8] and pancreatic NETs [9]. However, to the best of our knowledge, nomograms have yet to be used to predict the prognosis of neuroendocrine patients. Accordingly, this study attempts to develop and validate a nomogram model in predicting the survival probability of NET patients.

Materials and Methods
2.1. Patient and Tumor characteristics. In this study, 348 neuroendocrine tumor patients who were admitted to the Cancer Hospital of Guangxi Medical University, from January 1, 2013 to November 12, 2019, were enrolled. All patients were proven to have NETs, according to the surgical specimen results, histological differentiation, or immunohistochemical staining of the biopsy. This study was approved by the Ethics Committee of Cancer Hospital of Guangxi Medical University, and all patients' written consent was obtained.
The data was collected retrospectively from the patients' electronic or physical medical records.
2.2. Postoperative Follow-Up. All patients were followed up in outpatient clinics or by phone until December 31, 2019. The interval of follow-up was three months in the first two years after treatment and six months from 2 to 5 years. The overall survival (OS) rate was defined as the time from the start of treatment to death or the last exposure for any reason. This index was often considered to be the best end point of clinical efficacy in cancer clinical trials. If there was a small increase in survival time, it can be regarded as meaningful evidence of clinical benefit.

Data Preprocessing and Statistical Features of
Clinicopathology. The mice function in the mice package (https://cran.r-project.org/package=mice) was used to assume that all variables were related in this study. In addition, multiple regression models were used for multiple interpolation to supplement missing values. A univariable analysis was performed using the coxph function in the survival package (https://cran.r-project.org/package=survival), and the survminer package (https://cran.r-project.org/package=survminer) was adopted to obtained the variables related to OS. Furthermore, the forestplot function in forestplot package was used to draw the forest plot. As shown in Figure 1, equidistant sampling was used. In regard to the data of live patients, 8 patients were sampled each year, and less than 8 patients were used. For that of dead patients, less than 8 patients were present for more than 4 years. Accordingly, 12 samples were extracted in the first year, while 8 samples were taken in other years.

Selection of Effective Feature Indicators.
The createData-Partition function in the caret package was used (https://cran .r-project.org/package=caret) to divide the training set and test set according to the survival state, of which the ratio was 7 : 3. There were 131 alive and 113 dead samples in the training set, while there were 58 alive and 46 dead samples in the test set. The variables with p < 0:05 were selected for multivariable analysis using the coxph function in the survival package and survminer package in R.

2.5.
Validating the Performance of the Nomogram Model. The prediction ability of the nomogram model was assessed both internally (training set) and externally (validation set) using three methods: the operator characteristic curve (ROC) analysis, concordance index (C-index), and calibration plot. C-index was an index of probability of the concordance between predicted and actual situations, ranging from 0.5 to 1.0. A calibration plot was also made, which was a graph consisting of two curves, a 45-degree straight line (perfect match) and an irregular curve (calibration curve). The operator characteristic curve (ROC) plot with area under curve (AUC) was then derived to assess the actual predictive significance of the nomogram model based on the pROC package [10]. C-index was obtained using the concordance.index function in the survcomp package [11]. Finally, the 2 BioMed Research International calibration plot was obtained via the calibrate function in the rms pakcage (https://cran.r-project.org/package=rms).

Statistical
Analysis. All data were analyzed using the R software and were expressed as mean ± standard deviation. Student's t-test was used to compare the mean between the two groups, and p < 0:05 was considered to be statistically significant.     (Tables 3 and 4). Subsequently, beta coefficients from multivariate models were used to develop the nomogram (Figure 3).

Nomogram Predictive Ability and External Verification.
The C-index results demonstrated that the model possessed a good ability in predicting the survival rate of NET patients in both the training set (C-index was 0.891) and test set (C-index is 0.804). In addition, the ROC results verified that the performance of the model was robust (training set: AUC value was 0.823. test set: AUC value was 0.768) (Figure 4). Additionally, as shown in Figure 5, the 3-year survival rate of 244 patients in the training set

Discussion
In order to assess the biological behavior of NETs, the histopathological criteria used for diagnosis and classification have been widely established. However, due to heterogeneity of NETs, the lack of a recognized NET staging system has hindered the provision of prognostic information for patients. Currently, the most commonly used NET staging  For example, the current AJCC staging system only distinguishes 5-year OS between phase I and phase II (84% vs 72%. p = 0:01) and between phase 2 and phase 3 (72% vs 65%. p = 0:97). As a result, the third and fourth stages (65% vs 55%. p = 0:36) cannot be effectively distinguished [12]. In addition, these staging systems only focus on the survival rate of patients with NETs at specific sites. The purpose of the present research was to develop and preliminarily validate a nomogram for patients with NETs in order to accurately predict individual survival rates following treatment. Compared to traditional staging systems, a nomogram may serve as a more accurate and clinically viable tool in predicting cancer outcomes. Therefore, it is important to establish a nomogram model to predict the prognosis of all patients with NETs. In this study, five variables (lymphatic metastasis, international normalized ratio (INR), prothrombin time (PT), tumor differentiation, and the number of tumor metastatic sites) were used to establish the nomogram model. Here, lymphatic metastasis was an independent prognostic factor related to patient survival; however, the impact of lymphatic metastasis on patient survival remains controversial. A multicenter study demonstrated that the presence or absence of lymphatic metastasis has nothing to do with liver metastasis or increased risk of survival [13]. In addition, a prospective single-institution study indicated that lymphatic metastasis had no significant effect on survival. Existing studies have considered the relationship between the number of lymphatic metastases and survival; however, only using two lymph nodes as the critical value is ambiguous and limited [14]. At the same time, numerous other studies have shown that a major feature of neuroendocrine tumors is its tendency to metastasize to the lymphatic system [15]. Lymphatic metastasis reduced the survival rate and is closely related to mortality; its prognostic significance is similar to that of the Ki67 proliferation index [16]. In this study, the HR of lymphatic metastasis was found to be 1.56 (95% CI, 1.03-2.38). In addition, compared to patients with lymphatic metastasis, the survival rate of patients without lymphatic metastasis was found to be significantly higher (p = 0:0095).
Patients with malignant solid tumors usually present with aberrant activation of the blood coagulation system, resulting in an elevated incidence of thromboembolism [17]. PT and INR are two blood coagulation markers that can be used to predict the outcomes of certain cancers [18]. In this study, a multivariate analysis demonstrated that the HR of PT was 0.59 (95% CI: 0.4207-0.8755), indicating that for every percentage increase in PT, the risk of death reduced by 41%. In addition, the survival rate of patients with a longer PT was found to be significantly higher than that of patients with a shorter PT (p = 0:0036). The international normalized ratio (INR) refers to the ratio of prothrombin time of a patient to the normal control prothrombin time to the ISI power (ISI: international sensitivity index). When the same sample is tested in different laboratories with different ISI reagents, the PT value results are usually very different. However, the measured INR values are the same, increasing the comparability of the measured results [19]. In patients with malignant tumors, the coagulation index also changed; hence, it is possible to use INR to detect the prognosis of patients with NETs. The heterogeneity of tumors is divided according to the degree of tumor deterioration compared to normal tissue morphology. The WHO 2010 classification defined welldifferentiated neuroendocrine and poorly differentiated tumor based on the Ki67 proliferation index and mitotic index [20]. A NET is defined as a well-differentiated neuroendocrine neoplasm resembling normal gut-pancreas endocrine cells [21]. In contrast, a neuroendocrine carcinoma (NEC) is defined as a poorly differentiated, high-grade malignant neoplasm composed of small or large to intermediate cells [22]. In addition, the type of pathology (high, medium, and low differentiation) is recommended by various guidelines in predicting the prognosis of neuroendocrine tumor patients [23]. Therefore, the degree of tumor differentiation can be used as a variable in order to predict the prognosis of neuroendocrine tumor patients.
Among NETs, metastatic cancer refers to the spread of neuroendocrine tumors from the primary site to another part of the body, in which liver metastasis is more common. Studies have shown that liver metastasis is one of the predictors of poor prognosis in patients with NETs [24]. Up to about 20% of patients have metastasis at the tumor site with a significant difference in metastatic potential. Most patients with pancreatic and small intestinal neuroendocrine tumors are diagnosed with metastatic status [25]. Here, metastasis was observed to play an important role in the development of neuroendocrine tumors. In addition, the number of tumor metastatic sites was found to be closely related to the prognosis of patients. The OS rate of patients with two or less meta-static sites was observed to be significantly better than that of one tumor site (p = 0:00095).
Evidently, the ROC results of this prediction model demonstrated that the prediction performance of the model was sufficiently robust in the training set (AUC value = 0:823) and test set (AUC value = 0:768). Moreover, the C-index results illustrate that the model constructed in this study possessed a certain robustness. In addition, in other nomogram models that predict the survival rate of patients with NETs at specific sites, the prognostic risk factors were mostly the Ki67 proliferation index and CgA [26]. Additionally, in this study, the variables used to build a prediction model were lymphatic metastasis, INR, PT, tumor differentiation, and the number of tumor metastasis sites. As a result, this study may provide new insights into the prediction of the OS rate of NET patients.
Some limitations are present in this study; however, first, this investigation may be affected by limitations associated with all retrospective studies and single-center studies. However, as a relatively rare tumor, the number of cases studied was relatively large, which is of great significance. Second, although a nomogram was developed and validated that can predict the survival rate of NET patients, its reliability needs to be verified in a larger cohort. In addition, the risk factor analysis did not include all potential factors affecting the survival of patients suffering from NETs. At the same time, the collected cases were not large enough, and a larger cohort of data should be collected in order to optimize the performance of the model.

Conclusion
Based on five variables (lymphatic metastasis, international standardized ratio, prothrombin time, tumor differentiation, and the number of tumor metastatic sites), the proposed nomogram was shown to predict the 3-year and 5-year survival rates of patients with NETs. In addition, the above five variables were significantly found to be correlated with survival rate. The C-index (training set: 0.890, test set: 0.804), ROC curve analysis results (training set: 0.828, test set: 0.768), and calibration curve all verified the robustness of the model. Therefore, the proposed model may be clinically helpful in predicting the prognosis of patients suffering from NETs.

Data Availability
The raw data used to support the findings of this study are available upon author's request.

Conflicts of Interest
The authors declare that there is no conflict of interest regarding the publication of this paper.

12
BioMed Research International