Development and Validation of Novel Nomograms Using Serum Tumor Markers for the Prediction of Preoperative Histologic Grades in Gastroenteropancreatic Neuroendocrine Tumors

Background To develop and validate nomogram models for the preoperatively prediction of the histologic grade of gastroenteropancreatic neuroendocrine tumors (GEP-NETs) to provide appropriate treatments. Methods A total of 1014 participants, including 211 healthy controls, 293 patients with benign diseases, 299 patients with cancers, and 211 patients with GEP-NETs were included in the final analysis. Their sociodemographic and laboratory information, including serum tumor markers such as AFP, CEA, CA19-9, CA72-4, Cyfra21-1 and NSE were collected. Nomogram models were developed to preoperatively predict histologic grades of GEP-NETs. Results Among six serum tumor markers, only NSE was found to have a statistically significant association with the histologic grades in GEP-NETs (G1 vs. G2: p < 0.05; G2 vs. G3: p < 0.001; G1 vs. G3: p < 0.0001), which was combined with sex and age to develop the nomogram models. The first nomogram (to differentiate grade 1(G1) and grade 2/3 tumor (G2/G3)) showed a strong association to differentiate with an AUC of 0.747 (95% CI: 0.663-0.832) and 0.735 (95% CI: 0.624-0.847) in the training and validation datasets, respectively. The second nomogram (to differentiate G1/G2 and G3 tumors) showed a strong association to differentiate with an AUC of 0.827 (95% CI: 0.744-0.911) and 0.847 (95% CI: 0.744-0.950) in the training and validation datasets, respectively. The ROC, area under ROC curve (AUC), calibration curve and decision curve analysis (DCA) demonstrated the clinical usefulness of both models. Conclusions We proposed two novel nomogram models based on sex, age and serum NSE levels to preoperatively predict the histologic grades in GEP-NETs to assist the clinical decision-making.


INTRODUCTION
Neuroendocrine tumors (NETs) are heterogeneous malignancies arising from the diffuse neuroendocrine system. They can appear in various anatomic locations, but the majority of NETs are restricted to derivatives of the embryological gut, including the gastrointestinal (GI) tract and bronchopulmonary tree. Gastroenteropancreatic neuroendocrine tumors (GEP-NETs), including GI neuroendocrine tumors (GI-NETs) and pancreatic neuroendocrine tumors (pNETs) originate from enterochromaffin cells of the gutislets of Langerhans, respectively (1)(2)(3)(4)(5). Although GEP-NETs are rare, they comprise the second most common tumor of the digestive system after colorectal cancer. In the past few decades, the incidence rate of GEP-NETs has been increasing globally, which could be due to increased awareness and improvement in their detection methods (6)(7)(8).
According to the 2010 World Health Organization (WHO) grading system, well-differentiated NETs are classified as grade 1 (G1) and grade 2 (G2) tumors, and poorly-differentiated NETs are classified as the grade 3 (G3) tumors, based on the mitotic count and Ki-67 proliferation index (9). Tumor grade is a crucial determinant to guide the GEP-NETs management, but is usually determined on the postoperative specimens, which influencing the physician's decision making in clinical practice (10). Recently, tissue acquision by EUS-guided fine needle aspiration (EUS-FNA) has helped evaluate the preoperative histologic grade, but with controversial accuracy (11,12).
Various peptide hormones and biogenic amines secreted by GEP-NETs can enter systemic circulation, which could be used as biomarkers in the outpatient setting (13). At present, serum tumor markers are being extensively studied to provide future direction for the diagnosis, prediction and prognosis of the cancers (14). The neuron-specific enolase (NSE) is a cellspecific isozyme of the glycolytic enzyme enolase, which is highly specific for the neurons and peripheral nerve endothelial cells. Malignant neuron hyperplasia in the NETs may lead to an increase in the serum NSE level which can be used for the diagnosis, staging and treatment of these tumors, including GEP-NETs (15,16). In this study, we have proposed two novel nomogram models to evaluate the role of NSE in the preoperative diagnosis and grade prediction of GEP-NETs.

Study Design
We collected the demographic information of the study participants, including sex, age and test results of six serum tumor markers, including alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), cancer antigen 19-9 (CA19-9), cancer antigen 72-4 (CA72-4), cytokeratin 19 fragment 21-1 (Cyfra21-1) and NSE. We compared the distribution of all these serum tumor markers in the study participants, including the HCs, benign disease, cancer and GEP-NETs groups. Next, in the GEP-NETs group, we compared the distribution and differences of serum NSE in different histologic grades as follows: Low grade or Grade 1 [G1] tumors have a mitotic rate of 0 to 1 per 10 high power fields (HPF) and a Ki-67 index of 0% to 2%, the mitotic rate of tumors of intermediate grade (G2) is 2 to 20 per 10 HPF or 3 to 20% of Ki-67 index and the mitotic rate of tumors of high grade (G3) is greater than 20 per 10 HPF or Ki-67 index 20% (9). According to the basic principles of variable selection in clinical prediction modelling, we selected the candidate variables for the model by the univariate logistic regression analysis and clinical knowledge (17). Comprehensively considering the significant levels in the two models (G1 vs G2/, G1/2 vs G3), variables with significant difference (p < 0.05) and clinical significance were chosen. Then, we chose the full model as the final model. The cutoff value of variable was decided according to the maximum Youden index of the ROC curve, which was used to turn into a binary variable.

Statistical Analysis
Categorical variables are represented by frequency and proportion and continuous variables are represented by mean (standard deviation) and median (minimum and maximum). Because the distribution of the serum tumor markers in this study is right skewed, we transformed them to Normal distribution by taking log10, which made the prediction models more readable. The t-test or Mann-Whitney U test were used to evaluate the differences in the distribution of six tumor markers between the disease groups and the healthy control group. And NSE was compared in different grades of GEP-NETs.
Nomograms are based on the ratio of each regression coefficient to 0 to 100 points in the logarithmic regression conversion. The effect of the variable with the highest b coefficient (absolute value) is assigned 100 points. Add these points to the independent variables to get the total points and convert them into predicted probabilities. The predictive performance of the nomogram was measured by the area under the ROC curve (AUC) and the calibration curve with 1000 bootstrap samples. In addition, we performed the decision curve analysis, which calculates a clinical "net benefit" for the nomograms in comparison to default strategies of treating all or no patients. X-axis is preference, whose unit is High Threshold Probability. The Cost: Benefit Ratio help us see the relationship between preference and threshold probability easily. Y-axis shows the clinical decision net benefits after the benefits minus the disadvantages (18). R version 3.6.1 (http://www.rproject.org/) for all data analysis.

Characteristics of Participants
Data was collected on a total of 1014 individuals during the study period. There were 211 healthy controls, 293 patients with benign diseases, 299 cancer patients, and 211 patients with GEP-NETs in the cohort with male patient accounting for 50.24%, 56.31%, 63.55% and 50.24%, respectively. Other demographic variables and information about the levels of six serum tumor biomarkers for these groups are shown in Table 1. The mean ages of the healthy controls, benign diseases, cancer patient and GEP-NETs patients were 54.43, 55.72, 62.7, and 54.36 years, respectively. The serum NSE levels were the highest in the patients with GEP-NETs. Figure  S1 showed the violin plots of six tumor markers in four groups. Three disease groups were compared with the healthy control group, respectively. Among all the tumor markers, NSE was significantly different in GEP-NETs (p < 0.0001) and had a smallest overlap with other disease groups. In addition, six tumor markers for distinguishing GEP-NETs from healthy and other disease groups were shown in the Figures S2 and S3, respectively. Among them, serum NSE showed the best diagnostic performance.

Nomogram to Differentiate G1 From G2/3 in GEP-NETs
Two hundred eleven patients with GEP-NETs were randomly divided according to the ratio of 7:3. There were 133 people in the training dataset, including 48 people in the G1 group and 85 people in the G2/3 group. Seventy eight individuals were included in the validation dataset, with 34 people in the G1 group and 44 people in the G2/3 group ( Table 2). The results of the univariate logistic regression analysis showed that sex, age and NSE level was of clinical significance ( Table 1). The nomogram to differentiate G1 from G2/3 in GEP-NETs was constructed based on the full model ( Figure 2A). Age was used as a binary variable (<54.5 years and ≥54.5 years) according to the maximum Youden index of the ROC curve. NSE levels were transformed into the Normal distribution by taking log10. The AUC of the model reached 0.747 (95% CI: 0.663-0.832) and 0.735 (95% CI: 0.624-0.847) in the training and validation datasets, respectively ( Figures 2B, C). The calibration curve showed a high accuracy of the nomogram for predicting tumor pathologic grades both in the training and validation datasets ( Figures 2D, E). The DCA was used to demonstrate the clinical decision utility of the nomogram. The area under the decision curve in Figures 2F, G showed the clinical utility of corresponding strategies. The nomogram (red) showed more area than that the "treat all" (grey) or "treat none" (black) strategies, in both the training and validation datasets.  Table 3).
Comprehensively considering the results in the two models (G1 vs G2/, G1/2 vs G3), age, NSE and sex were chosen (Table S1). Then, we chose the full model as the final model and the nomogram to differentiate G1/2 from G3 in GEP-NETs was constructed ( Figure 3A). Age was used as a binary variable (<56.5 years and ≥56.5 years) and NSE levels were taken log10. The AUC reached 0.827 (95% CI: 0.744-0.911) and 0.847 (95% CI: 0.744-0.950) in the training and validation datasets, respectively ( Figures 3B, C). The calibration curve showed a high accuracy of the nomogram for predicting tumor pathologic grades both datasets ( Figures 3D, E).

DISCUSSION
In this study, we evaluated the role of serum NSE levels in the preoperative diagnosis and histologic grade prediction of GEP-NETs. Two novel nomogram models were established to predict the   (19). However, NSE alone is not sufficient for the diagnosis of NETs, as only 30% to 50% of NETs secrete NSE (20). In a study involving more than 200 patients with GEP-NETs, the sensitivity and specificity of NSE to distinguish NETs from nonendocrine tumors were only 39-43% and 65-73%, respectively (13). However, an article showed that NSE has specificity for NETs than other tumor markers. In their study, all tumors positive for an accepted neuroendocrine marker also expressed NSE (21). In our study, AUC for NSE to distinguish GEP-NETs from healthy individuals was 0.819 ( Figures S2). Nevertheless, in distinguishing GEP-NETs from other disease groups, the AUC was 0.657 ( Figures S3). Despite these results, we believe that NSE still has a potential in the diagnosis of GEP-NETS among the six serological tumor markers. In this study, serum NSE was found to be effective for GEP-NETs grade classification. GEP-NETs are heterogeneous in terms of origin, biological behavior with a malignant potential (5,22). In the past few decades, various classification systems based on the embryological origin or morphological differences have been proposed for GEP-NETs (23,24). The World Health Organisation (WHO) 2010 classified GEP-NETs in welldifferentiated (G1 and G2) tumors, while poorly differentiated neuroendocrine carcinoma (NEC) are considered equivalent to G3 tumors (25). Different grades of GEP-NETs have different clinical severity and prognosis with different treatment approach (26,27). In addition, different histologic grades have different prognosis based in the origin. In midgut GEP-NETs, the 5-year OS rates for G1, G2, and G3 tumors are 79%, 74%, and 40%, respectively (28). In pNETs, the 5-year OS of G1, G2, and G3 is 75%, 62%, and 7%, respectively (29). Therefore, it is crucial to predict the histologic grade of GEP-NETs preoperatively to help clinicians take decisive management actions effectively. A previous study proposed a combined nomogram model based on the radiomics signature and clinical-stage to distinguish G1 and G2/3 in pNETs for the treatment. In their study, parenchyma-sparing resections for G1 and a comprehensive treatment strategy including radical surgical resection with systematic chemotherapy was needed for patients with G2/3 to improve the survival (30). However, another study showed that the treatment strategies between G2 and G3 in pNETs should not be the same (31). These patients should receive surgical treatment in patients with limited metastatic disease, if technically feasible. Besides, targeted therapy with everolimus or sunitinib and somatostatin analogs (octreotide) is also used for advanced pNETs G1/2. Therefore, it is crucial to differentiate between G1/2 and G3 among pNETs than between G1 and G2/3 pNETs (31). Here, we developed two nomograms, one was used to distinguish G1 and G2/3 and the other was to distinguish G1/2 and G3.
A study showed that the average values of serum NSE for G1, G2, and G3 were 13, 17 and 21 mg/L in p-NETs, respectively (32). This was consistent with our results of NSE with significant differences in three grades of GEP-NETs. In our study, the nomogram differentiating G1/2 from G3 had a larger AUC than the nomogram differentiating G1 from G2/3. These results are consistent with the previous reports of advantages of NSE in diagnosing NETs with poor differentiation. In addition, elevation in NSE levels also reflected the overall survival of patients with GEP-NETs. Elevated serum NSE indicated the active disease, suggesting that the elevated NSE levels at the time of intial diagnosis is associated with poor prognosis (33). Therefore, the NSE can be used as a reliable diagnostic and prognostic markers in patients with GEP-NETs. We would also like to note some limitations of our study: (I) Relatively small sample size because it was a single-center study; (II) The information about the CgA was unavailable; (III) Inability to perform external validation of the data, and the conclusion in this study requires a larger multicenter validation analysis in future. In summary, we reassessed the role of serum NSE in the diagnosis and prediction of preoperative histologic grades in GEP-NETs. We developed two novel nomogram models based on sex, age and serum NSE levels, which can be used as a non-invasive and accurate assessment tool for GEP-NETs patients during preoperative period to help clinicians tailor treatment plans accordingly.

DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding author.

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by the First Affiliated Hospital of Nanjing Medical University (Nanjing, China) ethics review board. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

AUTHOR CONTRIBUTIONS
All authors contributed to the article and approved the submitted version.