Establishment and Validation of Prognostic Nomograms for Patients With Parotid Gland Adenocarcinoma Not Otherwise Specified: A SEER Analysis From 2004 to 2016

Background: Parotid gland adenocarcinoma not otherwise specified (PANOS) is a rare malignant tumor with limited data on its characteristics and prognosis. This research is aimed at characterizing PANOS and developing prognostic prediction models for patients with PANOS. Methods: Cases from 2004–2016 were selected from the Surveillance, Epidemiology, and End Results (SEER) Program database. Univariate and multivariate Cox regression were applied to ascertain the factors associated with survival. Competing risk analysis and Gray's tests were employed to analyze cancer-specific death. Propensity score matching (1:1) was conducted to reduce the influence of confounding variables. Results: A total of 446 patients with a median age of 66 years were selected, of which 307 were diagnosed with stage III/IV PANOS. The 5-year overall survival (OS) rate of all patients was 51.8%, and the median survival time was 66 months. Surgical treatment clearly improved survival time (p < 0.001). In the subgroup analysis, radiotherapy showed survival benefits in patients with stage III/IV disease (p < 0.001). Multivariate Cox regression analyses showed that age, T classification, N classification, M classification and surgery were independent prognostic indicators for OS; T classification, N classification, M classification and surgery were independent risk factors for cancer-specific survival (CSS). In addition, age was independently associated with other cause-specific death. Based on the results of multivariate analysis, two nomograms were developed and verified by the concordance index (C-index) (0.747 and 0.780 for OS and CSS) and the area under the time-dependent receiver operating characteristic (ROC) curve (0.756, 0.764, and 0.819 regarding for nomograms predicting 3-, 5-, and 10- year OS, respectively and 0.794, 0.789, and 0.806 for CSS, respectively). Conclusions: Our study clearly presents the clinicopathological features and survival analysis of patients with PANOS. In addition, our constructed nomogram prediction models may assist physicians in evaluating the individualized prognosis and deciding on treatment for patients.


INTRODUCTION
As a rare malignant tumor, parotid gland carcinoma accounts for 1-3% of all head and neck malignancies (1).Most head and neck malignancies are squamous cell carcinoma, but parotid gland carcinomas are more diverse, with 24 different histologic subtypes according to the 2017 World Health Organization (WHO) classification (2,3).The most common pathological types of parotid gland carcinomas are mucoepidermoid carcinoma and adenoid cystic carcinoma, which has been systematically studied (4)(5)(6).Adenocarcinoma not otherwise specified (ANOS) is also a major subtype (7)(8)(9)(10) which refers to carcinoma that has different levels of glandular differentiation in histology but cannot be attributed to a specific type.Its incidence ranges from the second to the fourth in parotid gland carcinoma based on different retrospective studies (7,8,(11)(12)(13).Although a few case series (14,15) have shown that the biological behavior of parotid gland adenocarcinoma not otherwise specified (PANOS) is highly malignant, large sample research regarding to its clinical features and long-time survival is still lacking due to the paucity of patients.
In order to fully evaluate the clinicopathological characteristics and prognosis of patients with PANOS, we extracted cases from the SEER database between 2004 and 2016 and conducted a comprehensive analysis.Meanwhile, we constructed two nomograms to help physicians predict the overall survival (OS) and cancer-specific survival (CSS) of these patients directly.

Statistical Analysis
The primary endpoint was overall survival and the secondary endpoint was cancer-specific survival.Univariate Cox regression analyses were performed to filter out significant variables related to OS. Kaplan-Meier curves were plotted and compared using the log-rank test for each significant variable.Competing risk analysis and Gray's tests were utilized to test the differences in cancer-specific death between subgroups.The cumulative incidence function curves were also depicted.Multivariate Cox regression analyses were subsequently conducted.
One-to-one propensity score matching (PSM) was performed using nearest neighbor matching method to construct a matched cohort including pairs of radiotherapy and non-radiotherapy subjects.The caliper value was 0.02.Age, gender, race, laterality, marital status, pathological grade, T, N and M stages, stage, surgery and chemotherapy were covariates used for matching.
Nomograms were constructed based on the results of multivariate analyses.Calibration curves were delineated using the Kaplan-Meier method with bootstrap to evaluate the

Patient Characteristics
Four Hundred and Forty Six PANOS Patients (2004-2016) Were Chosen From the SEER Database (Table 1
According to univariate analyses (Table 1, Figure 3), nine variables including age, gender, pathologic grade, T classification, N classification, M classification, stage, chemotherapy, and surgery were significant variables related to OS.The results of competing risk analysis and Gray's tests showed that the variables above were still statistically significant for cancerspecific survival (Supplementary Figure 1), except for age (p = 0.080).Additionally, age, T classification and stage were identified to be related to other cause-specific death (Supplementary Table 1).
The significant variables were then incorporated into multivariate Cox regression.Variable stage was not included due to its high multicollinearity with T, N, and M classification variables.The final results (Table 2) presented that age (p < 0.001), T classification (p < 0.001), N classification (p < 0.001), M classification (p < 0.001), and surgery (p < 0.001) were independent risk factors for OS, while sex (p = 0.630), pathologic grade (p = 0.174) and chemotherapy (p = 0.628) were excluded (Table 2).Besides, T classification (p < 0.001), N classification (p < 0.001), M classification (p = 0.002), and surgery (p < 0.001) were independent prognostic factors for CSS.Age (p < 0.001) was identified as an independent risk factor for other cause-specific death (Supplementary Table 1).

Propensity Score Matching
According to the univariate analysis, radiotherapy did not improve the survival of patients (p = 0.278).To identify the efficacy of radiotherapy, propensity score matching was performed according to the variables described in the methods.Before PSM, radiotherapy showed a trend toward white race (p = 0.037), high grade (p = 0.017), large tumor size (p = 0.010), positive nodal metastasis (p = 0.021), advanced stage (p < 0.001) and chemotherapy (p = 0.008) (Table 3).After PSM analysis, there were no significant differences in baseline characteristics.

Establishment and Validation of the Nomogram
The significant variables based on the final multivariate models were included to create the nomograms for OS and CSS. Figure 5 presented two nomograms which predict 3-, 5-, and 10-year OS and CSS.The C-indexes of the nomograms for OS and CSS were 0.747 and 0.780, which indicated good discrimination of the two nomograms.The calibration curves also showed that the predicted survival probability matched well with the observed survival probability at the 3-, 5-, and 10-year time points (Supplementary Figure 2).
We also used time-dependent receiver operating characteristic (ROC) curves and the area under the ROC curve (AUC) to validate the efficacy of nomograms.As shown in Figure 6, the area under the curve (AUC) values of ROC were 0.756, 0.764 and 0.819 regarding for nomogram predicting 3-, 5-and 10year OS, respectively.Likewise, the 3-, 5-, and 10-year AUC values of nomogram predicting for CSS were 0.794, 0.789 and 0.806, respectively.

DISCUSSION
In this study, we analyzed data from a large cohort of 446 patients with parotid gland adenocarcinoma not otherwise specified and constructed two nomograms.Our findings may present a comprehensive viewpoint on the characteristics and prognosis of PANOS patients.
Our study found that some clinical features of PANOS were in accordance with previously published reports, including a male predominance, a tendency to lymph node invasion, a propensity for high-grade tumors and more patients who were diagnosed with advanced stage III/IV disease (8,14,15).
With respect to survival, the 5 and 10-year overall survival rates of PANOS patients were 51.8 and 36.7%, which were significantly lower than those of mucoepidermoid carcinoma and adenoid cystic carcinoma (4-6).According to similar SEER analyses, Sun et al. (4) showed that the 5 and 10-year OS rates of mucoepidermoid carcinoma were 83.2 and 73.6%, and the 5-year OS rates of adenoid cystic carcinoma in Tasoulas's study (6) were 81%.These findings indicate that PANOS has a relatively poorer prognosis among major parotid gland carcinoma subtypes, which calls for our attention.
Surgical resection is still the principal treatment for parotid gland carcinomas which has considerable effects on patient survival (16)(17)(18).In this study for PANOS patients, surgery notably prolonged OS among patients with PANOS (p < 0.001).The median overall survival of patients who underwent surgery increased by nearly 5 years compared to those who did not.Multivariate Cox analyses also revealed that surgery was an independent favorable prognostic factor both for OS and CSS.
Many studies have shown that radiotherapy can prolong survival of parotid gland carcinoma patients (19,20).In our study, the characteristics between radiotherapy and nonradiotherapy patients were imbalanced before propensity score matching analysis and univariate analysis showed that radiotherapy was not significantly correlated with OS (p = 0.278).After PSM analysis was conducted, the final results showed that radiotherapy can significantly improved survival (p = 0.03).Therefore, radiotherapy still plays a significant role in the survival of PANOS patients.
Previous studies have shown that age is an important prognostic factor for survival in patients with PANOS (8,21), and our study also found that younger patients had a better prognosis.In our study, age was an independent prognostic factor of OS and other cause-specific death, but it had no effect on CSS.This may be explained by the fact that PANOS has a relatively longer survival time compared with other cancers (3), and that older patients may suffer from death caused by competing events like cardiovascular events, which may hinder the occurrence of cancer-specific death.
Unsurprisingly, T, N, and M classifications were independent risk factors both for OS and CSS.A larger range of tumor extension, more involved lymph nodes, or distant metastasis were correlated with shortened survival time.
It is well known that the American Joint Committee on Cancer (AJCC) Staging is a commonly used prognostic tool for malignancies (22).Nevertheless, this staging system is not specially designed for individuals and did not include important prognostic factors such as age, tumor differentiation and the treatments that patients have received.The nomogram includes a variety of cancer-related risk factors and can individually predict the survival rate of patients in a visual way (23,24).Therefore, we established and validated two nomograms to predict the 3-, 5-, and 10-year survival rates in a quantitive way, which may assist physicians in evaluating patient's prognoses for PANOS.
This study has certain limitations.Firstly, in the process of patient selection, many patients were excluded due to the absence of information such as pathological grade and stage, which may have led to the extraction of inprecise OS and CSS for the patients.Secondly, the lack of data on chemotherapy regimens and clinical symptoms, such as pain or tenderness may limit the ability to identify their influences on the prognosis of PANOS patients.Thirdly, the pathological classification of parotid carcinoma is a challenge because its subtypes are so diverse and we cannot examine the original clinicopathological information, which may lead to biased information.

CONCLUSIONS
In summary, a population-based method was used to present the clinicopathological features and prognosis of patients with PANOS based on data extracted from the SEER database.We demonstrated that age, T, N and M classifications and surgery were independent prognostic factors for OS; T, N and M classifications and surgery were independent risk factors for CSS.In addition, age was independently correlated with other cause-specific death.Moreover, we developed two nomograms predicting the 3-, 5-, and 10-year OS and CSS in patients with PANOS, which can visually and effectively evaluate the prognosis of individuals.

FIGURE 1 |
FIGURE 1 | The flow diagram of the selection process for the study cohort.

FIGURE 2 |
FIGURE 2 | Kaplan-Meier curves of the 446 patients in the cohort.(A) Overall survival; (B) Cancer-specific survival.

FIGURE 4 |
FIGURE 4 | Kaplan-Meier curves of overall survival before (A) and after (B) propensity score matching analysis based on radiotherapy.

TABLE 1 |
Demographic and clinicopathological characteristics of patients and univariate Cox regression analyses of overall survival.

TABLE 2 |
Multivariate Cox regression analyses of overall survival and cancer-specific survival.
HR, hazard ratio; CI, confidence interval; Ref, reference.Bold value means the p-value of a statistically significant factor.

TABLE 3 |
Baseline characteristics of patients divided by radiotherapy in the regular and matched groups.