A nomogram for predicting overall survival in patients with Merkel cell carcinoma: A population-based analysis

Background: Merkel cell carcinoma (MCC) is a rare neuroendocrine skin cancer with increasing incidence and poor prognosis. We sought to develop and validate a nomogram to estimate overall survival (OS) of MCC patients. Methods: 1863 MCC patients between 2010-2015 from the Surveillance, Epidemiology and End Results (SEER) database were randomly divided into the training and validation cohort. Independent prognostic factors determined by Cox regression analysis in the training cohort were used to establish a nomogram. We evaluated prognostic performance using the concordance index (C-index), area under receiver operating characteristic curve (AUC) and calibration curves. Decision curve analysis (DCA), net reclassication index (NRI) and integrated discrimination improvement (IDI) were used to compared the the nomogram’s clinical utility with that of the staging system. Results: eight independent prognostic factors were incorporated in the nomogram. The C-index of the nomogram was 0.744, which was superior to the C-index of AJCC TNM Stage (0.659). The AUC was greater than 0.75 and the calibration plots of this model exhibited good performance. Additionally, the positive NRI and IDI of nomogram versus the staging system illustrated that the nomogram had better predictive accuracy than the staging system (P<0.001) and the DCA showed great clinical usefulness of the nomogram. MCC patients were perfectly classied into three risk groups by the nomogram, showing better discrimination than the staging system. Conclusions: We developed and validated a nomogram to assist clinicians in evaluating prognosis of MCC patients.


Introduction
Merkel cell carcinoma (MCC) is a rare neuroendocrine skin cancer that often presents as a solitary cutaneous or subcutaneous nodule on sun-exposed areas in the advanced age population 1 .
Immunosuppression, chronic sun exposure and advanced age are major risk factor [2][3][4] . Approximately 80% MCC cases are associated with Merkel cell polyomavirus (MCPyV) infection and the remaining 20% are associated with chronic sun exposure [5][6][7] . Since the rst description by Toker in 1972 8 , the incidence of MCC increased rapidly and this trend was sustained into the new millennium 9-12 . The increased incidence was considered to be related to reduced misdiagnosis as the new pathological diagnosis technology of cytokeratin-20 staining was in introduced in the 1990s 13 . Besides, the increase in at-risk population of elderly people, immunosuppression individuals is another important risk factor 12,14 .
MCC is highly malignant and aggressive, with more than 1/3 patients dying from MCC, which is the second leading cause of skin-cancer death following melanoma 15,16 . Local failure, regional recurrence as well as distant metastases remain primary causal factors of the high mortality and poor prognosis 17,18 . Approximately 26%-36% MCC patients have nodal diseases and 6-16% of patients present with distant metastases at the time of diagnosis 19 . During their disease course, up to 25-50% patients present local or locoregional recurrence depending on their stages and nodal diseases [20][21][22][23] . Surgery is the the primary treatment modality for MCC. If surgery is not feasible, radiation therapy is an effective strategy to control disease. Chemotherapy was the only treatment option for advanced-stage or refractory MCC patients before immunotherapies was demonstrated to be effective in MCC in several recent clinical trials 19 . However, despite major advancements in the understanding of MCC biology and treatment, the clinical outcomes of MCC are still poor, suggesting that an individual prediction model is needed to facilitate better treatment strati cation and outcome evaluation. However, to the best of our knowledge, there is no individual prognostic model for MCC owing to its rarity.
Currently, the most commonly used standard for prognostication in MCC is the American Joint Committee on Cancer (AJCC) staging system. However, an obvious shortcoming of AJCC staging system predicting survival was its low accuracy. Since other factors such as age, gender and site, are also associated with patients' outcome, a personalized predictive model for MCC patients is warranted.
A nomogram is a reliable tool to predict and quantify an individual probability of a certain clinical event by integrating prognostic and determinant factors, so that it could predict patient outcomes accurately and facilitate personalized medicine. Therefore, in current study, we sought to develop and validate a nomogram and built a risk strati cation system for MCC patients based on a large set MCC dataset from the Surveillance, Epidemiology and End Results (SEER) database. Besides, we compared the predictive accuracy of the survival nomogram with that of the American Joint Committee on Cancer (AJCC) stage.

Cohort population
A retrospective study was conducted based on the information from the SEER database, a public available cancer statistics database, which is constitutive of 18 cancer registries in the United States and covers about 28% of the total population of the United States (https://seer.cancer.gov/data/). Informed consent was waived for the use of public data from the SEER.
The SEER*Stat software (Version 8.3.6) was used to recruit MCC patients from 2010 to 2015. All cases were diagnosed as MCC by histology con rmation with the International Classi cation of Diseases for Oncology, Third Edition (ICD-O-3) histologic codes 8247/3 (Merkel cell carcinoma). According to the Primary Site -labeled codes, primary sites were classi ed into four sites as follows: NOS/overlapping codes (C448-C449), head and neck (C440-C444), trunk (C445), extremities (C446-C447). We selected patients over 18 years of age and those with only one primary tumor. Patients with unknown staging or unknown follow-up were excluded. The patient screening owchart was showed in Supplemental gure 1.
Patient demographics including age, gender, race, marital status, primary sites, tumor stage, treatment [primary site surgery, sentinel lymph node biopsy and/ or lymph nodes removal (SLNB and or LN removel), chemotherapy and radiation] were obtained from the database. Tumor stage in SEER database between 2010 to 2015 was recorded according to the AJCC Cancer Staging Manual, 7th edition. The endpoint of our study was overall survival (OS), which was de ned as the time from cancer diagnosis to the time of death from any cause or of the last follow-up.

Statistical Analyses and Nomogram Development
All included MCC patients were randomly assigned to the training and validation cohort in a ratio of 7:3 using the random sampling function in version 3.6.2 of R software. Patient characteristic between the training and validation cohort were compared by descriptive statistics. Categorized variables were analyzed by Chi-square tests and continuous variables were compared using the t test. Univariate and multivariate Cox models were performed to identify variables that signi cantly affect overall survival (OS) in the training cohort. Based on these prognostic factors in training cohort, we established a nomogram to predict the 3-year and 5-year OS rate for MCC patients with the "rms" package. The discrimination and calibration of the nomogram were validated in the training and validation cohort. To evaluate discriminative ability, concordance index (C-index) and area under curve (AUC) value were calculated. Typically, the C-index and AUC value greater than 0.7 suggested relatively favorable discrimination. Calibration ability was evaluated by calibration plots. The integrated discrimination improvement (IDI) and the net reclassi cation improvement (NRI) were calculated to compare the accuracy of the established nomogram with that of AJCC staging system. Decision curve analyses (DCA) were performed to assess the clinical usefulness and bene ts of the nomogram. A risk strati cation system was built according to the total points of each patient in training cohort. We classi ed all patients into the low-, intermediate-, and high-risk groups with a similar number of cases. Kaplan-Meier method were used to generate survival curves and the Log Rank test was performed to compare the differences among the curves. All statistical tests were two-sided and a P value < 0.05 was considered statistically signi cant. All statistical analyses were conducted using the statistical software packages R version 3.6.2 (http://www.R-project.org, The R Foundation) and SPSS statistics version 23.0 (IBM SPSS Statistics, New York, United States).

Patient characteristics
A total of 1863 MCC patients from 2010 to 2015 were enrolled in our study, who were randomly divided into the training (n=1307) and validation (n=556) cohort by a ratio of 7:3. Patient characteristics were summarized in Table 1. The average age of MCC patients in the total, training, and validation cohorts were 73.92 years, 74.02 years and 73.70 years, respectively. In the total cohort, the majority of cases were male (n=1145, 61.5%) and white (n=1760, 94.5%). There were 41.9% (n=780) and 39.0% (n=726) MCC occurring in skin of extremities and skin of head and neck, respectively. Most patients (n=1567,84.1%) in the total cohort study received primary site surgery. 64.3% (n=1198) patients underwent SLNB and or LN removel and 55.7% (n=1037) patients were treated with radiation. However, only 13.5% (n=251) patients underwent chemotherapies. Baseline characteristics between the training and validation cohort were balanced (all P>0.05).

Independent prognostic factors for OS in training cohort
The median follow-up time was 43.0 months (95% CI: 40.9-45.1 months) and the median OS was 59.0 months (95% CI: 51.6-66.4 months) in the training cohort. Univariate and multivariate Cox regression analysis were conducted to screen signi cant prognostic factors for OS in the training cohort. In the univariate analysis (Table 2), nine variables were signi cantly associated with OS, including age, gender, marital status, primary site, tumor stage, primary site surgery, SLNB and or LN removel, radiation and chemotherapy (all P <0.05). Then we performed multivariate analyses to identify factors identi ed in the univariate analyses. Age, gender, marital status, primary site, stage, SLNB and or LN removel, radiation were independent risk factors for prognosis of MCC patients ( Table 2).

Development and validation of Nomogram
Based on all independent prognostic indicators for OS in the training cohort, we construct a prognostic model to predict 3-and 5-year OS for MCC patients. The prognostic model was virtually presented in the form of a nomogram ( Figure 1) and was validated using a dependent validation cohort. A C-index and ROC curves were used to evaluate the discrimination performance. The C-index of OS prediction was 0.744 (95% CI: 0.722-0.766) in the training set and 0.737 (95% CI: 0.706-0.768) in the validation set. The ROC curves (Figure 2) showed that the AUC values for predicting 3-and 5-year OS were over 0.75 in the both training and validation cohorts. The calibration curves (Supplemental gure 2) showed optimal consistencies between the predicted and observed survival probability in the both training and validation cohorts. Overall, our nomogram had good discrimination and calibration ability and was validated in the validation cohort.
Comparison of clinical value between nomogram and the 7 th AJCC TNM stage system AJCC tumor stage system was traditionally used to predict prognosis strati cation for MCC patients in clinical practice. As tumor stage in SEER database between 2010 to 2015 was recorded according to the AJCC Cancer Staging Manual, 7th edition, we sought to evaluate whether the nomogram that included the7 th AJCC TNM stage information and other prognostic factors could perform better than 7 th AJCC TNM stage system alone in stratifying OS. Firstly, the C-index for OS prognosis by the nomogram was 0.744, which was signi cantly higher than the C-index of 0.659 for OS prognosis by the 7 th AJCC tumor stage alone, suggesting that our nomogram had better accuracy in predicting OS for MCC patients in training cohort. And the result was also con rmed in the validation cohort. We further calculate NRI and IDI value to evaluate the accuracy between the established nomogram and the 7th AJCC TNM stage alone ( Table 3) The decision curve analysis was performed to compare the clinical bene ts among the nomogram and the 7 th AJCC TNM stage system (Figure 3). The decision curves displayed that if the threshold probability of a patient is > 15%, the established nomogram in prognosticating OS yielded more bene t than that of than all patients dead scheme or none patient dead scheme in the both training and validation cohorts.
Moreover, in this range, the nomogram could better predict the 3-and 5-year OS than the 7 th AJCC TNM stage system, as it added more net bene ts compared with the 7 th AJCC TNM stage system.

Risk strati cation based on the nomogram
A risk strati cation system was built based on the total points of each patient in the training cohort. MCC patients were classi ed into three risk groups: the low-risk group (total points ≤ 96), the intermediate-risk group (96<total points ≤132) and the high-risk group (total points >132). As the (Supplemental gure 3A-C) showed, the 7 th AJCC tumor stage system was relatively unsatisfactory in stratifying MCC patients between stages II and III in the training, validation cohorts and whole population. However, OS in three risk groups was accurately differentiated in the training cohort, validation cohort and whole population (Supplemental gure 3 D-F). These results suggested that the risk strati cation system could perfectly classi ed patients into three risk groups and showedgreater discrimination than he 7 th AJCC tumor stage system.

Discussion
MCC was a rare but highly malignant tumor, until now its prognosis and treatment treatment decision were mainly based on the conventional TNM stage system. However, it is well known that the stage system only takes the anatomical extent of the disease into consideration and not think about patient clinicopathological characteristics and treatment which also affected patient prognosis, so that the TNM stage system was unable to completely re ect the accurate prognosis and personal feature. Nomograms for some cancers have been established and shown to be more accurate than the conventional staging systems for predicting survival [24][25][26][27][28] . Nomogram for MCC was not yet established. Therefore, in current study, we established a nomogram for MCC patients based on the data from SEER database, which included eight signi cant prognostic factors that were selected by univariate and multivariate Cox analysis. According to the standard deviation along nomogram scales, we could intuitively nd that the 7th AJCC TNM stage was the most important factor, followed by SLNB and or LN removal, age, radiation, gender, marital status and primary site.
In fact, some important prognostic factors for MCC patients, including age, gender, marital status and primary site, have been proposed in existing literatures. Importantly, in current study, we fully took these factors into consideration and incorporated them into the nomogram. For example, a large population study of 6908 MCC patients reported that age, gender, primary site and marital status were important factors affecting mortality of MCC 29 . Similarly, a study of 3048 MCC patients reported that age older than 75 years and male sex showed negative effect on OS 30 . Moses Tam et al. also demonstrated that women with MCC showed improved survival compared to man with MCC, even after propensity score-matched analysis 31 . Some studies thought the underlying cause is because women have stronger innate and adaptive immune responses than men 32,33 . These results consistent with our study suggested that demographic and clinicopathological characteristics of MCC patients were strongly associated with OS. When predicting personal prognosis, these factors should be considered.
Patient treatments were another important prognostic factors. In present study, undergoing SLNB and or LN removal was associated with improved OS. It is well known that lymph node invasion was the major route of metastasis in MCC and patients with lymph node metastasis have accelerated disease progression, associated with worse survival 34,35 . SLNB and or LN removal was a valuable tool to assess the regional LN status. SLNB is recommended for patients with clinically negative lymph nodes, which has been shown to improve survival (30). Well, LN removal was an important treatment for patients with clinically positive lymph nodes or SLNB-positive and it also could quantify regional metastatic lymph nodes and assess exact node stage. A recent population study showed that number of metastatic LNs was the dominant nodal factor for survival in patients with MCC 36 . Therefore, SLNB and or LN removal was a strong favorable predictor for OS in patients with MCC as it permitted more accurate staging and more appropriate prognosis and management.
Besides, receiving radiotherapy was also a signi cant favorable prognostic factor. Previous study has shown that MCC is very responsive to radiotherapy, which could control the local disease in 75-85% of cases [37][38][39][40] . Radiotherapy has been used as the de nitive, adjuvant, and palliative treatment of patients with MCC within a multidisciplinary framework, which was highly effective in providing locoregional control bene ts in many literatures 41 . Local control may further translate into survival bene ts. Therefore, receiving radiotherapy as an effective treatment for MCC showed positive effect on survival.
Chemotherapy was not associated with improved OS and it was a negative factor for OS in univariate cox analysis in present study. This was because patients with chemotherapy in our study had a higher rate of adverse prognostic features, such as stage III-IV disease (stage III: 59.3% vs 31.9%, stage IV: 34.1% vs 5.6%, P<0.001), male sex (68.2% vs 60.4%, P=0.027), and extremities primary (26.9% vs 5.9%, P<0.001), compared with patients without chemotherapy (Supplemental Table 1). These adverse prognostic features may be responsible for worse prognosis in patients with chemotherapy rather than chemotherapy itself. Besides, only 182 (13.9%) patients in the training cohort underwent chemotherapy, which made it di cult to evaluate its role correctly. And it is noticed that chemotherapy for MCC patients showed short duration of response with the reported median progression-free survival of 61 days and 94 days in two studies, as well as, high rates of toxic death 42-44 .
Notably, primary site surgery was an important component for MCC, however it was not included in the nomogram. This might be because in present study most patients (84.1%) had undergone surgery, making its effects hard to be analyzed properly. In addition, the management of surgery, including its extent and margins, should be decided on an individual basis and these surgery related factors were not discussed in detail in our study. In fact, wide local excision of the primary tumor is the standard of care and a 1-2 cm excision margin down to the muscle fascia or the pericranium was recommended in guideline 19 . Therefore, we could not be interpreted as meaning that surgery has no bene t on survival.
Overall, in current study, we established a prognostic nomogram for MCC patients, which integrates demographic and clinicopathological characteristics, patient treatment. The nomogram performed well in predicting overall survival, with good discrimination (C-index, 0.744; AUC>0.75) and calibration in the primary cohort and was validated in the validation cohort (C-index, 0.737; AUC>0.80). When compared with the conventional TNM stage, our nomogram had increased accuracy for predicting OS, re ecting by that the C-index of nomogram was 0744 greater than 0.659 for the conventional TNM stage. Besides, the positive NRI and IDI of the nomogram versus the staging system further suggested that the nomogram could predict survival more accurately than the conventional TNM stage alone. Furthermore, DCA curves showed that the established nomogram predicted survival with better clinical bene t and utility compared with the conventional staging system. Moreover, patients were classi ed into low-, intermediate-, and high-risk groups according to their nomogram total points. The Kaplan-Meier curves clearly showed signi cant differences in OS among the three risk groups with better discrimination than the conventional TNM stage. Overall, our nomogram was a valuable tool to assist clinicians in predicting prognosis of MCC patients and facilitate personalized medicine.
However, there are some limitations in present study. Firstly, we were unable to evaluate the prognostic role of immunity therapy and surgical margins in MCC due to the retrospective nature and limited data availability of SEER database. Besides, it would be better if there was a multicenter clinical validation cohort to evaluate the external utility of the nomogram. However, as MCC is very rare, it is di cult for us to collect a multicenter clinical validation. But in current study we evaluated a large set of samples including the data from 18 medical centers registered in the SEER database, which represent the populations of different areas.

Conclusion
In conclusion, our nomogram had better accuracy, better clinical utility, and more precise prognosis prediction than the conventional staging system, which may be a valuable tool for predicting survival of patients with MCC.