Predictive value of a nomogram for melanomas with brain metastases at initial diagnosis

Abstract Background Estimation of incidence and prognosis of melanomas with brain metastases (MBM) at initial diagnosis based on a large cohort is lacking in current research. This study aims to construct an effective prognostic nomogram for newly diagnosed MBM. Materials and Methods Patients diagnosed with melanomas from Surveillance, Epidemiology, and End Results program between 2010 and 2014 were enrolled in our study. Risk factors predicting brain metastases (BM) were identified using logistic regression analysis. Cox regression analysis was performed to identify prognostic factors of overall survival (OS). Nomogram for estimating 6‐, 9‐, and 12‐month OS was established based on Cox regression analysis. The discriminative ability and calibration of the nomogram were tested using C statistics, calibration plots, and Kaplan‐Meier curves. Results Sixty‐two thousand three hundred and sixty‐nine melanoma patients were enrolled, including 928 with BM. Sex, marital status, insurance status, subsite, surgery of primary sites, radiation, chemotherapy, bone metastases, liver metastases, and lung metastases were associated with MBM at initial diagnosis. On multivariable Cox regression, the following eight variables were incorporated in the prediction of OS: age, unmarried status, absence of surgery to primary sites or unknown, absence of radiation or unknown, absence of chemotherapy or unknown, with bone metastases, with liver metastases, and with lung metastases. The nomogram showed good predictive ability as indicated by discriminative ability and calibration, with the C statistics of 0.716 (95% CI, 0.695‐0.737). Conclusions The incidence and prognosis of MBM patients were well estimated in this study based on a large cohort. The nomogram performed well and could be a useful tool to predict prognosis.


| INTRODUCTION
In the United States, melanoma is the fifth most common cancer with a rapidly increasing incidence of 96 480 new cases in 2019. 1 The 5-year overall survival (OS) rate of cutaneous melanoma is 91.8% (2008-2014). 2 However, long-term survival rate in patients with distant metastatic melanoma has been less than 10%, 2 with a median survival time of only 6-9 months. 3 It is estimated that 40% of patients with melanoma initially present with localized disease, 9% with regional disease, and 4% with distant metastatic disease. 1 Malignant cutaneous melanoma has the third highest incidence of brain metastasis among all types of cancer, following lung and breast cancers. 4 According to previous studies, approximately 10% of patients with malignant melanoma and 40%-60% of patients with metastatic melanoma ultimately develop brain metastases (BM). [5][6][7] The related factors for melanoma with BM (MBM) incidence and prognosis such as age, depth of invasion, location, systematic disease, number of intracranial and extracranial metastases, etc, have all been reported in previous studies but remain controversial. 5,[8][9][10] Nomogram survival prediction has been widely used in cancer research including hepatocellular carcinoma, 11 adrenocortical carcinoma, 12 and colorectal cancer. 13 Yet, to our knowledge, it has not been reported as a related factor on MBM.
Therefore, our study aims to evaluate the incidence, risk, and prognostic factors of newly diagnosed MBM patients using the Surveillance, Epidemiology, and End Results (SEER) database. Nomogram survival prediction was visualized using Cox regression model due to its ease of use and ability to facilitate management-related decision-making. 14 To our knowledge, we have established the first nomogram to predict the probability of survival rate for MBM.

| Patient
Information about BM of melanoma patients at initial diagnosis from the SEER program between 2010 and 2014 was obtained. From the SEER database, 106 739 patients pathologically diagnosed with melanomas were identified. The followings were the exclusion criteria: (a) patients with other primary cancer(s) (N = 36 010); (b) patients diagnosed at autopsy (N = 176); (c) patients with missing or incomplete information about survival time (N = 4731); BM (N = 2226); bone, liver, and lung metastases (N = 4); (d) patients younger than 18 years old (N = 295). Based on these criteria, 63 297 patients remained for incidence analysis. But only 928 patients were diagnosed with MBM and used for survival analysis. Patients with unknown factors (N = 101) were removed; and therefore, only 827 patients were eligible for nomogram prediction ( Figure 1). Informed consent was not required because SEER data contained no personal identifying information. The variables included in the study were age, sex, race, marital status, insurance status, subsite, surgery of primary sites, radiation, chemotherapy, bone metastases, liver metastases, lung metastases, BM, and survival months. Eight hundred and twenty-seven MBM patients were then randomly divided into two groups: the training set (N = 414) and the validation set (N = 413). Training cohort was used to construct the nomogram and validation cohort was served as validation. There was no significant difference between the two groups ( Table 1).

| Statistical analysis
Whole numbers and proportions were used to describe categorical variables, and means with interquartile ranges (IQRs) were used to describe continuous variables unless indicated otherwise. Chi-square test and Fisher's exact test for categorical variables, and Student's t test for continuous variables were applied to compare baseline characteristics. Multivariable logistic regression was calculated to determine the potential risk factors associated with BM. Overall survival was defined as the length of time from diagnosis to any cause of death. Kaplan-Meier method was used to calculate the OS, and differences were examined by the log-rank test. Variables associated with OS were collected based on clinical importance and predictors identified in previous articles. Cox regression analysis was performed to identify the associations of relevant clinicopathological variables with OS. Hazard ratios (HR) with 95% confidence intervals (CIs) were then calculated. 15 The validated variables were incorporated into the nomogram to predict the probability of 6-, 9-, and 12-month OS rates for patients with melanomas and BM at initial diagnosis using the rms package in R software. 16 The performance of the model was assessed by the discriminating ability and calibration ability. The discrimination of this model was assessed by C statistics. 17 Calibration was evaluated using a calibration plot, which compared the actual probabilities and the nomogram-predicted probabilities. Bootstrap sample 12 was used to correct for overfitting bias. Kaplan-Meier curves were used to further assess calibration by plotting over stratified patient scores predicted from nomograms in the dataset. 12 A twotailed P < .05 was considered statistically significant. Statistical analyses were performed with the R 3.4.3 software.

| Nomogram and model performance
Nomogram predicting OS of the melanoma patients with BM is presented in Figure 4. The following eight

F I G U R E 2 Multivariate logistic
regression for melanoma with brain metastases. CI, confidence interval; Surg, surgery independent prognostic factors were incorporated into the nomogram: age, marital status (married or unmarried), surgery to primary sites (yes or no/unknown), radiation (yes or no/unknown), chemotherapy (yes or no/unknown), bone metastases (yes or no), liver metastases (yes or no), and lung metastases (yes or no). Lower total points based on the sum of the nomogram were related to a better prognosis. For example, a melanoma patient with 60 years of age, unmarried status, evidence of bone metastases, and surgery of primary sites would have a total of 120 points assigned (50 points for age, 32.5 points for unmarried status, 0 points for surgery to primary sites, 37.5 points for bone metastases), for a predicted 6-, 9-, and 12-month OS of 65.0%, 57.5%, and 50.0%, respectively. Kaplan-Meier curves based on the predicted probability of OS were plotted to further evaluate the discriminative ability of the model. These curves were stratified by the predicted probability of the group calculated from the nomogram: low-risk group, middlerisk group, and high-risk group. Patients of high-risk group had a substantially worse outcome compared with patients of low-risk group and middle-risk group (P < .001) ( Figure  5). The nomogram-predicted median OS revealed good estimation when compared with the actual survival based on Kaplan-Meier method. A C statistics of 0.716 (95% CI 0.695-0.737) was used to assess the discrimination of the model. The accuracy of the model and potential model overfit were assessed by two approaches: (a) comparison between the training cohort, used to create the nomogram, with the validation cohort ( Figure 6A-C), (b) bootstrap validation with 1000 resamplings from the whole population ( Figure 6D-F). The calibration plots, displaying the probability of survival at 6, 9, and 12 months after diagnosis, showed a strong correlation between the nomogram-predicted probabilities and the observed probabilities.

| DISCUSSION
Lung cancer, breast cancer, and melanoma are the three most common sources of BM in patients. 18 The lower morbidity of melanoma reflects the high brain metastasis tendency; and therefore, renders greater importance to developing more effective therapies. However, further understanding of the clinical characteristics, risk, and prognostic factors of BM in melanoma is still required to improve research in this area. The salient and novel findings of our study were as follow. First, to our knowledge, this study contained the largest cohort of MBM patients. In the initial diagnosed melanoma cohort, 1.47% had BM. Melanoma patients who were presence with bone or lung metastases had greater opportunities of developing BM at diagnosis. Secondly, melanoma with BM was a F I G U R E 3 Univariable and multivariate Cox regression for analyzing prognostic factors for patients with melanoma and brain metastases at initial diagnosis | 7583 LIU et aL.
fatal disease with a median OS of 4 months. Generally, with melanoma there appears to be a relationship between parenchymal CNS metastases and leptomeningeal metastasis (LM): 87%-96% of melanoma patients with BM are associated with LM. 19,20 Compared to the median OS of 4 months for MBM patients found in our study, outcomes for melanoma patients with LM are worse with a median OS of 6-8 weeks without tumor-specific treatment. On the other hand, LM-directed treatment including targeted therapy and immunotherapy may only prolong the median OS to 1.7-2.5 months. 19,20 However, further exploration of this comparison is limited due to the lack of detailed information about parenchymal CNS metastases F I G U R E 4 Nomogram predicting survival in patients with melanoma and brain metastases at initial diagnosis. The nomogram to predict overall survival was created based on eight independent prognostic factors F I G U R E 5 Kaplan-Meier curves demonstrating survival in patients with melanoma and brain metastases at initial diagnosis according to groups of predicted survival. A, Kaplan-Meier curves demonstrating survival of the validation cohort; B, Kaplan-Meier curves demonstrating survival of the whole cohort and LM in the SEER database. Third, in this study, we constructed a nomogram that numerically facilitated individualized prediction of OS in melanoma patients with BM at initial diagnosis relied on patient-related and tumor-related factors. It can be used for patient-consulting on prognostic information, as well as to help physician make individualized clinical decisions combining with the AJCC staging system. Indeed, when stratified into groups, the nomogram was able to identify distinct groups of patients having different risks of death. Most importantly, our nomogram presented good discriminative ability, with a C statistic of 0.716 (95% CI 0.695-0.737). Accurate risk stratification of patients with melanoma is important due to the heterogeneity of patient prognoses.
This study is the first study to predict the prognosis of MBM through a nomogram. However, there are still several limitations to our study. First, the SEER database underestimates the total diagnosed cancer cases because it only collects information on newly diagnosed cancer cases. Consequently, patients who have BM later in their disease course would not be included in the data. This explains the lower proportion of BM, 1.47%, found in our study compared to that observed in the whole melanoma population, approximately 10%. 5,21,22 It is important to know that the risk and prognostic factors of BM only reflect the characteristics of this cohort, which are not equal to those for the whole melanoma population. Therefore, further studies are needed to explore the potential differences between this cohort and those who were diagnosed with BM during follow-up. The second study limitation is the unavailable information on the number of BM, Karnofsky performance status, comorbidities, extracranial disease, LM, aggregate brain tumor volume or BRAF status, some of which have been confirmed to be potential important prognostic indicators. [23][24][25] Third, although our nomogram was validated using two methods and presented strong agreement between the nomogram-predicted survival and the actual survival, external validation is needed in the future to validate the recommended nomogram. F I G U R E 6 Calibration plot comparing predicted and actual survival probabilities at 6-, 9-, and 12-mo followup. A-C, The plot for the prediction of 6-, 9-, and 12-mo overall survival between training cohort and validation cohort. D-F, The 1000-sample bootstrapped calibration plot for the prediction of 6-, 9-, and 12-mo overall survival by validation. The red line represents the ideal fit; the black line represents the actual fit

| CONCLUSION
In conclusion, our study developed a convenient nomogram, which may offer prognostic assessment for individual MBM. Additional studies are required to determine whether it can be applied to other patient groups.