Preoperative Risk-Stratication of Pediatric Malignant Brainstem Glioma based on a validated nomogram: A SEER based Analysis

Purpose: Pediatric patients diagnosed with brainstem malignant gliomas (BSMGs) have a poor prognosis. Our study aimed to construct and validate a prognostic nomogram to predict the cancer ‐ specic survival preoperatively (CSS) rates and then a risk-stratication given to these patients. Methods: From 1998 to 2016, we extracted patients' data from Surveillance Epidemiology and End Results (SEER) database. A total of 1160 patients were enrolled and randomly divided into training and validating groups. Subsequently, the Cox regression analysis was used to screen variables. Then, the nomogram was constructed. Lastly, we calculated C-indexes and plotted calibration curves and the utility of decision curve analyses (DCAs) to assess our survival model's benets. Result: Here, after multivariate cox regression analysis, we established four variables for constructing nomogram for CSS rates. Subsequently, the C-index, the area under the receiver operating characteristic curve, and calibration curves were used to conrm the nomogram's good performance. DCAs of the nomogram indicated that both groups obtained good 1-, 3-, and 5-year net benets. Conclusion: The nomogram model for preoperatively predicting CSS provided a convenient and practical tool to assess pediatric patients' prognosis with BSMG.

nomogram to preoperatively predict patient survival rate, guide clinical prognosis and treatment decisions.

Methods
We extracted patients' data from the SEER database , which was released in April 2019, based on the November 2018 submission using the SEER*Stat software version 8.3.8. The inclusion criteria for all enrolled cases included: primary-site codes; C71.7, brain stem; and morphology codes such as 9380/3 (Glioma, malignant), 9401/3 (Astrocytoma, anaplastic), 9440/3 (Glioblastoma, NOS), and 9451/3 (Oligodendroglioma, anaplastic). This study focused on patients who were diagnosed between 1998 and 2016 and were younger than 18 years. On the other hand, the following exclusion criteria were used: 1. Survival months were unknown; 2. Not the rst tumor.
Subsequently, we extracted the following variables: age, sex, race, age/years at diagnosis, histologic type, survival months, and cause-speci c mortality. Consequently, we grouped all variables into different subgroups to make data suitable for analysis using the Rstudio software and to optimize the model. These subgroups included: sex (male, female), race (white, black, and other), age at diagnosis (<= 9, > 9), tumor size (<= 25mm, >25mm, and Unkown), and histologic type (low-grade, high-grade, and mixed), except survival months (de ned as a continuous variable). We used the X-tile software to de ne the groups of age/years at diagnosis and tumor size. Our primary outcomes were cancer-speci c survival (CSS), and all patients were randomly divided into training and validation groups at a ratio of 7:3. Then, a log-rank test was applied to assess the survival difference between these two groups. Afterward, we did a multivariable Cox and backward stepwise Cox regression analysis to pick out prognostic variables in the training group to t the model. Thereafter, the C-index was calculated in both the training and validation groups to assess internal validation. Later on, the following packages such as; "rms", "foreign", and "survival" in RStudio software were used in building nomogram for predicting survival rates of CSS in pediatric patients with BSMGs and were over 1-, 3-, and 5-year.
Our study also applied a calibration method with bootstrapping values to demonstrate the correlation between the actual and the predicted probability, and the operation characteristic curve (ROC) of both the training and the validation groups were plotted. Lastly, the area under the receiver operating characteristic curve (AUC) was used to assess the nomogram's prediction accuracy, whereas the decision curve analyses (DCAs) were plotted to test the clinical value of the predictive models. Notably, all of the statistical analyses were performed using the RStudio software version 1.3.1093 (http://www.rstudio.com) Our study protocol was approved by the Ethical Committee of the rst hospital a liated to Jinan University, and all data deposited in the SEER database are freely available to all users without any restrictions. Table 1 illustrates the baseline demographic characteristics of the two groups. From 1998 to 2016, a total of 1160 patients matched our inclusion criteria as outlined above. We used the X-tile software to calculate the following optimal cutoffs; for the age at diagnosis ( Figure 7A), tumor size ( Figure 7B), and years at diagnosis ( Figure 7C). Of all patients, 819 (70.60%) patients were younger than nine years, whereas the remaining 341 (29.40%) were older than nine. Among these patients, 562 (48.45%) were male, and a total of 310 patients (26.72%) were diagnosed before the year 2003. Moreover, white was the predominant race. The tumor size in 166 (14.31%) patients was smaller than 25mm, whereas that of 449 (38.71%) patients was unknown. In 19 (1.64%) patients, the tumors were located in the left, and 18

Patient Baseline Characteristics
(1.55%) in the right.

Multivariate cox regression results
At a ratio of 7:3, patients were randomly divided into training and validating groups, respectively. Next, we used the log-rank test to explain no signi cant survival difference between the training and validation groups (p=0.9). Subsequently, signi cant variables in the results of the stepwise-backward Cox regression model and multivariate-Cox regression analyses of CSS were entered into the construction of a nomogram. As illustrated in Figure 1 and Figure 2 (A-D), signi cant variables contained age at diagnosis, years at diagnosis, tumor size, and histological grade.

Nomogram construction and validation
As illustrated in Figure 2, the constructed nomogram showed that tumor size was the most vital factor in uencing prognosis and was followed by the histological grade, age at diagnosis, and years at diagnosis, respectively. Subsequently, we applied the C-index to assess the performance of the nomogram. Here, the CSS prediction results showed that the C-index value in the training group was 0.649, whereas that of the validating group was 0.652. Among the training groups, 1-, 3-, and 5-year relative survival AUCs were 0.695, 0.803, and 0.803 ( Figure 3A). On the other hand, as illustrated in Figure   3B, the validation group AUC values for the different years were 0.680, 0.835, and 0.843, respectively. As demonstrated in Figure 4, the calibration curves show excellent consistency between nomogram in CSS predictions and actual observations in 1-, 3-, and 5-year survival rates. Moreover, as shown in Figure 5, the DCA curves of 1-, 3-, and 5-year show that the CSS model could generate net bene ts in the training and validation groups. Additionally, we conducted a risk group strati cation of the total risk score calculated using the nomogram. Pediatric patients with BSMG were divided into two categories: low risk (total points <= 92.3) and high risk (total points > 92.3). A Kaplan-Meier product-limit estimate was applied to estimate the survival curve between two risk categories ( Therefore, there is a need to discover a new complete prognostic scoring system for an accurate and useful prognosis assessment. A nomogram is a graphical representation of complex mathematical formulas. (Grimes, 2008) The nomogram showed much more convenience and accuracy when used in clinical applications. A higher Cindex means a better predictive value. (Wang and Li et al., 2013;Zhu and Zhu et al., 2016) The nomogram utility has been assessed in many elds, like bladder cancer, (Yang and Bai et al., 2020) esthesioneuroblastomas, (Jiang and Lin et al., 2020) non-metastatic breast cancer, (Qian and Jia et al., 2020) or in neurosurgery of cancers like glioblastoma (Li and He et al., 2020) and so forth. Of note, this is the rst research to introduce nomogram into pediatric primary BSMG before the onset of therapies.
It is commonly believed that the prognosis of cancer patients depreciates with age. Previous studies of adult patients with brainstem gliomas have shown that younger age is a prognostic factor for improved survival. (Guillamo and Monjour et al., 2001;Theeler and Ellezam et al., 2015; Esteve-Codina and Alameda et al., 2020) Doyle et al. reported that those patients whose age was more than 50 years signi cantly signi ed a worse prognosis than younger cohorts. (Doyle and Khalafallah et al., 2019) However, in pediatric patients, the situation was different. Our study showed that the patients older than nine years had a better prognosis. In our opinion, this could be due to the patient's younger age, which decreased their tolerance to the disease and treatment measures. Notably, the research by Sun et al. supported our ndings. Here, a signi cantly better prognosis was observed in patients whose age was more than ten years. (Sun and Xu et al., 2020) Regarding the size of the tumor, the large size (de ned as tumors with dimensions greater than 25mm) was an independent predictor of mortality in our patient cohort. Maxwell et al. established that the tumor size, which was larger than 34mm was a signi cantly worse predictor of mortality in his cohort. (Maxwell and Luksik et al., 2018) Large size (>2 cm), enhancement, and invasion of adjacent brainstem structures on head MRI are signs of more aggressive tectal lesions (Rengachary and Rauf, 2012). Besides, the brainstem contained an extremely high density of critical structures. (Scott, 2004;Mukherjee and Antar et al., 2020) Nevertheless, as the tumor diameter increased, these critical structures are more likely to be compressed, which partly de ned the surgiCal complexity.
Interestingly, pediatric BSMGs diagnosed after 2003 had a signi cantly better prognosis than that before 2003. It has been gradually recognized that brainstem gliomas comprise a heterogeneous group of tumors, and the patients could bene t from introducing the concept of precision and individualized therapeutic methods. (Wong and Brown et al., 2019) As the study progressed, a better understanding of the biology of BSMGs would improve treatment options and outcomes undoubtedly. (Reyes-Botero and Mokhtari et al., 2012) In previous studies, the biopsy is warranted in cases only when tumor progression. (Stark and Fritsch et al., 2005) Nevertheless, subsequent research suggested that brainstem stereotactic biopsy in children was a safe procedure and could alter the treatment recommendations. (Pincus and Richter et al., 2006) There is minimal long-term bene t from chemotherapy, and the mainstay of treatment for BSMGs is radiotherapy. After 2003, it has been demonstrated that the use of radiochemotherapy may improve treatment e cacy. (Marcus and Dutton et al., 2003;Sanghavi and Needle et al., 2003;Bernier-Chastagner and Grill et al., 2005; All of the abovementioned facts can directly lead to better survival among pediatrics diagnosed after 2003.

According to the 2016 World Health Organization (WHO) of the Central Nervous System Grading
Standard, gliomas are divided into four grades (I-IV), depending on the histological and molecular features. (Louis and Perry et al., 2016) High-grade gliomas (WHO grade III-IV) are more aggressive and invasive than low-grade (WHO grade I-II), and high-grade brainstem gliomas have shown poorer survival rates in pediatrics (Jallo and Biser-Rohrbaugh et al., 2004;Sun and Wan et al., 2013), which is in line with our ndings. Lesions with focal enhancement on head MRI in pediatric BSMGs were found mostly lowgrade glioma. (Lesniak and Klem et al., 2003) Necrosis and inhomogeneous contrast enhancement were regarded as signi cant predictors of high-grade brainstem gliomas. (Moharamzad and Sanei Taheri et al., 2018) Recurrent mutations of a regulatory histone-H3F3A, (Schwartzentruber and Korshunov et al., 2012) gain of H3K27 methylation at p16INKA4A, (Chan and Fang et al., 2013) and upregulation of MYCN in G34R/v mutated (Bjerke and Mackay et al., 2013) were associated with carcinogenesis, tumor progression, and the risk of pediatric high-grade gliomas and had a signi cant effect on survival.
We divided pediatrics patients into low and high-risk groups according to their total points calculated from the nomogram. Kaplan Meier survival analysis showed signi cant differences in cancer-speci c survival between low and high-risk groups. Thus, particular attention was given to these pediatric BSMGs with total points higher than 92.3 or tumor diameters larger than 2cm because of their poor outcomes.
In this study, we performed calibration and validation after the nomogram was constructed. Here, the receiver operating characteristics (ROC) were plotted using the calibration method, and the respective areas under the receiver operating characteristic curve (AUC) were calculated. The obtained calibration curve and high AUC values con rmed that our nomogram model was reliable, and its predicting capability was robust. Therefore, using these models, we can predict pediatric patients' survival with BSGM more accurately and provide the basis for individualized treatment for these patients. Finally, the DCAs of the nomogram in 1-, 3-, and 5-year show that net bene ts could be obtained in both the training and validation cohorts.
This study has some drawbacks. First, our study was based on the SEER database. This database lacks some vital information on tumor biological data, such as MGMT (O (6)-methylguanine-DNA methyltransferase), (Jovanović and Mitrović et al., 2019) promoter methylation, and IDH (isocitrate dehydrogenase) mutations, (de Quintana-Schmidt and Alvarez-Holzapfel et al., 2015) or H3 K27M, (Chi and Tarapore et al., 2019) which have substantial prognostic value in pediatric patients suffering from glioma. Secondly, it has no detailed description of the relationships with neurovascular structures and the precise location of the lesions in the brainstem. Thirdly, this is a retrospective study and cannot de ne causality. However, compared to other reported studies, it has the largest sample size, a strength.

Conclusion
Nomogram constructed with four variables in this study has a good performance. It is convenient, practical, and can help make the individual clinical decision and assess pediatric patients' risk with primary brainstem malignant glioma. This preoperatively predictive model was objective and accurate in predicting 1-, 3-, and 5-year CSS rates before treatment.  ROCs curve for nomograms. AUCs of the nomograms to predict cancer-speci c survival at 1-, 3-, and 5year using training dataset (A) and validation dataset (B).

Figure 4
The calibration curves in training (A, C, E) and validation (B, D, F) cohorts show the nomogram with excellent predicting ability in 1-, 3-, and 5-year cancer-speci c survival