Deep learning models for predicting the survival of patients with medulloblastoma based on a surveillance, epidemiology, and end results analysis

Deep learning models for predicting the survival of patients with medulloblastoma based on a surveillance, epidemiology, and end results analysis


Play all audios:

Loading...

Medulloblastoma is a malignant neuroepithelial tumor of the central nervous system. Accurate prediction of prognosis is essential for therapeutic decisions in medulloblastoma patients. We


analyzed data from 2,322 medulloblastoma patients using the SEER database and randomly divided the dataset into training and testing datasets in a 7:3 ratio. We chose three models to build,


one based on neural networks (DeepSurv), one based on ensemble learning that Random Survival Forest (RSF), and a typical Cox Proportional-hazards (CoxPH) model. The DeepSurv model


outperformed the RSF and classic CoxPH models with C-indexes of 0.751 and 0.763 for the training and test datasets. Additionally, the DeepSurv model showed better accuracy in predicting 1-,


3-, and 5-year survival rates (AUC: 0.767–0.793). Therefore, our prediction model based on deep learning algorithms can more accurately predict the survival rate and survival period of


medulloblastoma compared to other models.


Medulloblastoma is an embryonal tumor that arises from the cerebellum and has the potential to spread throughout the nervous system. It is the most common type of paediatric embryonal tumor,


with an incidence ranging from 5 to 11 cases per 1 million individuals1,2. According to current international consensus, there are four subgroups of medulloblastoma: Wingless (WNT), Sonic


Hedgehog (SHH), group 3 (G3), and group 4 (G4)3. Multimodal therapy, which includes surgery, external beam irradiation, and/or cytotoxic chemotherapy, can result in survival rates ranging


from 50 to 80% based on clinical staging4. Certain prognostic features, such as age at diagnosis, extent of resection, histological subtype, and molecular subgroup classification, have been


found to affect survival predictions in individual patients.


Previous studies have used the Cox proportional-hazards model (CoxPH) to evaluate the survival rate of medulloblastoma patients5,6,7. This model incorporates survival outcomes and time as


target variables, allowing for the simultaneous analysis of multiple factors' impact on survival time. It is extensively used for predicting outcome events when the survival distribution of


the analyzed data is unknown8. A nomogram is a commonly used method for quantifying and combining important clinical characteristics of patients to calculate the probabilities of outcome


events based on the CoxPH model9. However, the model assumes that each predictor variable has the same effect throughout the follow-up time, which ignores variations in their impact on


individual patients at different times. Therefore, a new method is required to improve the accuracy of predicting the survival rate of cancer patients.


In recent years, computer and information technology have shown revolutionary potential for artificial intelligence (AI) in the healthcare industry10,11,12. Machine learning models have


stronger nonlinear modeling capabilities compared to traditional linear models and can better capture complex relationships among clinical variables. The analysis of these models can provide


accurate personalized survival predictions and decision-making support for treatment strategies to improve patient survival rates13,14. Deep learning is a subfield of machine learning that


involves discovering the distributed features of sample data by learning the underlying laws and levels of representation15,16. Neural networks are at the heart of deep learning algorithms


and consist of input, hidden and output layers that can be used to solve complex, multi-factor and non-linear problems. Deep learning-based models have become highly effective predictors of


clinical outcomes across various disease domains due to the continuous advancements in deep learning research techniques and the abundance of biomedical big data. Jiang et al.17 demonstrated


the use of an artificial neural network model to predict the survival rate of patients diagnosed with pancreatic neuroendocrine neoplasms, by leveraging clinical information. Katzman et


al.18 integrated deep learning with a multilayer neural network architecture, known as the DeepSurv model, resulting in a personalized treatment recommendation system that showed remarkable


performance.


To our knowledge, there is a lack of research combining deep learning techniques with the study of medulloblastoma. Therefore, this study aimed to fill this research gap by utilizing data


obtained from the Surveillance, Epidemiology, and End Results (SEER) database, which contains information on patients diagnosed with medulloblastoma in the United States. And then the


DeepSurv model was used to evaluate their survival rates.


The data of this retrospective cohort study were obtained from the SEER database, which encompasses information from 18 cancer registries representing approximately 28% of the entire US


population19. This database offers extensive and detailed patient data, including demographic characteristics, tumor-related information, cause of death, and survival duration. The SEER*Stat


software (version 8.3.6) was used to identify patients with medulloblastoma. The dataset covering the years 2000 to 2019 in the United States was accessed.


The patients included in the study had to meet the following criteria: (1) a confirmed pathological diagnosis of medulloblastoma; (2) identification of medulloblastoma cases based on the


third edition of the International Classification of Diseases for Oncology (ICD-O3) using specific ICD-O-3 codes for histopathology, including 9,470/3 for medulloblastoma, NOS; 9,471/3 for


desmoplastic nodular medulloblastoma; and 9,474/3 for large cell medulloblastoma. Furthermore, patients were required to have a known survival status and time. Afterwards, they were randomly


divided into a training group and a testing group at a 7:3 ratio. A flowchart in Fig. 1 illustrates the process of patient selection.


Study profile and analysis pipeline. Patients with a diagnosis of medulloblastoma as primary tumor in the SEER database 2000–2019 with complete follow-up data. The entire dataset was divided


7:3 into training (n = 1,625) and test (n = 697) sets. The CoxPH and RSF models were constructed directly from the training set data. When constructing the Deepsurv model, we used grid


search and fivefold cross-validation for hyperparameter tuning on the training dataset. Finally, the performance of the models was evaluated in the testing set (n = 697) using several


metrics.


Several parameters were collected from the samples, including age at diagnosis, sex, race, histological type, tumor size, surgery, chemotherapy, radiation therapy, and survival time. To


evaluate the prognostic value of age and tumor size in patients with medulloblastoma objectively, the patients were categorized into two groups based on the optimal cutoff values obtained


using the X-tile software (https://x-tile.software.informer.com, Yale School of Medicine, New Haven, CT, United States). Age cutoff values of ≤ 3 years and > 3 years, and tumor size cutoff


values of ≤ 3.4 cm, > 3.4 cm, and/or unknown were utilized. For detailed visual representations, please refer to Fig. 2.


The X-tile analysis was conducted to determine the best cutoff points for the variables of age and tumor size. (A) X-tile plot of training sets in age. (B) The cutoff point highlighted using


a histogram of the entire cohort. (C) Kaplan–Meier plot showing the distinct prognosis determined by the cutoff point. (D) X-tile plot of training sets in tumor size. (E) The cutoff point


highlighted using a histogram. (F) Kaplan–Meier plot showing the prognosis determined by the cutoff point. The low subset is depicted in gray, while the high subset is shown in blue.


This study selected three models for training: DeepSurv, RSF, and CoxPH. DeepSurv is a deep feedforward neural network used to predict patients’ survival time or survival probability. It


employs a multi-layer neural network to capture the complex nonlinear relationship between patients’ survival probability and input features. This study utilized deep-learning calculations


based on the DeepSurv calculation method described by Katzman et al.18 to predict the survival outcome of patients diagnosed with medulloblastoma. The term RSF refers to Random Survival


Forests, which is a survival analysis method based on random forests. When constructing a random survival forest, subsets of samples and features are randomly selected, and multiple decision


trees are built using these subsets20. Each decision tree splits the samples based on features in the nodes and determines the optimal splitting based on the evaluation of survival time


differences. The predictions from multiple decision trees in the random survival forests are combined to obtain the final survival prediction. The CoxPH is a semi-parametric regression model


used to analyse survival data and estimate the risk of event occurrence. The Cox proportional-hazards model is used to compare the relative risks of events between different groups and


study the impact of various factors on event occurrence. The model functions by modeling the relationship between time and event occurrence as a function of hazard ratios.


We performed hyperparameter tuning in the Deepsurv model using grid search and fivefold cross-validation on the training dataset, selecting the parameter with the highest average C-index in


the cross-validation as the optimal parameter.


For the implementation of the algorithms in this research, CoxPH and RSF were implemented using the Python package “Scikit-learn (version 0.24.1)” and DeepSurv was implemented using the


open-source Python package “Tensorflow-gpu (version 2.6.2)”.


The study evaluated the model’s performance using several metrics, including C-index, Brier score, integrated brier score (IBS), receiver operating characteristic (ROC) curves, and area


under the curve (AUC) values.


The C-index is a commonly used metric for evaluating the accuracy of survival predictions21. It measures the concordance or correlation between the predicted survival risk and the actual


observed survival time. A C-index of 0.5 indicates random predictions, while a value of 1.0 indicates perfect predictions. The Brier score assesses the mean squared difference between the


observed patient statuses (event occurrence or censoring) and the predicted survival probabilities. It ranges from 0 to 1, with 0 indicating a perfect match between predictions and


observations. In practice, models with Brier scores less than 0.25 are considered useful22,23. The IBS is a metric that evaluates the overall performance of a survival model across all


available time points24. It takes into account the model’s sensitivity and specificity to time-dependent events, providing a comprehensive measure of predictive accuracy. Receiver Operating


Characteristic (ROC) curves are frequently used to assess a model’s sensitivity and specificity at various discrimination thresholds. The ROC curve plots the true positive rate against the


false positive rate. The Area Under the Curve (AUC) values, which range from 0 to 1, are computed to quantify the overall performance of the model. A higher AUC indicates better


discrimination ability. This study calculated AUC values to assess the model's performance at different time points: 1, 3, and 5-year survival rates.


In the clinical data, continuous variables are expressed as mean ± standard deviation (SD), while categorical variables are described using frequencies and percentages. Statistical tests


such as chi-square tests and unpaired t-tests are used to compare variables between groups.


This study analysed data from 2,322 medulloblastoma patients registered in the SEER database between 2000 and 2019. Table 1 presents the demographic features of the patients, with 869 cases


(37.42%) being female and 1,453 cases (62.58%) being male. The racial distribution was as follows: 185 patients (7.97%) were Black, 1,939 (83.51%) were White, and 198 (8.53%) belonged to


other races. Regarding the subtypes of medulloblastoma, 329 patients (14.17%) had desmoplastic/nodular medulloblastoma (DMB), 1,866 (80.36%) had medulloblastoma, not otherwise specified (MB,


NOS), and 127 (5.47%) had large-cell/anaplastic medulloblastoma (LC). In terms of surgical interventions, 1,616 patients (69.60%) underwent total resection, 244 (10.51%) underwent subtotal


resection, 343 (14.77%) underwent local excision or biopsy, and 119 (5.12%) did not undergo surgery. Of the patients, 1,849 (79.63%) received chemotherapy, 1,766 (76.06%) underwent radiation


therapy, and 713 (30.71%) died. The cutoff values for age and tumor size were determined using X-tile analysis (Fig. 2). Specifically, 324 patients (13.95%) were ≤ 3 years old, and 1,998


patients (86.05%) were older than 3 years. Regarding tumor size, 314 patients (13.52%) had tumors ≤ 3.4 cm, 1,269 patients (54.65%) had tumor size > 3.4 cm, and the tumor size was unknown


for 739 patients (31.83%).


The predictive model was generated by partitioning the complete dataset into two mutually exclusive subsets. 70% of the dataset was allocated for the training set, while the remaining 30%


was used for the testing set. Model generation was performed on 1,625 randomly assigned patients from the training set, while the accuracy of the model was estimated using 697 randomly


assigned patients from the testing set. No statistically significant differences in characteristics were found between the two groups (refer to Table 1). Additionally, survival outcomes


showed no differences between the two groups (refer to Fig. S1).


The CoxPH model was developed using the training set (refer to Fig. 3). Only variables that showed statistical significance in the univariate analysis were included in the multivariate


analysis. The survival of medulloblastoma patients was significantly affected by non-surgical treatment, LC, white race, tumor size ≤ 3.4 cm, total resection, age > 3 years, chemotherapy,


and radiotherapy. Furthermore, the survival of the patients was significantly associated with these features in the multivariate analysis. The collinearity analysis also revealed a high


correlation between age and radiotherapy, as well as between chemotherapy and radiotherapy (refer to Fig. S2). Ultimately, we included seven features (age, race, tumor size, histological


type, surgery, chemotherapy, and radiotherapy) in the model development.


Univariate & Multivariable CoxPH analyses. Variables are sorted in descending order of hazard ratio (HR). Red represents a value above 1, while blue represents a value below 1. *p