INTRODUCTION
Glioblastoma multiforme (GBM) isocitrate dehydrogenase (IDH)-wildtype is a highly aggressive form of primary malignant brain tumor that causes a significant challenge in terms of treatment and management. It is the most common type of primary brain and central nervous system tumor in adults, accounting for a staggering 45.2% of all malignant cases [18]. Despite advancements in neurosurgery, radiotherapy, and chemotherapy, survival rates remain low at 12 to 15 months postdiagnosis. Improved strategies are urgently needed to enhance patient outcomes and quality of life in the face of this challenging brain cancer [35]. GBM, though rare, holds considerable prominence among primary malignant brain tumors in adults, constituting about 60% of cases in this demographic, with an estimated annual incidence of approximately three to four cases per 100000 individuals in the United States [37]. The high prevalence of GBM highlights its significant implications for patients, families, and healthcare systems. Predicting overall survival (OS) in GBM is critically important because it helps in determining treatment approaches to individual patient needs. By accurately forecasting survival outcomes, clinicians can prioritize more aggressive treatments for patients with potentially shorter survival duration and consider more conservative approaches for those likely to live longer. This strategy not only maximizes the treatments efficacy but also enhances quality of life by avoiding unnecessary side effects in patients with better prognoses. Thus, OS prediction plays a vital role in optimizing therapeutic strategies and improving overall patient care in GBM treatment.
GBM, IDH-wildtype, a form of GBM, shows variable outcomes. Lower-grade IDH-wildtype astrocytomas resemble higher grades, prompting classification review [23,25]. GBM, IDH-wildtype (90% of cases) affect older individuals with a median survival of 10-15 months [12]. Treatment includes extensive surgery, radiation, and temozolomide chemotherapy, with subsequent cycles to consolidate therapy. GBM remains formidable despite these efforts, driving ongoing research for more effective treatments to enhance patient outcomes [19]. Traditional prognostic factors for GBM, IDH-wildtype, such as age and surgical extent, often provide limited insights into tumor biology.
In recent years, radiomics and radiogenomics have shown promising outcomes. Radiomics enhances OS prediction accuracy by extracting detailed quantitative features from magnetic resonance imaging (MRI) scans, revealing tumor heterogeneity and microenvironment characteristics beyond human visual perception. This technique quantifies complex tumor attributes, enriching our understanding of disease behavior and treatment response [21]. These radiomic features, encompassing aspects like tumor texture, intensity, and shape, can provide valuable insights into tumor aggressiveness and response to treatment. By integrating these features with traditional prognostic factors, radiomic models can offer a more comprehensive picture of a patient’s GBM, leading to more accurate predictions of survival outcomes [6]. GBM OS prediction is essential for treatment optimization. Radiomics delves deeper into tumor details from MRI scans, revealing hidden insights beyond conventional factors. Machine learning, such as support vector machine (SVM), analyses radiomic features to uncover links between tumor characteristics and patient outcomes. This approach can revolutionize clinical practice by enabling personalized treatment and efficient trial design for GBM, IDH-wildtype [21].
This study aimed to create a predictive model for estimating OS using radiomic data and clinical variables like age, gender, and MGMT (O6-methylguanine-DNA methyltransferase) status. We preprocessed the data to eliminate less informative features and use recursive feature elimination (RFE) to highlight crucial survival predictors. An SVM model was then be trained with k-fold cross-validation to forecast OS outcomes. The model’s effectiveness measured by a survival index calculated from a sigmoid function, correlating survival predictions with actual survival durations. Additionally, integrated clinical data with the survival index to assess how demographic factors influence predictions. Kaplan-Meier curves illustrated the model’s survival probability predictions.
Our study seeks to optimize treatment outcomes for GBM, IDH-wildtype through a comprehensive approach integrating radiomics and machine learning techniques. Through meticulous pre-processing and feature selection, we aim to develop a robust predictive model. By correlating predicted survival likelihood with actual survival days and analyzing demographic factors’ impact, the accuracy, and personalized treatment strategies could be enhanced. This approach revolutionizes clinical decision-making, offering valuable insights for improving patient outcomes in this challenging cancer subtype.
MATERIALS AND METHODS
This study aimed to create a predictive model for estimating OS using radiomic data and clinical variables like age, gender, and MGMT status. We preprocessed the data to eliminate less informative features and used RFE to highlight crucial survival predictors. An SVM model was then trained with k-fold cross-validation to forecast OS outcomes. The model’s effectiveness is measured by a survival index calculated from a sigmoid function, correlating survival predictions with actual survival durations. Additionally, integrated clinical data with the survival index to assess how demographic factors influence predictions. Kaplan-Meier curves illustrated the model’s survival probability predictions [26].
Data collection
The data used in this study is publicly available Brain Tumor AI Challenge and the University of Pennsylvania Health System (UPHS) glioblastoma (UPENN-GBM) datasets. Waiver for written patient consent was not sought from the Institutional Review Board (IRB) as it is not required by the national legislature for retrospective studies of existing data (e.g., registry data). Multi-parametric MRI (mpMRI) data from routine clinical and radiologic examinations are included in the UPENN-GBM [3,4,8] cohorts. Six hundred thirty patients from the UPENN-GBM collection were first taken into consideration. However, after applying the inclusion criteria, which called for patients to have received preoperative scans, including baseline mpMRI at the time of diagnosis, and to be at least 18 years old, a subset of 611 patients was found [3,4,8]. Based on these datasets, 574 patients were included in the analysis because OS status data were available, which is days of survival after surgery. This represents a comprehensive assessment of survival outcomes within the study population. Four structural MRI sequences were included in these scans : T1 with gadolinium (T1GD) (T1-CE is also called), native T2-weighted (T2), native T1-weighted (T1), and T2-FLAIR [3,4,8] (Fig. 1). Table 1 presents the distribution of data categorized by gender and MGMT status, along with the proportions of short-term and long-term survivals among these groups. We have considered 400 patients for the discovery cohort to get the appropriate model, and 174 patients’ data was used for independent testing (replication cohort). In the replication cohort, we added clinical parameters and radiomics.
The entire GBM and its subregions, including the tumor core (TC) labeled as 1, edema (ED) labeled as 2, and enhancing tumor (ET) labeled as 4, are identified in Fig. 2.
A boxplot representing the connection between survival days and survival outcomes classified as long and short survival is shown in Fig. 3. By comparing the distribution of survival days between the two groups, this visualization makes it possible to identify any potential variations in the length of survival.
MRI data acquisition
Within the UPENN-GBM dataset, the MRI data encapsulates patient information, with each patient linked to four distinct modalities : T1, T2, FLAIR, and T1GD. Each modality is presented in a 3D format, comprising 155 slices. In MRI imaging, various technical parameters are crucial for obtaining clear and precise images. For T1 images, the echo time (TE) and repetition time (TR) are generally shorter, enhancing the contrast between different tissues. T2 images use longer TE and TR, highlighting fluid and inflammation by making fluid appear bright. Fluid-attenuated inversion recovery (FLAIR) images, which are helpful in detecting lesions near the of the brain’s ventricles, suppress the fluid signal to make abnormalities more visible by using a long inversion time. Contrast-enhanced T1 (T1GD) images involve the use of a gadolinium contrast agent to improve the visibility of blood vessels and tumors, adjusting TE and TR similar to standard T1 images to capture the contrast uptake in tissues. These parameters help in differentiating various tissue types and pathologies, providing crucial diagnostic information [4].
Image pre-processing
Each patient in our study was described by four modalities : T1, T2, FLAIR, and T1GD (T1-CE). Loading the images and related segmentation data for each modality was the initial stage in our pre-processing workflow. Only slices with segmentation information were selected [28]. Initially, we found the beginning and ending slices in the segmented data with nonzero-pixel values greater than 300 to calculate the slice range for each modality. This was done to ensure that only slices, including the brain region, were considered, leaving out those that did not [14]. This facilitated faster feature extraction and helped lower noise in the data [22]. A bounding box of each modality’s slice was taken from the segmented image data once we determined each modality’s slice range [15,38]. This made it more likely that only the pertinent portion of each slice was analyzed, leaving out the remainder.
Noise in the data was minimized, and the feature extraction process was refined by cropping each modality’s slice according to segmented data and applying intercubic interpolation. This type of interpolation, using a piecewise cubic polynomial, estimates values between known data points, reducing noise introduced during cropping and smoothing the images for more accurate analysis [17].
Feature extraction
In this study, we used proven approaches to thoroughly extract radiomic characteristics from medical imaging data. To improve image quality and lower noise, the collected data were carefully pre-processed, which included resampling to a consistent voxel size and intensity normalization. Segmentation methods were used to delineate the regions of interest (ROI) associated with tumor lesions. From these identified ROIs, a comprehensive set of radiomic features was extracted. This included texture features such as the gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), and local binary patterns (LBP), alongside fundamental statistical metrics like mean, median, standard deviation, and skewness. These features are critical for analyzing the textural and intensity variations within the tumor, providing valuable insights into its heterogeneity (Supplementary Table 1). Next, the most informative and discriminative features were determined using a combination of statistical analysis and feature selection [20]. The next step involved feature selection, combining statistical analysis and machine learning techniques to find the most discriminative and informative features. Lastly, before being used in later analyses for predictive modeling of OS prediction, the chosen characteristics were normalized to guarantee consistency and comparability across the dataset. In our study, we were able to get important insights into tumor characteristics and prognosis because of our meticulous approach to feature extraction [9]. Detailed flow of the paper is presented in Fig. 4.
Feature selection
Low variance feature removal was a crucial pre-processing step, aimed at eliminating features with minimal discriminatory power. By filtering out features with low variance across samples, noise was reduced, and focus was placed on characteristics contributing significantly to the predictive task. This technique enhanced model interpretability and prevented overfitting, improving the robustness of radiomics-based predictive models [20].
RFE was used in the study to improve predictive modeling with SVM. With its iterative removal of the least significant features, RFE served as a smart feature selection method. A collection of features that best discriminated between clinical outcomes was successfully determined by combining RFE and SVM. This method not only reduced feature space dimensionality but also effectively captured the most important tumor shape and heterogeneity features. Consequently, RFE contributed greatly to the development of radiomics-based predictive modelling in oncology by improving model performance and interpretability [13]. Initially having 1728 radiomics features for all the modalities (T1, T2, FLAIR, and T1GD), the feature set was reduced to 1111 by removing low variance features. Subsequently, upon applying RFE, the 200 (Supplementary Table 1) were the most relevant features.
Model development
We utilized a SVM model to calculate the survival index for all patients in our dataset.
Survival index
In our study, the survival index is derived from the SVM model by calculating the distance of each case from the decision hyperplane, followed by the application of a sigmoid function to these distances. This method transforms the SVM outputs into a probability metric that quantitatively assesses patient survival likelihood, offering a critical tool for refined prognosis evaluation in clinical settings [10]. We employed a tenfold stratified k-fold cross-validation strategy [2,32] to evaluate the performance of our SVM model. The dataset was split into training and testing sets for each fold, and the SVM model was trained on the training set before being assessed for performance on the testing set. To ensure that every data point was used for testing precisely once, we repeated this procedure 10 times. The Standard Scaler from the scikit-learn module was used to apply the same scaling to both sets, preventing information leakage between them [33]. The same scaler was used to transform the training and testing sets after fitting it on the training set. Within each fold, a grid search over a range of hyperparameters for the SVM model was conducted [28]. The regularization parameter C and the kernel type (linear, poly, rbf, or sigmoid) were the hyperparameters adjusted [17,34]. To enable the estimation of class probabilities, utilized to construct the receiver operating characteristic (ROC) area under the curve (AUC) score, the SVM model’s probability parameter was set to true [2,7]. The ROC AUC score was calculated to assess the SVM model’s performance on the testing set after fitting it on the training set of each fold. Additionally, the ROC curve was calculated and plotted for every fold. Lastly, the AUC score and mean ROC curve were plotted on a different graph, spanning all folds.
Based on the greatest mean ROC AUC score overall folds, the best SVM model was chosen, and its hyperparameters and scaling technique were used to evaluate the final model on the unseen data. SVM utilized a decision function to categorize data points into distinct classes. In binary classification scenarios, this function computed the signed distance of each data point from the decision boundary, often referred to as the hyperplane. Following the training of an SVM model, the decision function generated output values for every data point in the dataset. These output values represented the confidence levels or scores assigned by the model to each data point belonging to a particular class. To derive the survival index from these output values, they were passed through a sigmoid function. The sigmoid function transformed the output values into a range between 0 and 1, providing a measure of the likelihood or propensity of each data point for a specific outcome, such as survival [16]. This process enabled the interpretation of SVM output values as survival indices, facilitating the assessment of individual data points’ relative risk or probability of survival based on the SVM model’s predictions.
Integration with clinical data
We augmented our dataset by incorporating clinical parameters such as age, gender, MGMT status, and survival days. These additional features provided valuable insights into the patient cohort, allowing for a comprehensive analysis considering demographic characteristics, molecular biomarkers, and clinical outcomes. The integrated analysis was measured through Kaplan-Meier (KM) curves.
Accuracy parameters
We computed parameters such as AUC ROC, accuracy, precision, recall, and many other evaluation metrics to offer a more thorough and trustworthy evaluation of the model’s performance. The precision determines the fraction of correct positive predictions, recall quantifies the model’s capacity to identify all positive classes accurately, which evaluates the frequency of predictions. A harmonic balance between recall and precision is reached by the F1-score, which is the harmonic mean of these two metrics. Furthermore, Log-rank test (chisquare statistics) was used to analyze the OS prediction model followed by Cohen’s d effect size.
KM curve
In our study, the KM curve is employed to illustrate the survival probabilities of glioblastoma patients based on their prognostic indices derived from the SVM model. This statistical method allows us to plot the survival functions effectively, providing a clear visual representation of the time-to-event outcomes for different patient groups. By grouping patients according to their calculated survival indices, the KM curves enable us to observe and compare the survival experiences across these groups. This visualization not only underscores the discriminative power of our SVM model in predicting patient outcomes but also serves as a crucial tool for verifying the model’s accuracy in a clinical context. The KM curves thus play a fundamental role in our analysis, offering valuable insights into the effectiveness of the survival indices at stratifying patients in terms of survival probabilities.
Programming software
In this study, we utilized Python as our primary programming language, leveraging the TensorFlow and Keras libraries to build and manage our machine learning models. These libraries are well-suited for handling large datasets and complex computations, making them ideal for developing robust predictive models in medical research. This software setup facilitated the efficient execution of our feature extraction and model training processes.
RESULTS
Model performance
The model’s separability, or how well it divides the two groups, is measured by the AUC ROC. In essence, it evaluates the model’s ability to distinguish between methylated samples and those not, or between short and long survivors.
Fig. 5 shows the ROC curve for the SVM model trained with 10-fold cross-validation. It visualizes the model’s performance in distinguishing between classes, displaying sensitivity (true positive rate) versus 1-specificity (false positive rate) across various threshold values. The shape of the curve, along with an AUC of 0.84 (95% confidence interval [CI], 0.80-0.90) with (p<0.001), provides insight into the predictive accuracy and discrimination capability of the model. With an AUC of 0.84, the model demonstrates moderate predictive performance, indicating its ability to distinguish between positive and negative cases.
To guarantee the model’s accuracy and dependability in producing useful predictions in real-world circumstances, its performance is assessed using a variety of measures.
Statistical analysis
Following Fig. 6 shows a boxplot that illustrates the relationship between the survival index and the long and short survival categories. The detection of any differences in survival index between long and short survival groups was made easier with this graphic, which helped explain how the survival index fluctuated across individuals with varied survival outcomes.
The analysis using the Log-rank test yielded a p-value of 0.00029, indicating a statistically significant difference in survival distributions among the groups studied. Additionally, the Cohen’s d-effect size was 1.20, suggesting a significant difference between short and long survivors group means of the replication cohort. These results emphasize both statistical significance and practical importance in the observed differences.
Survival analysis
The results in the following KM survival curves (Fig. 7) demonstrate a statistically significant difference in survival probabilities between two groups i.e., short and long survivals. Over the timeline, extending up to 2500 days, the group predicted to have long survival consistently shows higher survival probabilities than the short survival group. This is quantitatively supported by the p-value of less than 0.0001, indicating a highly significant difference between the survival outcomes of the two groups. Additionally, the number at risk decreases more rapidly in the short survival group over time, which aligns with their lower survival probability, further validating the effectiveness of the predictive model used to categorize these groups.
DISCUSSION
The study aimed to enhance treatment outcomes for GBM, IDH-wildtype by integrating radiomics analysis and machine learning techniques. Radiomic features were extracted from medical imaging data, followed by applying RFE to identify important predictors for survival estimation after removing low-variance features [30]. These selected radiomic features were combined with OS data and clinical variables like gender, age, and MGMT status to develop a comprehensive predictive model. Machine learning algorithm, specifically SVM, was employed using k-fold cross-validation (k=10) to train and optimize the predictive model [27] and obtained an AUC of 0.84 (95% CI, 0.80-0.90) with (p<0.001). This method ensured robustness and generalizability by assessing model performance across multiple training and testing iterations. The best performing SVM model was then used to calculate a survival index, transforming a sigmoid function to enable continuous likelihood assessment of survival probabilities.
The boxplots compared survival index and survival days distributions between short-term and long-term survival groups in GBM, IDH-wildtype patients, showing potential as a prognostic indicator for personalized treatment strategies. This visualization offered insights into the model’s accuracy in predicting actual survival outcomes. Furthermore, a classification algorithm was applied after merging clinical data with the transformed survival index to predict individual patient survival outcomes. The Log-rank test produced a p-value of 0.00029, demonstrating statistically significant differences in survival rates across the groups analyzed. Furthermore, with a Cohen’s d value of 1.20, the effect size is considered large, highlighting a significant and meaningful distinction between the means of the groups. These findings underscore the relevance and significance of the observed disparities. The study’s findings were reinforced by generating a KM curve (Fig. 7) with p-value <0.0001, illustrating the difference between short and long survivors [36]. This visualization highlighted the model’s prognostic capabilities and potential clinical utility in guiding personalized treatment strategies for GBM, IDH-wildtype patients. Comparing these findings with existing literature reveals significant advancements in GBM prognosis and treatment planning. While traditional methods predominantly rely on clinical parameters, incorporating radiomics enables a more nuanced understanding of tumor heterogeneity and its impact on patient outcomes [24]. To date, several studies have been done for OS estimation such as by Saxena et al. [28] exploring the prediction of MGMT status and OS using hybrid radiomics signatures derived from mpMRI in a machine and deep learning framework. Fathi Kazerooni et al. [10] emphasize the synergistic value of clinical measures, radiomics, and genomics for AI-based prediction of OS in GBM, IDH-wildtype, integrating conventional and deep learning methods. Das et al. [9] propose a deep learning-based ensemble model for brain tumor segmentation using multi-parametric MR scans, enhancing efficiency in tumor identification and analysis. Bakas et al. [4] introduce the UPENN-GBM cohort, providing a comprehensive dataset with advanced MRI, clinical, genomics, and radiomics information for predictive and diagnostic assessments. One more study applies ensemble machine learning to rank clinical features and predict survival in GBM, highlighting the importance of computational intelligence for personalized treatment planning [6].
The distinct contributions of our study from exsisting works are crucial to advancing personalized medicine for GBM, particularly in patients with IDH-wildtype, a notably aggressive and treatment-resistant subgroup. Our work implemented an important survival index, a sophisticated tool derived from a combination of advanced radiomic features. This index profoundly enhances the ability of medical professionals to predict patient prognosis more accurately, thereby allowing for more personalized and effective treatment strategies. By employing a comprehensive array of radiomic features extracted from multi-parametric MRI, our approach enables a nuanced analysis of tumor heterogeneity, crucial for precise survival prediction. Additionally, a sophisticated SVM classifier, enhanced by a multistage feature selection using RFE, refines the model further, improving both its predictive power and clinical interpretability. Our rigorous validation through cross-validation techniques assures the model’s reliability and applicability across different clinical settings, ensuring it can substantially impact managing this challenging cancer type.
Our study optimizes treatment outcomes for GBM, IDHwildtype patients and represents a focused and innovative approach. By employing SVM classifier with radiomics features extracted from mpMRI scans along with the clinical parameters, our work promises reasonable results in predicting OS outcomes. Integrating clinical parameters such as age, gender, and MGMT promoter methylation status, our study refines survival estimates and provides personalized treatment planning tools. Unlike existing methods, our SVM-based radiomics approach achieves a reasonable cross-validated AUC of 0.69, effectively categorizing patients into short and long survivors. We incorporated survival index to enhance the robustness of our model. The integration of clinical data further enhances prediction accuracy, demonstrating the potential of our model in improving patient outcomes [1] and guiding precise therapeutic interventions for GBM, IDH-wildtype [5]. While the study on using radiomics features to predict OS in IDH-wildtype GBM patients shows promise, there are some important points to consider for future improvements. The dataset used, which included mpMRI scans from 574 patients, may not fully represent all the variations seen in GBM, IDH-wildtype cases [28]. This could affect how well the predictive model works across different types of patients [11]. Although the study used methods like RFE to pick out the most useful radiomics features, it’s important to check if these features will work well in other patient groups or with different imaging techniques [31]. Additionally, while the model got a boost from including clinical details like age, gender, and MGMT promoter methylation status, there could be other important factors (like performance status or treatment history) that were not considered. Understanding why the model makes its predictions can also be tricky, since it is based on complex machine learning methods. Testing the model with data from other sources is crucial to see how well it holds up in different situations. Lastly, it would be helpful to connect the radiomics features more directly like how these features relate to disease behaviour or response to treatment. By working together across different areas of expertise, researchers can refine these models to guide treatment decisions for GBM, IDH-wildtype patients better, ultimately making a real impact in clinical practice. Apart from various positive outcomes, our study also includes limitation such as exclusion of extent of resection (EOR) in our study which is a one of the factor in predicting GBM outcomes. However, this is attributed to unavailability in our dataset due to its retrospective nature. In our future studies, we will address this by seeking a large multi-institutional cohort with comprehensive EOR information, which will help improve the accuracy of our model.
CONCLUSION
This study demonstrates the effectiveness of an SVM-based radiomics model in OS prediction for GBM, IDH-wildtype patients. By integrating radiomics features with clinical data, the model significantly improves personalized prognosis and treatment planning. Key findings include its ability to distinguish between short and long-term survival prospects accurately. The study highlights the need for broader validation and suggests future research should include additional biomarkers to enhance predictive accuracy. Overall, this research illustrates the potential of machine learning in enhancing clinical decisions for complex GBM treatments.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation 
Supplement
Print 
