An ensemble learning approach for ai-based classification of paraganglioma/ pheochromocytoma, low grade glioma, and glioblastoma tumors

Abstract

AIM: To propose a weighted vote-based ensemble classification method to classify paraganglioma/pheochromocytoma, low-grade glioma, and glioblastoma tumors-conditions that present with similar symptoms-against other central nervous system tumors using clinical and molecular data. MATERIAL and METHODS: This study utilized clinical and molecular data from The Cancer Genome Atlas database of the United States National Cancer Institute. Initially, categorical variables were transformed into numerical values, and class distribution imbalance was addressed through oversampling. The dataset was split, with 80% used for training across 10 different classical classification algorithms and the remaining 20% reserved for testing. A weighted vote-based ensemble classification algorithm was developed using six classifiers, artificial neural networks, logistic regression, extra trees, random forest, gradient boosting, and extreme gradient boosting, selected for their high classification accuracy. Additionally, feature importance analysis identified the most critical risk factors within the dataset. RESULTS: The proposed algorithm achieved an accuracy of 90.4% and an area under the receiver operating characteristic curve of 0.968, indicating strong classification performance. CONCLUSION: The findings from this study suggest that the proposed method could be a valuable tool for supporting treatment planning in central nervous system tumor cases.