Approximately 80,000 individuals in the United States are diagnosed annually with primary brain tumors, with gliomas accounting for nearly 25% (20,000 cases annually). Gliomas range from low-grade tumors median-survival sixteen years to high-grade malignancies glioblastoma multiforme (GBM)–median-survival eight months. In 2021, The World Health Organization released a revised classification of gliomas based on molecular features. Molecular information has greatly enhanced precision diagnosis and facilitated treatment decisions. However, as molecular profiling techniques advanced, molecular mutation landscape data for gliomas continued to expand, necessitating testing across the growing number of mutations that numbers at least twenty by now. Such expensive molecular testing imposes a massive burden on cost and resources.

Using publicly available TCGA glioma data, we implemented a voting-based feature selection framework integrating seven complementary feature selection methods. The selected features were evaluated using an ensemble of six machine-learning(ML) classifiers, producing 57 baseline ensemble configurations, which were expanded to 171 through the addition of hard voting, soft voting, and stacking strategies. The voting based and ensemble method was compared with standard Lasso-method for feature efficiency and diagnostic accuracy.

The voting-based framework reduced the optimal feature set to eight molecular and clinical variables from the original set of 23 features, compared to fourteen features selected by Lasso, while maintaining similar predictive performances. Stacking ensembles—particularly SVM+SGD and AdaBoost+SVM+SGD—achieved the highest mean accuracy (0.881), and stacking ensembles using Lasso selected features reached a comparable peak accuracy (0.882) for glioma classification.

These findings demonstrate that voting-based feature selection and ensemble method significantly outperforms Lasso in feature efficiency while maintaining the same high predictive accuracy. By using a streamlined molecular feature set, this approach achieves cost-effective glioma grading with high diagnostic accuracy, while improving equity in healthcare access.

Sarvesh Amatya is a tenth-grade student at Alexander W. Dreyfoos School of the Arts in West Palm Beach, Florida, where he is currently ranked first in a class of 365 students. He has been pursuing research in artificial intelligence since seventh grade and has presented his work at the Junior Science and Humanities Symposium (JSHS) and local chapters of the International Science and Engineering Fair (ISEF). This year, he will present his research at the Florida state level. In addition to his academic pursuits, Sarvesh is an accomplished pianist who has received awards in national competitions and performed at Carnegie Hall in New York.

For this project, he worked under the mentorship of Dr. Ashoo Joshi and will be presenting the research on her behalf.