Integrated Predictive Modeling for Brain Tumor Outcomes

Abstract

Brain tumors, most commonly gliomas and meningiomas, are abnormal masses of cells growing at the brain. We employ a Random Survival Forest model and an XG Boosting model, two ensemble types, to predict brain tumor patient survival from clinical data. XGBoost outperformed RSF with C-indices of 0.7143 and 0.6928, respectively, and Cramér's V values of 0.5270 and 0.4188, respectively. Both models identified diagnosis as the most important feature and stereotactic methods as the least important. They also outperformed other prominent models tested on datasets with similar cardinality, and their accuracy can be improved with a larger sample size and the inclusion of omic data in training. The results are applicable in future ensemble model design and oncology survival forecasting.