Toward Improved Glioma Mortality Prediction: A Multimodal Framework Combining Radiomic and Clinical Features

Abstract

Gliomas, especially diffuse gliomas, remain a major challenge in neuro-oncology due to their highly heterogeneous nature and poor prognosis. Accurately predicting patient mortality is essential for improving treatment strategies and outcomes, yet current models often fail to fully utilize the wealth of available multimodal data. To address this, we developed a novel multimodal predictive model that integrates diverse magnetic resonance imaging (MRI) sequences—T1, T2, FLAIR, DWI, SWI, and advanced diffusion metrics such as high angular resolution diffusion imaging (HARDI)—with detailed clinical data, including age, sex, tumor genetic markers, and WHO CNS tumor grade. Using the UCSF Preoperative Diffuse Glioma MRI (UCSF-PDGM) dataset, our study introduces an innovative framework that integrates deep learning (e.g., VGG16 for extracting embeddings from a diverse array of MRI modalities, including standard sequences and advanced diffusion metrics) with machine learning algorithms (e.g., XGBoost) to combine imaging and clinical data. This approach captures complementary insights that surpass the capabilities of both single-modal models and previous multimodal methods, which often rely on predefined radiomic features or limited integration of data types. Our results demonstrate significant improvements in predictive accuracy for glioma mortality, showcasing the value of integrating raw imaging embeddings with detailed clinical variables. By providing a more comprehensive understanding of tumor behavior and patient outcomes, our study advances glioma prognosis and supports the development of more personalized and effective treatment strategies.