Radiogenomics for Glioblastoma Survival Prediction: Integrating Radiomics, Clinical, and Genomic Features Using...

Area of Science:

• Neuro-oncology
• Artificial Intelligence in Medicine
• Radiomics

Background:

• Glioblastoma (GBM) is a highly aggressive brain tumor with variable prognosis.
• Accurate survival prediction is crucial for personalized treatment strategies.
• Current prognostic models often lack integration of diverse data types.

Purpose of the Study:

• To develop and validate an advanced machine learning model for improved glioblastoma survival prediction.
• To integrate imaging phenotypes, clinical characteristics, and molecular markers for a comprehensive prognostic assessment.
• To compare the performance of a dense neural network (Dense NN) against traditional machine learning algorithms.

Main Methods:

• Analysis of two multi-institutional cohorts (UPENN-GBM and UCSF-PDGM) comprising isocitrate dehydrogenase (IDH) wild-type GBM cases.
• Extraction of radiomic features from MRI data using specialized tools.
• Application of machine learning algorithms including Random Forest, XGBoost, LightGBM, and an optimized Dense NN.
• Hyperparameter tuning and model optimization for the Dense NN architecture.

Main Results:

• The optimized Dense NN model achieved superior predictive performance compared to tree-based algorithms.
• Concordance indices (CI) of 0.86 (UPENN-GBM) and 0.83 (UCSF-PDGM) were obtained with the Dense NN.
• MGMT promoter methylation was associated with significantly longer median survival (504 days vs. 329 days).

Conclusions:

• An integrative, non-invasive approach combining AI, imaging, clinical, and molecular data enhances glioblastoma survival prediction.
• The optimized Dense NN model offers a robust framework for personalized therapeutic strategies in GBM.
• This methodology advances precision medicine by improving prognostic accuracy for GBM patients.