Machine-Learning-Based Survival Prediction in Glioblastoma Using Graph-Theoretical Analysis of Structural Network...

Area of Science:

• Neuroscience
• Medical Imaging
• Computational Biology

Background:

• Glioblastoma is a highly aggressive brain tumor known for infiltrating white matter and disrupting brain connectivity.
• Current prognostic models often overlook the widespread effects of glioblastoma on the brain's structural connectome.
• This study aims to improve glioblastoma survival prediction by incorporating network analysis of brain connectivity.

Purpose of the Study:

• To integrate graph-theoretical analysis of diffusion tensor imaging (DTI)-derived structural connectomes with machine learning (ML).
• To enhance the prediction of overall survival (OS) in newly diagnosed glioblastoma patients.
• To explore the utility of connectome-derived metrics in glioblastoma prognostication.

Main Methods:

• Whole-brain structural connectomes were constructed from preoperative DTI data of 871 glioblastoma patients.
• Connectomes were weighted using tract count and quantitative anisotropy (QA).
• Graph-theoretical network metrics were extracted and combined with clinical data; ML models were trained and validated.

Main Results:

• Random forest, adaptive boosting, and KStar models demonstrated strong validation performance.
• Random forest models predicted OS beyond one year with high accuracy (0.862-0.874) and AUROCs (0.909-0.929).
• Key predictors included network strength and clustering coefficient, with significant nodes concentrated in the temporal lobe.

Conclusions:

• Graph-theoretical analysis of structural brain network disruption, coupled with ML, accurately predicts glioblastoma OS.
• These findings support a network-based understanding of glioblastoma.
• Connectome-derived metrics can potentially complement existing prognostic frameworks for glioblastoma.