Computational Approaches To Epilepsy: Graph Networks For Localization And Multi-Model Approach For Prediction

Epilepsy is one of the world's most common neurological diseases, affecting over 50 million people. Nearly a third of these individuals have drug-resistant epilepsy (DRE), a condition characterized by a lack of response to antiseizure medications. For these individuals, the primary remaining treatments are surgical intervention or neuromodulation. The success of these interventions depends mainly on solving two critical challenges: first, precisely localizing the seizure onset zone (SoZ) for resection or stimulation targets, and second, accurately predicting seizures for advisory or therapeutic devices. This dissertation addresses both challenges by developing and validating two distinct computational frameworks based on the retrospective analysis of intracranial EEG (iEEG) recordings. First, this thesis establishes a generalizable, biomarker-driven framework for identifying the SoZ. It demonstrates that epileptic networks possess common, population-level signatures. By leveraging effective connectivity measures based on frequency-domain convergent cross-mapping (FD-CCM) and graph centrality, a robust model is developed. This approach is validated on challenging interictal recordings, demonstrating that integrating these novel graph features with established biomarkers in a hybrid classifier achieves state-of-the-art performance with 90% accuracy, providing a robust and generalizable tool for pre-surgical evaluation. Second, the dissertation pivots to the challenge of seizure prediction, where it develops a patient-specific, multi-model framework to account for the high degree of intra-subject seizure heterogeneity. This approach uses unsupervised learning to cluster seizures based on their unique preictal signatures and trains an ensemble of specialized classifiers. The power and flexibility of this framework are demonstrated through two distinct feature engineering approaches. Its effectiveness was first validated using a comprehensive set of power spectral density (PSD) features, achieving a mean sensitivity of 98.54% and reducing the false positive rate by nearly 50% compared to the unclustered approach. In parallel, a novel biomarker, the Absolute Mean Instantaneous Frequency Difference (AMIFD), was introduced and also demonstrated strong predictive performance (92.08% sensitivity), confirming the framework's robustness.

Author dc:creator

• Balaji, Sai Sanjay

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/11299/279267

OAI identifier oai:identifier

oai:conservancy.umn.edu:11299/279267