Neural Mosaics: Detecting Aberrant Brain Interactions using Algebraic Topology and Generative Artificial Intelligence

Epilepsy affects over 50 million persons worldwide, with less than 50% achieving long-term success following surgery. Traditional electrophysiology signal-based seizure detection methods are resource-intensive, laborious, and overlook multifocal brain interactions. Algebraic topology methods, particularly persistent homology, offer robust representations of complex brain interaction patterns. Leveraging persistent homology and the Google Gemini Pro Vision 1.0 large language model (LLM), we present a novel prompting template to classify topological structures computed from intracranial electroencephalography (iEEG) recordings from refractory epilepsy patients. This study marks the first use of persistence diagrams as input to a LLM for analyzing brain interaction dynamics. Our results indicate that simply prompting LLMs with persistence diagrams is insufficient for accurate seizure detection. Nonetheless, unlike traditional approaches using machine learning algorithms for EEG classification, our approach does not require large volumes of representative training data or brittle hyperparameter tuning, which highlights the promise of more scalable analyses in the future.