NeuroGRIP Enhances EEG Seizure Diagnosis with Medical Knowledge

Diagnosing seizures from electroencephalogram (EEG) signals remains a significant challenge due to the complex neural dynamics and potential for spurious connections in inter-channel modeling. While spatial-temporal graph neural networks (STGNNs) have advanced EEG brain network representation, their purely data-driven nature often results in graph structures that lack clinical plausibility and interpretability. Researchers have developed NeuroGRIP, a novel retrieval-augmented graph refinement framework designed to address these limitations. NeuroGRIP incorporates external medical knowledge to calibrate and refine the noisy EEG graphs produced by STGNNs. It achieves this by constructing a large-scale, domain-specific knowledge base from authoritative clinical guidelines, leveraging large language models to extract structured biomedical entities and relations. The framework aligns STGNN-generated EEG node embeddings with this knowledge graph, performing semantic queries to retrieve relevant medical evidence. This evidence is then used to assign confidence scores to predicted edges, enabling the pruning of medically implausible connections from the original graph. Extensive experiments on TUSZ and CHB-MIT datasets demonstrate that NeuroGRIP not only boosts seizure detection accuracy but also significantly enhances interpretability by grounding predictions in clinically validated knowledge, marking a step towards explainable clinical diagnosis.