A Taxonomy-Driven Survey of AI for Seizure Detection: Thalamic Signals, Phase Dynamics, and Translational Gaps

Seizure prediction and postictal recovery remain critical challenges in epilepsy care, particularly in real-world, resource-constrained settings. This survey presents a taxonomy-driven synthesis of 150+ peer-reviewed studies spanning seizure phase modeling, thalamic EEG biomarkers, edge inference, and clinical AI integration. We introduce an eight-axis framework covering neurophysiological foundations, machine learning advances, wearable inference pipelines, large language model (LLM) assistants, and privacy-preserving architectures. A key focus is the emerging role of thalamic stereo-EEG (SEEG) as a high-fidelity substrate for modeling seizure transitions and informing closed-loop interventions. Unlike prior reviews, this work explicitly unifies preictal and postictal phase dynamics with modern AI tools—such as federated learning, explainable deep learning, and agentic reasoning. We highlight gaps in dataset diversity, clinical interpretability, and cross-center generalization, while proposing a translational roadmap toward ethical, explainable, and deployment-ready seizure care.