Hybrid AI for EEG-based Seizure Detection: A PRISMA-guided Narrative Review of Spatiotemporal, Multimodal, and Explainable Models
Mihir Sojitra1, Archi Dhamelia2, Manuel Alanís-Bernal3
1B.J. Medical College, Ahmedabad, India, 2MGM Medical College Navi Mumbai, 3Neurology, Vall d'Hebron University Hospital
Objective:
To evaluate recent advances in hybrid artificial intelligence models that address the inherent challenges in automated seizure detection from electroencephalography, including signal heterogeneity, noise, and complex spatiotemporal patterns.
Background:
Despite technological progress, automatic seizure detection from EEG remains challenging due to signal non-stationarity and low signal-to-noise ratios. Traditional machine learning methods are likely to fail in capturing both spatial and temporal aspects of ictal activity. Hybrid AI models that combine complementary algorithms may well offer higher detection performance.
Design/Methods:
We conducted a narrative PubMed search from 2019-2025 and identified 28 primary research articles employing hybrid AI frameworks for EEG-based seizure detection. The frameworks were categorized into spatiotemporal fusion, graph-based connectivity, explainable AI, and light-weighted detection frameworks.
Results:
Hybrid architectures demonstrate significant advances compared to single-algorithm solutions. Spatiotemporal models utilizing convolutional and recurrent neural networks achieved over 99% accuracy by concurrently extracting spatial patterns and temporal patterns of seizure activity. Graph neural networks accurately modeled dynamic functional connectivity patterns for seizure localization. The integration of attention mechanisms and explainable AI methods provided clinically interpretable insights into model decisions, emphasizing salient EEG channels and features without compromising high accuracy. Lightweight autoencoder models enabled real-time detection on wearable devices with more than 98% accuracy, meeting requirements for practical implementation.
Conclusions:
Hybrid AI architectures represent a significant leap towards automated seizure detection, even if significant challenges remain before clinical application. Future efforts must tackle computational needs of embedded systems, increase model generalizability to diverse patient populations and recording environments, and provide standardized interpretability frameworks to form clinical trust. Interdisciplinary collaboration between neurologists, data scientists, and engineers will be essential to advance these exciting technologies into clinical use as useful tools for improved patient care.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.