Epileptic Seizure Detection in EEG Signals: A Survey of Deep, Graph, Federated and Explainable Learning

Epilepsy affects tens of millions of people, and manual inspection of long electroencephalogram (EEG) recordings delays diagnosis and limits continuous monitoring. Although deep learning has advanced automated seizure detection and prediction, reported performance is mostly subject-specific, single-dataset and segment-based, and rarely reflects clinical conditions. This paper surveys 22 studies published from 2024 to August 2026 covering convolutional, recurrent, transformer, graph-based, federated and explainable approaches, and compares their methods, datasets, results and limitations. Accuracies of 94-99% are common on CHB-MIT and Bonn, and a pooled meta-analysis reports sensitivity of 0.96 and specificity of 0.97. However, accuracy falls to about 90% in cross-subject testing, event-based metrics are seldom reported, heavy transformer and graph models are rarely tested on edge hardware, and explainability is not validated by neurologists. Eight research gaps are identified, leading to a problem statement for a generalisable, lightweight and clinically interpretable detector suited to wearable and hospital monitoring.