A Review of Machine Learning and Deep Learning Trends in EEG-Based Epileptic Seizure Prediction

Wu, Yitong ORCID: https://orcid.org/0009-0006-2103-5837, Su, Eileen Lee Ming ORCID: https://orcid.org/0000-0001-9366-5404, Wu, Mingyu, Ooi, Chia Yee ORCID: https://orcid.org/0000-0003-2307-4048 and Holderbaum, William ORCID: https://orcid.org/0000-0002-1677-9624 (2025) A Review of Machine Learning and Deep Learning Trends in EEG-Based Epileptic Seizure Prediction. IEEE Access, 13. pp. 159812-159842. ISSN 2169-3536

Preview

Published Version Available under License Creative Commons Attribution. Download (3MB) | Preview

Abstract

Epileptic seizures impair patients’ health and quality of life, and electroencephalography (EEG)-based prediction enables timely intervention. Early work on epileptic seizure prediction employed linear models, whereas recent studies employ more advanced machine learning and deep learning models. We review 149 studies (2018—May 2025), summarizing trends and gaps across the full pipeline in EEG-based epileptic seizure prediction. We find that four persistent challenges related to this domain remain. First, severe dataset imbalance, with overreliance on Children’s Hospital Boston–Massachusetts Institute of Technology EEG dataset (CHB-MIT), minimal use of intracranial electroencephalography (iEEG) datasets, and infrequent integration of public and private datasets. Second, preprocessing practices remain rigid, often involving segmentation into windows of 10 s or less, fixed preictal horizons of 30 to 60 min, basic filtering and undersampling procedures. Third, limited model diversity, as most studies use convolutional neural networks (CNNs), with scant exploration of Transformers or graph neural networks (GNNs), dimensionality reduction is predominantly implicit and few linear baselines. Fourth, validation and deployment are fragmented, with fixed warning horizons, single evaluation protocols, missing clinical metrics, static thresholds or k-of-n voting, minimal interpretability, and few real-world implementations. Future work should assemble heterogeneous EEG datasets, adopt adaptive multiscale preprocessing, and build unified benchmarks that pit CNN, Transformer, GNN, and related models against well-tuned linear baselines—advancing flexible non-convolutional designs beyond the time-frequency paradigm, mandating ablations that contrast implicit with explicit, task-aware dimensionality reduction, and pairing these evaluations with clinically aligned validation, dynamic risk-aware thresholds, embedded interpretability, and prospective testing on energy-efficient edge hardware in ward and home settings.