Machine learning approaches for prediction of epilepsy risk across clinical pathways: a systematic review.
Authors
Affiliations (4)
Affiliations (4)
- Department of Medicine, Division of Neurology, Queen's University, 76 Stuart St, Kingston, Ontario, K7L 2V7, Canada.
- Acquired Brain Injury (Epilepsy Program), Nova Scotia Health Authority, 6287 Alumni Crescent, Halifax, Nova Scotia, B3L 2C2, Canada.
- Department of Psychology, Queen's University, 62 Arch Street, Kingston, Ontario, K7L 3N6, Canada.
- Department of Medicine, Division of Neurology, Queen's University, 76 Stuart Street, Kingston, Kingston, Ontario, K7L 2V7, Canada.
Abstract
Machine learning (ML) and deep learning (DL) models are increasingly being
explored for individualized epilepsy risk prediction after a first unprovoked seizure (UFS) and
after acute brain insults such as stroke or traumatic brain injury. We systematically evaluated
their predictive performance, input modalities, validation strategies, methodological quality,
and translational readiness across these two clinical pathways.
Approach. PubMed, Scopus, IEEE Xplore, and Web of Science were searched for Englishlanguage human studies published between January 2005 and October 2025. Eligible studies
used ML or DL to predict seizure recurrence after UFS or epilepsy development after
acute brain insult using clinical, neuroimaging, electrophysiological, electronic-health-record,
or multimodal data. Two reviewers performed blinded duplicate screening, followed by
duplicate data extraction using a CHARMS-aligned form. Risk of bias and applicability were
independently assessed using PROBAST+AI across the Participants, Predictors, Outcome,
and Analysis domains.
Main results. Thirteen studies met the eligibility criteria: six addressed UFS and seven
addressed post-insult epilepsy. Reported AUCs for the best-performing models ranged from
0.60 to 0.93, with the highest discrimination observed in models using high-dimensional
neuroimaging, unstructured clinical text, or multimodal data. These inputs included MRI
morphometric asymmetry, clinical free text, EEG, diffusion MRI, resting-state fMRI, and
multimodal fusion. In the three studies that directly compared modality combinations,
multimodal models improved AUC by approximately 0.04-0.10 over the best single-modality
counterpart. Model credibility was strongest when independent validation, transparent
feature handling, and calibration assessment were reported.
Significance. ML/DL approaches show clear potential for earlier, individualized epilepsy
risk stratification, particularly when complementary clinical, electrophysiological, and
neuroimaging data are integrated. Future studies should prioritize prospective multi-site
validation, standardized EEG/MRI data structures, transparent multi-metric reporting, and
reproducible model documentation aligned with TRIPOD+AI and PROBAST+AI.