Seizure forecasting with multiple timescales and features.
Saved in:
| Title: | Seizure forecasting with multiple timescales and features. |
|---|---|
| Authors: | Liu, Yueyang (AUTHOR), Soto‐Breceda, Artemio (AUTHOR), Karoly, Philippa (AUTHOR), Grayden, David B. (AUTHOR), Cook, Mark J. (AUTHOR), Freestone, Dean R. (AUTHOR), Schmidt, Daniel (AUTHOR), Kuhlmann, Levin (AUTHOR) |
| Source: | Epilepsia (Series 4). Apr2026, Vol. 67 Issue 4, p1778-1791. 14p. |
| Subjects: | Electroencephalography, Machine learning, Electrophysiology, Time series analysis, Ensemble learning, Epilepsy |
| Abstract: | Objective: Forecasting epileptic seizures is a difficult task. Studies of seizure prediction have investigated many different EEG features, but none of them have been useful enough to be applied in clinical practice beyond trials. Moreover, most of these features have been applied to short‐term intracranial EEG (iEEG) recordings, limiting the possibility of reliable statistical evaluation. This paper proposes a machine learning algorithm to forecast an epileptic seizure 2‐4 mins before seizure. This allows patients to seek help, or stimulation devices to work. Methods: This paper investigates a large subset of features from the past and present to unravel which features and feature analysis methods will yield the best performance on long‐term iEEG recordings (from 14 patients with focal epilepsy) and thus provide the most reliable step toward clinical utility. Specifically, this study implements a multiple long‐time scale cycle feature analysis framework for seizure forecasting that considers the state‐of‐the‐art time series features of critical slowing down (autocorrelation and variance) as well as interictal epileptiform discharge (IED) / spike rate, High Frequency Activity (HFA), seven different univariate features, and three Neural Mass Model (NMM) features based on brain dynamics. Results: Seizure phase histograms of all the features are then analyzed to investigate each feature's potential for seizure forecasting by evaluating corresponding synchronization indices (SI) on fast (40 minutes to 2 days) and slow (2+ days) wideband time scales. Out of all combinations considered, the overall performance comparison across patients highlights that 'autocorrelation + variance + NMM + spike rate' features achieve the highest average AUC of 0.83, showcasing its performance in forecasting seizures. Significance: A model is proposed that has a similar performance compared to the state‐of‐the‐art method, without the need of selecting the best channel prior to model building. Light is also shed on the comparative performance on long‐term recordings of many of the seizure forecasting features considered in the past. [ABSTRACT FROM AUTHOR] |
| Copyright of Epilepsia (Series 4) is the property of Wiley-Blackwell and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Psychology and Behavioral Sciences Collection |
Be the first to leave a comment!
