Abstracts

Rationale: In individuals with drug-resistant epilepsy (DRE), seizures are unpredictable, often occurring with little or no warning. The unpredictability causes anxiety and much of the morbidity and mortality of seizures.  Recently, machine learning (ML) models have been developed to predict seizures with high accuracy using high-dimensional electroencephalography (EEG) recordings. One reason for this success is the ability to identify a pre-seizure period, which is not detectable by the domain experts but can be advanced ML models such as deep learning (DL) algorithms. As a result, seizures can be predicted by detecting the transition from the non-seizure regime to the pre-seizure regime.

Methods: In this work, we formulate the seizure prediction problem as an instance of a change point detection (CPD) problem and consider several variants under different settings, including the quickest CPD, transient CPD, and retrospective CPD frameworks. The EEG data are discretized to 1 second windows, but instead of treating the data as a bag of windows, we remain faithful to the temporal order of the windows. Our approach to CPD is based on estimating the probability density ratios using ML algorithms without explicitly computing the ratios. We can analytically establish that density ratios are sufficient statistics for performing CPD.

Results: We show the feasibility in simulations and apply the methods to Intracranial EEG recordings of 102 seizures of mesial temporal lobe onset were analyzed from 19 patients with DRE. We compare different frameworks and the associated penalty functions on the experimental data.

Conclusions: The CPD-based approach to seizure prediction captures the temporal information in the EEG signal; it is sensitive to temporal fluctuations and can detect the changes with minimal delay after the change point. This is a promising approach for real-time deployment.

Funding: None