Abstracts

Rationale: Interictal epileptiform discharges (IEDs) can be associated with transient cognitive impairment (TCI) in persons with epilepsy. TCI due to IEDs is variable and may interfere with daily life, including driving. Prediction of impaired responses to visual stimuli may be especially relevant for fitness-to-drive evaluations. Previously, our lab trained classifiers to predict, from EEG, impaired reactions in absence epilepsy. We aim to apply this approach to IEDs resulting from a variety of epilepsies, in hopes of expanding the relevance of such a classifier to a more heterogeneous patient population reflective of those seen in outpatient epileptology clinics.

Methods: Routine scalp EEG recordings of 27 adolescent and adult patients, including patients with genetic generalized, non-lesional focal, and structural epilepsy, as well as epilepsy of unknown origin, were reviewed. IEDs occurring during a visual stimulus and/or response were examined. The visual stimuli were a flash with a button-press response, and an on-road obstacle to be avoided in a laptop driving game. IED onset and offset, and the categorization of the IED as spared or impaired (preserved reaction(s) or missed reaction to a stimulus) were recorded. Using a preliminary dataset of IEDs (466 IEDs, 30 impaired), quantitative features were extracted from during-IED time windows of the respective EEGs. The features were tested for significance (Wilcoxson rank sum, p < 0.001) between the spared and impaired IEDs. Feature set 1 included only duration, raw spike power, and raw wave power; feature set 2 included all significant features (Wilcoxson rank sum, p < 0.001). Feature sets 1 and 2 were used together with spared and impaired labels to train support vector machine (SVM) and linear discriminant analysis (LDA) classifiers. 10-fold cross-validation was used to assess classifier performance for each feature set; for the test set, spared predictive value (SPV) was calculated as the number of IEDs correctly classified as spared, divided by the total number of IEDs classified as spared (correctly and incorrectly). Mean SPV was calculated for 5 trials.

Results: The following features differed (p < 0.001, Wilcoxon rank sum) between spared and impaired IEDs: raw spike power, raw wave power, duration, Hjorth activity, Hjorth complexity, mean RMS, Delta power, Theta power, Alpha power, Beta power, low Gamma power, and temporal variance. The best performance, in terms of mean SPV, was seen in the LDA classifier using feature set 2, with a mean 95.8% ± 0.4% SPV achieved.

Conclusions: LDA predicted spared and impaired reactions to visual stimuli associated with IEDs from a variety of epilepsies. The high mean SPV achieved suggests similar machine learning approaches could, in the future, aid prediction of IED-associated TCI.

Funding: Please list any funding that was received in support of this abstract.: YA supported by the Yale College Dean’s Office Hahn Fellowship and Dufault Fellowship; HK by the European Union's Framework Program for Research and Innovation Horizon 2020, Marie Sklodowska‐Curie Grant Agreement No. 99791.