Abstracts

Rationale: Recent studies have reported promising performances regarding the use of higher-order spectral analysis (HOSA) as a new precursor of ictal activity in intracranial electroencephalography (iEEG) recordings. Bispectrum is a HOSA measuring the level of non-linear phase and amplitude cross-frequency coupling between all pairs of frequency components of a signal. Its capacity to measure nonlinearities while retaining phase information allows the bispectrum to capture complex epileptic activity and renders it appropriate for the study of epileptic iEEG. While our team and others have demonstrated the potential of iEEG HOSA for detecting and predicting epileptic seizures, these studies were limited to short-term human and canine iEEG recordings and classification/analysis strategies were based on all available electrode channels. In this study, we statistically validated bispectrum features as biomarkers of epileptic activity and performed channel selection prior to classifying preictal and ictal iEEG recordings in the only existing long-term (total recording duration: 12.2 years) human iEEG database [1].

Methods: Bispectrum-derived features (magnitude (Mave), entropy (E1), squared entropy(E2)) were extracted from 16-contact iEEG recordings of 12 patients who participated in the NeuroVista Melbourne University Seizure Prediction Trial [1]. A total of 1195 seizures were analyzed with the 60 sec preceding each seizure (preictal period). A non-overlapping 30-sec window was used for feature extraction resulting in 29186 samples. The minimum redundancy maximum relevance (mRMR) algorithm was applied to the extracted features from the training sets (first 70% of seizures) of each patient to rank the 16 channels in order of importance. Patient-specific support vector machine (SVM) classifiers were then trained to classify preictal and ictal segments using the best 2,4,6,8, and 10 channels as well as the whole channel set.

Results: Student t-tests showed that on average, 12/16 channels present significant differences between preictal and ictal bispectrum feature values (Bonferroni-corrected p < 0.01). SVM classifiers were capable of distinguishing preictal and ictal segments without significant loss of accuracy after channel selection. With 2, 4, 6, 8, 10, and all 16 channels, average classification accuracies on held-out samples (30%) were 81.6%, 83.6%, 82.9%, 83.7%, 85.3%, and 85.9% respectively. Overall, for each of the features tested, the average loss of classification accuracy across all 12 patients were 0.34% (Mave), 0.33% (E1), and 0.07% (E2) per channel removed.