Abstracts

Rationale: The goal is to determine if Benford's Law can be used to extract features from EEG data for seizure detection and to compare these manual features with automatic features generated by a CVAE as well as the validated methods of phase-locked high gamma (PLHG) and Epileptogenicity Index (EI). CVAEs are used to generate a continuous latent space (resembling an embedding vector from NLP) that can be trained to emphasize specific spectral features of the data, and thus contain more clinically verifiable information than other compression algorithms. Classifier performance on latent space can guide feature selection by acting as an ideal feature set.

Methods: Two entropy metrics were designed to quantify deviations from Benford’s Law, one in the time domain (TDBE) and one in the frequency domain (FDBE), over one second epochs in intracranial EEG (iEEG) recordings from patients with medication refractory focal epilepsy undergoing a surgical work-up. The newly designed Benford Entropy (BE) metrics were studied individually and grouped into various feature ensembles. Classification between ictal and preictal epochs was performed using a k-nearest neighbor (kNN, k=10) classifier for each feature combination. A CVAE was also designed to automatically learn a characteristic embedding. As for BE, CVAE classification was performed using kNN over the model’s latent space.

The dataset consisted of 516 leads taken from 164 seizures across 35 individuals: 31 seizures from 12 patients were from UW; 94 seizures from 15 patients were from the Epilepsiae Database (Klatt et al 2012); 39 seizures from 8 individuals were provided by Mayo (Kini et al 2016). Each lead was selected by a board-certified neurologist and cropped to 2 mins, with seizure onset being one min into the segment. The preictal labels were assigned to every 1sec before seizure onset, and the ictal labels were assigned to every 1sec after seizure onset, creating an even amount of preictal and ictal data points. Each patient consented to the study, and it was approved by the respective IRBs.

Results: The TDBE and FDBE classifiers (by themselves) obtained similar AUC scores to the baseline method EI, but PLHG obtained a much higher score. AUC also increases when combining BE with other methods, with all four precomputed metrics together performing better than any subset. The CVAE latent space classifier obtained a higher AUC than the precomputed metric ensemble. The highest performing classifier used all four precomputed features with the latent space. All classifier scores are included in the Table 1 (P < 0.05).