Abstracts

Rationale: EEG-based automated seizure detection methods have been investigated for objective seizure quantification by extension, individually tailored treatments. Recently, the efforts have been made to practically be used in real-time EEG monitoring and closed-loop responsive therapeutic devices. For improving the efficiency and reliability of the seizure detection, the selection of discriminative EEG features of (pre-) ictal onset is imperative. Methods: This is a preliminary study conducted with the intracranial EEG (icEEG) data of two different seizure types; Type A with ambiguous seizure onset focus followed by rapid generalization and Type B with highly localized seizure onset focus followed by sequential ictal pattern evolution, respectively. EEG data of 7 subjects were converted into the frequency domain by using short-time Fourier Transform (STFT), to calculate the power spectrum density (PSD), which was partitioned into various frequency bands including delta (δ: 2-4 Hz), theta (θ: 4-8 Hz), alpha (α: 8-13 Hz), beta (β: 13-25 Hz), low gamma (γ: 25-55 Hz), high gamma (65-125 Hz), ripples (125-250 Hz), and fast ripples (250Hz-500Hz). STFT was computed with a Hamming window of 2 seconds (3200 samples) and shifted every 1/16 second. The PSD mapping data of each channel and each frequency band were labeled as “baseline” or “seizure” segment by support vector machine (SVM) classifier with median thresholds (Th1) obtained from 3 different training set of seizure period. Test EEG signals were preprocessed with STFT and then put into the trained classifier model to predict baseline or seizure epoch. If the classification results remain as seizure longer than 2 seconds, we considered the events as seizures. Results: Our classifier aims to detect seizure with the accuracy more than 95 % (mean 98.5 ± 1.4 %) in addition to short (± 3 sec, mean - 0.45 ± 0.83 sec) detection latency from the visual analysis among the electrodes over seizure onset zone. Among the pattern classifier algorithms of Type A, 61.2% were survived with the above cut-off values, and most of them adapted to be trained by a set of alpha to low gamma frequency bands. Which segment or how many times the classifier has trained the seizure has not had a significant impact on the performance. When we reviewed the result of Type B, only 1.4% classifiers were survived with the above cut-off values and the trained segment of the seizure determined the adopted frequency band. The number of times trained did not differ in performance between classifiers. Conclusions: Our results demonstrated that the efficiency of the automated EEG classifier based on the machine learning method is determined by selecting the EEG features of (pre-) ictal onset which was considered as the most reasonable for the purpose of the algorithm by an expert with neurophysiological knowledge. This computer-aided reproducible system for automatic quantification can serve as valuable clinical tools because the EEG signals are nonlinear and non-stationary, and the subtle and invisible statistical characteristics of ictal EEG precede visually analyzable rhythm changes. Funding: This research was supported by the Basic Science Research Program through the National Research Foundation of Korea funded by the Ministry of Science and ICT (2017R1A2A2A05069647, 2019M3C1B8090803 and 2019M3C1B8090802 to H.W.L.).