Abstracts

Rationale: Identifying interictal epileptiform transients (ET) on scalp electroencephalography (EEG) is time-consuming and expert-dependent. A fast and reliable automated or semi-automated detection system would be a useful aid for clinicians. The classification of an EEG as diagnostically ‘positive’ is more clinically relevant than detecting the number of ETs per se. In this study, we develop an automated identification system based on a database of annotated interictal patterns appearing on clinical EEGs. Methods: We analyzed 256 routine EEGs recorded using the 10-20 International electrode system. 156 EEGs (93 diagnostic and 63 ET-free EEG) were extracted at Massachusetts General Hospital (MGH), Boston and 100 EEG (50 epileptic and 50 ET-free EEG) were recorded at National University Hospital (NUH), Singapore. The EEGs were notch filtered (50 Hz for NUH data and 60 Hz for MGH data), high pass filtered (1 Hz Butterworth, order = 4), and the Common Average Referential (CAR) montage was applied. The EEG data was down-sampled to 128 Hz after applying an anti-aliasing filter. A power-based artifact rejection mechanism removed ‘noisy’ EEG segments. The ETs in the MGH datasets (specifically spikes and spike-wave complexes) were cross-annotated by two neurologists. We developed a two-step system to classify EEGs: detection of ETs, then classifying the EEG as “ET free” or “not ET-free”. A Convolutional Neural Network (CNN) was designed for ET detection. The ETs and non-ETs (background) are extracted as 500-millisecond waveforms (64 samples), and these waveforms were applied as the training inputs for the CNN. The CNN takes a 500-millisecond segment of EEG and provides an output between 0 and 1, a higher value indicating the presence of an ET. The different parameters of the CNN were optimized by nested cross-validation. We obtain 19 CNN outputs (for 19 channels) for a particular instant of time. The 19 CNN outputs were combined to provide a single decision for that epoch. We employ the average of the maximum two values of CNN outputs as the combining rule. This approach improves the ET detection accuracy by incorporating spatial information and removing certain channel localized artifacts such as electrode popping. Next EEG classification was implemented based on the features extracted from ET detection. We employed a Support Vector Machine (SVM) with Gaussian kernel for EEG classification. We extracted features from the different time-instant CNN outputs of an EEG. Each feature was defined as the fraction of time-instant CNN outputs that belongs to a particular threshold range. We divide the CNN output range [0,1] into 10 equal ranges: [0,0.1), [0.1,0.2), …, [0.9,1]. The best features were selected based on p-values. The SVM was trained with the best features and the different parameters of SVM (soft margin cost function parameter and gamma) were optimized by applying Bayesian optimization. We divided the MGH data into two sets, one for training the ET detector and one for training the SVM EEG classifier. The performance of the ET detector was evaluated on the same MGH set applied for training the SVM. The SVM EEG classifier was tested on the ‘blinded’ NUH dataset. Results: The EEG classifier was trained based on the best features extracted from the ET detector on the MGH data. We achieved an accuracy of 88% for classifying the 100 NUH EEGs as containing ETs or ET-free. Conclusions: This study demonstrates the reliability of an EEG identification system across data from different hospitals. We achieved a mean accuracy of 88% for classifying EEGs as diagnostic or spike-free. This system would aid non-EEG neurologists in diagnosing epilepsy more efficiently, thus improving patient care. Funding: NHIC-I2D-1608138