Seizure Prediction: A Visual Approach

Abstract

Activity preceding the onset of epileptic seizures has been an elusive subject for neuroscience research, without a clear grasp of what patterns might be responsible. In this work, we present an out of the box approach to this problem, trying to mimic the visual inspection process that a trained physician might do to locate the beginning of a pre-ictal state in an EEG plot. We explore different data labeling methods for the posterior training of a Convolutional Neural Network, taking into account only visual characteristics for classification. Ten second images (300x400 px) were synthesized from scalp EEG recordings belonging to 10 epileptic patients from the public Physionet CHB-MIT database. A tortuosity measure was taken for each one-second window, for each channel (23 channels in 10–20 bipolar configuration). Unsupervised clustering methods in conjunction with the mean and the standard deviation of the tortuosity sets were used to identify pre-ictal states; interictal states were selected according to the same proximity criteria used for the Kaggles Melbourne University AES/MathWorks/NIH Seizure Prediction Challenge. The proposed labelling method indentified 28 posible pre-ictal states across 10 patients. Data from pre-ictal states and interictal states was used to train, and test, a Convolutional Neural Network classifier for each of the 8 patients selected. A classification accuracy of 99.29% was achieved for the best patient; however, an accuracy of 46.93% was also obtained for the worst patient. Mean performance across patients was 76.03%, a 52.07% improvement over chance.