Supervised domain generalization for integration of disparate scalp EEG datasets for automatic epileptic seizure detection

Authors

K Ayodele
W Ikezogwo
M Komolafe
Philip O. Ogunbona, University of WollongongFollow

RIS ID

142833

Publication Details

Ayodele, K., Ikezogwo, W., Komolafe, M. & Ogunbona, P. (2020). Supervised domain generalization for integration of disparate scalp EEG datasets for automatic epileptic seizure detection. Computers in Biology and Medicine, 120

Abstract

© 2020 Elsevier Ltd Numerous automatic epileptic seizure detectors (ESDs) with excellent performances have been reported, but they generally experience performance degradation when tested with real-life clinical data. This has been blamed on the scarcity of high-quality training data, which leads to models that generalize poorly. There is consequently interest in methods to improve the quality and quantity of training data for ESDs. This study used a domain generalization approach to combine data from two different datasets for training an ESD, which was thereafter tested on a third dataset. A subspace of the CHB-MIT and TUSZ scalp EEG seizure datasets was extracted using transfer component analysis, based on a reproducing kernel Hilbert space approach. We then used the Azimuthal Equidistant Projection to transform 3D electrode coordinates into 2D space, followed by interpolation using the Clough–Tocher technique to generate 16x16 rasters. We thereafter generated feature vectors, each of which was a sequence of 17 ten-layer 16x16 raster arrays. The vectors were used to train a recurrent-convolutional neural network. The network had a 128-unit long short-term memory layer with inputs from 17 parallel networks each with three stacks of convolutional layers. Testing was based on a private 26-subject dataset, combined with randomly selected subsets of the CHB-MIT and TUSZ datasets. A combined sensitivity of 74.5% was achieved, along with a false positive per hour rate of 0.84, and a latency of 2.32 s. Detection sensitivity on the private dataset was 72.5%. These results compare favorably with results of large-scale validation studies in literature and confirm the viability of this approach to increasing the size of training datasets for ESDs.