Self-supervised deep learning and EEG categorization

Deep learning has the potential to be used to improve and streamline EEG analysis. At the present, classifiers and supervised learning dominate the field. Supervised learning depends on target labels which most often are created by human experts manually classifying data. A problem with supervised learning is intra- and interrater agreement which in some instances are far from perfect. This can affect the training and make evaluation more difficult. This thesis includes three papers where self-supervised deep neural networks were developed. In self-supervised learning, the input data to the networks themselves contain structures that are used as targets for the training and no labeling is necessary. In paper I, deep neural networks were trained to increase the number of-, or to recreate missing EEG-channels. The performance was at least on the same level as that of spherical interpolation, but unlike in the case of interpolation, missing data does not have to be identified manually first. Papers II and III involved developing deep neural networks for clustering analysis. The networks produced two-dimensional representations of EEG data and the training strategy was based on the principle of t-distributed stochastic neighbor embedding (t-SNE). In paper II, comparisons were made to parametric t-SNE and EEG-features obtained from time-frequency methods. The deep neural networks produced more distinct clustering when tested on data annotated for epileptiform discharges, seizure activity, or sleep-wakefulness. In paper III, the newly developed method was used to compare annotations of epileptiform discharges. Two experts performed independent annotations and classifiers were trained on these, using supervised learning, which in turn produced new annotations. The agreement when comparing two sets of annotations was not larger between the two experts than between an expert and a classifier. The analysis showed that differences in the annotations by the experts infl
Funding: The work was funded by Saad Nagi (grants RÖ-974228, RÖ-962769, and RÖ-941377), Magnus Thordstein (grant RÖ-986017), and Håkan Olaus-son (grants LIO-936017 and RÖ-941359).