Machine-Learning-Based Diagnostics of EEG Pathology

Machine learning (ML) methods have the potential to automate clinical EEGanalysis. They can be categorized into feature-based (with handcraftedfeatures), and end-to-end approaches (with learned features). Previous studieson EEG pathology decoding have typically analyzed a limited number of features,decoders, or both. For a I) more elaborate feature-based EEG analysis, and II)in-depth comparisons of both approaches, here we first develop a comprehensivefeature-based framework, and then compare this framework to state-of-the-artend-to-end methods. To this aim, we apply the proposed feature-based frameworkand deep neural networks including an EEG-optimized temporal convolutionalnetwork (TCN) to the task of pathological versus non-pathological EEGclassification. For a robust comparison, we chose the Temple UniversityHospital (TUH) Abnormal EEG Corpus (v2.0.0), which contains approximately 3000EEG recordings. The results demonstrate that the proposed feature-baseddecoding framework can achieve accuracies on the same level as state-of-the-artdeep neural networks. We find accuracies across both approaches in anastonishingly narrow range from 81--86\%. Moreover, visualizations and analysesindicated that both approaches used similar aspects of the data, e.g., deltaand theta band power at temporal electrode locations. We argue that theaccuracies of current binary EEG pathology decoders could saturate near 90\%due to the imperfect inter-rater agreement of the clinical labels, and thatsuch decoders are already clinically useful, such as in areas where clinicalEEG experts are rare. We make the proposed feature-based framework availableopen source and thus offer a new tool for EEG machine learning research.