Machine-learning-based diagnostics of EEG pathology.

Area of Science:

• Computational neuroscience
• Medical informatics
• Artificial intelligence in healthcare

Background:

• Clinical electroencephalogram (EEG) analysis is crucial for diagnosing neurological conditions.
• Manual EEG interpretation is time-consuming and requires specialized expertise.
• Machine learning (ML) offers potential for automating and standardizing EEG analysis.

Purpose of the Study:

• To develop a comprehensive feature-based ML framework for EEG pathology decoding.
• To conduct an in-depth comparison between feature-based and end-to-end ML approaches for EEG analysis.
• To evaluate the performance of these methods on a large, publicly available EEG dataset.

Main Methods:

• Development of an elaborate feature-based framework for EEG analysis.
• Application of the feature-based framework and deep neural networks (including a temporal convolutional network - TCN) to pathological EEG classification.
• Utilized the Temple University Hospital (TUH) Abnormal EEG Corpus (v2.0.0) for robust comparison.

Main Results:

• The proposed feature-based framework achieved classification accuracies comparable to state-of-the-art deep neural networks.
• Accuracies for both feature-based and end-to-end methods ranged narrowly between 81% and 86%.
• Analysis revealed that both approaches leveraged similar data features, such as delta and theta band power in temporal regions.

Conclusions:

• Feature-based ML decoding can achieve performance on par with advanced deep learning models for EEG pathology classification.
• Current EEG pathology decoders show potential clinical utility, especially in resource-limited settings, despite performance ceilings possibly due to label subjectivity.
• The open-source release of the feature-based framework provides a new resource for the EEG machine learning research community.