Deep learning-based electroencephalography analysis: a systematic review.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) signal interpretation is complex, often requiring extensive training and advanced signal processing.
• Deep learning (DL) offers potential for automated feature extraction from raw EEG data.
• The comparative advantage of DL over traditional EEG analysis methods is an ongoing research question.

Purpose of the Study:

• To review and analyze the trends in the application of DL to EEG data.
• To identify promising DL approaches and application domains in EEG research.
• To provide recommendations for future EEG studies utilizing DL.

Main Methods:

• Systematic review of 154 DL-based EEG papers published between January 2010 and July 2018.
• Extraction and analysis of data on EEG datasets, preprocessing, DL model design, results, and reproducibility.
• Categorization of studies across application domains like epilepsy, sleep, and brain-computer interfacing.

Main Results:

• EEG data size varied significantly; over half of studies used publicly available data.
• A shift towards inter-subject analysis was observed.
• Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) were prevalent DL models.
• A median accuracy gain of [Formula: see text] was observed for DL over traditional methods.
• A significant lack of reproducibility was noted due to unavailable data and code.

Conclusions:

• DL methods demonstrate a notable performance improvement in EEG analysis compared to traditional approaches.
• Reproducibility is a critical issue hindering scientific progress in DL-based EEG research.
• Recommendations are provided to enhance the reproducibility and sharing of DL-EEG research, including the development of public benchmarking resources.