Deep learning-based electroencephalography analysis: a systematic review

Abstract: Electroencephalography (EEG) is a complex signal and can require several years of training to be correctly interpreted. Recently, deep learning (DL) has shown great promise in helping make sense of EEG signals due to its capacity to learn good feature representations from raw data. Whether DL truly presents advantages as compared to more traditional EEG processing approaches, however, remains an open question. In this work, we review 156 papers that apply DL to EEG, published between January 2010 and July 2018, and spanning different application domains such as epilepsy, sleep, brain-computer interfacing, and cognitive and affective monitoring. We extract trends and highlight interesting approaches in order to inform future research and formulate recommendations. Various data items were extracted for each study pertaining to 1) the data, 2) the preprocessing methodology, 3) the DL design choices, 4) the results, and 5) the reproducibility of the experiments. Our analysis reveals that the amount of EEG data used across studies varies from less than ten minutes to thousands of hours. As for the model, 40% of the studies used convolutional neural networks (CNNs), while 14% used recurrent neural networks (RNNs), most often with a total of 3 to 10 layers. Moreover, almost one-half of the studies trained their models on raw or preprocessed EEG time series. Finally, the median gain in accuracy of DL approaches over traditional baselines was 5.4% across all relevant studies. More importantly, however, we noticed studies often suffer from poor reproducibility: a majority of papers would be hard or impossible to reproduce given the unavailability of their data and code. To help the field progress, we provide a list of recommendations for future studies and we make our summary table of DL and EEG papers available and invite the community to contribute.

• The paper systematically reviews 156 EEG studies, revealing a median 5.4% accuracy improvement of deep learning models over traditional methods.
• It details diverse preprocessing techniques and deep learning architectures such as CNNs and RNNs, highlighting the trend toward inter-subject classification.
• It emphasizes reproducibility issues with limited data and code sharing, and recommends standardized reporting practices to enhance research quality.

Overview of "Deep learning-based electroencephalography analysis: a systematic review"

The paper "Deep learning-based electroencephalography analysis: a systematic review" provides a comprehensive review of the application of deep learning (DL) to electroencephalography (EEG) data analysis, encompassing publications from January 2010 to July 2018. The analysis encompasses 156 papers, presenting trends in DL methodologies, data characteristics, application domains, and reproducibility concerns.

Data Characteristics and Collection

The survey spans a wide range of data characteristics:

• Quantity: EEG data used in the reviewed studies varies widely, from datasets containing a few minutes of recording to thousands of hours. The number of examples seen by a network ranges from tens to millions.
• Subjects: The number of subjects per study varies significantly, with a median of 13. Notably, larger datasets with thousands of subjects also exist, demonstrating that scaling issues are being addressed.
• Recording Parameters: The reviewed studies employ various EEG recording devices, with some using consumer-grade devices (e.g., Emotiv EPOC, Muse), while others use research-grade systems (e.g., BioSemi, BrainVision). Sampling rates typically range between 100 and 1000 Hz.
• Data Augmentation: Techniques like overlapping windows and adding noise are commonly used to augment the limited EEG data, albeit with varying levels of impact on performance.

EEG Processing Methods

The review identifies several preprocessing steps and artifact handling techniques:

• Preprocessing: Common steps include downsampling, band-pass filtering, and windowing. While some studies apply minimal preprocessing aiming to let DL models learn from raw data, others use extensive preprocessing to enhance signal quality.
• Artifact Handling: Around 46% of the studies do not handle artifacts explicitly, suggesting that DL models can tolerate or learn to ignore certain types of noise inherent in EEG data.

Deep Learning Methodologies

The review categorizes DL methodologies into several components:

• Architectures: Convolutional Neural Networks (CNNs) are most frequently used, followed by Recurrent Neural Networks (RNNs) and Autoencoders (AEs). Notably, deeper models typically found in computer vision tasks (e.g., VGG, ResNet) are less common in EEG analysis due to the smaller dataset sizes.
• Training Procedures: Around 50% of studies employ standard optimization techniques without pre-training, while approximately 25% use pre-training methods such as unsupervised feature learning.
• Regularization and Optimization: Regularization techniques like dropout and weight decay are often used to improve generalization. Adam and Stochastic Gradient Descent (SGD) are the most common optimizers. There is a notable lack of detailed hyperparameter search methodologies in most studies.
• Subject Handling: There is a growing trend towards inter-subject classification, moving beyond the traditional intra-subject models. This approach aims to leverage larger and more varied datasets, improving the generalization capabilities of DL models.

Performance and Model Comparison

The paper examines:

• Baseline Comparisons: About 68% of studies compare their DL models to traditional EEG processing pipelines, while 34% compare to other DL approaches. The median accuracy improvement of DL models over traditional baselines is reported as 5.4%.
• Reporting Practices: There is a high degree of variability in how performance metrics and validation procedures are reported, making direct comparisons challenging. Standardizing these reporting practices would enhance the field's progress.

Reproducibility

A significant concern is the reproducibility of results:

• Data Sharing: More than half of the studies use private datasets, limiting reproducibility. Sharing internal recordings would substantially benefit the research community.
• Code Sharing: Only 19% of studies make their source code available. Providing code and detailed descriptions of models and training procedures is crucial for reproducibility.

Recommendations

The paper concludes with recommendations to enhance the reproducibility and quality of DL-EEG research:

1. Clearly describe the architecture and provide diagrams or tables.
2. Use clear and unambiguous terminology when describing data.
3. Leverage existing public datasets to benchmark new models.
4. Include state-of-the-art baselines to contextualize improvements.
5. Share data and code to enable result replication and further research.

Conclusion and Future Directions

The review highlights the promise of DL in improving EEG analysis, particularly through end-to-end learning and better generalization across subjects. Future work should focus on expanding public datasets, standardizing evaluation metrics, and adopting best practices for data and code sharing to propel the field forward.

Supplementary Material: The authors also provide supplementary material online, including data items and code to encourage ongoing contributions and updates to the review. This initiative aims to maintain the review's relevance and accuracy as new research emerges.