# Deep learning-based electroencephalography analysis: a systematic review

**Authors:** Yannick Roy, Hubert Banville, Isabela Albuquerque, Alexandre Gramfort, Tiago H. Falk, Jocelyn Faubert  
**arXiv:** [1901.05498](https://arxiv.org/abs/1901.05498)  
**Date:** 2019-01-16  
**Summary page:** https://www.summarizepaper.com/en/paper/1901.05498v2/

## Key points

- Deep learning (DL) is a powerful method for analyzing brain activity and interpreting complex EEG signals.
- A systematic review examined 156 studies from January 2010 to July 2018 that applied DL to EEG data in areas such as epilepsy and sleep disorders.
- Key findings include variations in the amount of EEG data used, with CNNs being the most commonly used DL architecture (40%) followed by RNNs at 14%.
- Concerns about reproducibility were highlighted due to data and code unavailability in many papers.
- DL approaches showed a median accuracy improvement of 5.4% over traditional baselines in relevant studies.

## Summary

, , , , 


In the field of electroencephalography (EEG) analysis, deep learning (DL) has emerged as a promising tool for interpreting complex EEG signals. A recent systematic review conducted by Yannick Roy, Hubert Banville, Isabela Albuquerque, Alexandre Gramfort, Tiago H. Falk, and Jocelyn Faubert delves into the application of DL to EEG data in various domains such as epilepsy, sleep disorders, brain-computer interfacing, and cognitive/affective monitoring. The review encompasses 156 papers published between January 2010 and July 2018.

 is a powerful method for analyzing brain activity and  has emerged as a promising tool for interpreting its complex signals. In a recent <kd> systematic review</ kd>, researchers examined 156 studies from January 2010 to July 2018 that applied DL to EEG data in areas such as epilepsy and sleep disorders. The study highlights the potential of DL in extracting meaningful features from raw EEG data but also raises questions about its comparative advantages over traditional EEG processing methods. Through an extensive analysis of the reviewed papers, several trends and interesting approaches were identified to guide future research endeavors. Key findings from the review include variations in the amount of EEG data utilized across studies, ranging from less than ten minutes to thousands of hours. Convolutional neural networks (CNNs) were found to be the most commonly used DL architecture (40%), followed by recurrent neural networks (RNNs) at 14%, typically with 3 to 10 layers. Additionally, nearly half of the studies trained their models on either raw or preprocessed EEG time series. is a significant concern in many papers due to unavailability of data and code, hindering replication efforts within the scientific community. One notable outcome revealed by the analysis was a median accuracy improvement of 5.4% achieved by DL approaches over traditional baselines in relevant studies. To address these challenges and propel further advancements in DL-based EEG analysis, the authors provide a set of recommendations for future research initiatives. They also make their comprehensive summary table of DL and EEG papers accessible for contributions from researchers in the field. Overall, this systematic review sheds light on the current landscape of DL applications in EEG analysis while emphasizing the importance of  and  within the research community to drive innovation and progress in this evolving domain.