Hybrid Deep Learning (hDL)-Based Brain-Computer Interface (BCI) Systems: A Systematic Review.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Brain-Computer Interface (BCI) systems are increasingly reliable due to Artificial Intelligence (AI).
• Hybrid deep learning (hDL), combining multiple deep learning algorithms, has emerged as a significant advancement in BCI research over the last five years.

Purpose of the Study:

• To review 47 papers on hDL-based BCI systems published between 2015 and 2020.
• To extract trends and highlight key aspects of hDL applications in BCI.

Main Methods:

• Searched four major scientific databases: Google Scholar, PubMed, IEEE Xplore, and Elsevier Science Direct.
• Extracted and analyzed data items including databases, applications, training methods, tasks, pre-processing, normalization, features, DL architectures, layer counts, and optimization approaches.

Main Results:

• Electroencephalography (EEG) is the most utilized technique, with hDL showing potential to mitigate its inherent Signal-to-Noise Ratio (SNR) drawbacks, as pre-processing was used in only 21.28% of cases.
• Temporal features yielded high accuracy (93.94%), while spatial-temporal features were most common (33.33%).
• Convolutional Neural Network-Recurrent Neural Network (CNN-RNN) was the most frequent architecture (47%), with many studies using fewer layers for efficiency.

Conclusions:

• A summary table of hDL-based BCI papers is provided to aid the scientific community.
• Open challenges include exploring neuroimaging techniques beyond EEG (e.g., fNIRS), understanding pre-processing's role in accuracy, and investigating novel architectures like RNN and Deep Belief Networks (DBN).