Relating EEG to continuous speech using deep neural networks: a review.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalography (EEG) records brain responses to continuous speech.
• Linear models are currently used to measure neural tracking of speech but have limitations due to assumed linearity.
• Deep learning models offer a potential alternative for analyzing nonlinear EEG-speech relationships.

Purpose of the Study:

• To review and analyze deep learning studies relating EEG to continuous speech.
• To identify common methodological pitfalls in existing research.
• To propose requirements for a standard benchmark in EEG-speech decoding analysis.

Main Methods:

• Systematic review of 29 deep learning-based studies on EEG-speech decoding.
• Critical analysis of methodologies, including cross-validation, data leakage, and model complexity.
• Identification of best practices for model evaluation and benchmarking.

Main Results:

• Identified recurrent methodological issues such as biased cross-validations and data leakage.
• Found that many models are over-fitted due to disproportionate data size relative to complexity.
• Highlighted the need for public datasets, common evaluation metrics, and standardized match-mismatch tasks.

Conclusions:

• Deep learning shows potential for EEG-speech decoding, but rigorous methodology is crucial.
• Standardized benchmarks and transparent reporting are necessary for advancing the field.
• Addressing identified pitfalls will improve the reliability and generalizability of EEG-speech decoding models.