Recognition of EEG Signals from Imagined Vowels Using Deep Learning Methods

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCI) leverage electroencephalographic (EEG) signals for communication, particularly with imagined speech.
• Analyzing and classifying complex EEG signals related to imagined speech remains a significant research challenge.
• Existing methods require improvement for accurate recognition of nuanced imagined speech patterns.

Purpose of the Study:

• Develop a novel Deep Learning (DL) algorithm, CNNeeg1-1, for recognizing EEG signals during imagined vowel tasks.
• Create a specialized imagined speech database (BD2) comprising 50 Spanish native speakers imagining vowels (/a/,/e/,/i/,/o/,/u/).
• Evaluate and compare the CNNeeg1-1 algorithm's performance against benchmark DL algorithms (Shallow CNN, EEGNet) using both open-access (BD1) and the new (BD2) databases.

Main Methods:

• Implementation of a new Deep Learning algorithm, CNNeeg1-1, designed for EEG signal classification.
• Establishment of a new imagined vowel speech database (BD2) with 50 participants.
• Comparative performance analysis using mixed variance analysis of variance for intra-subject and inter-subject training, contrasting CNNeeg1-1 with Shallow CNN and EEGNet on BD1 and BD2.

Main Results:

• The CNNeeg1-1 algorithm demonstrated superior performance in classifying imagined vowels compared to Shallow CNN and EEGNet.
• CNNeeg1-1 achieved an accuracy of 65.62% on the open-access BD1 database.
• CNNeeg1-1 reached an accuracy of 85.66% on the newly developed BD2 database for intra-subject training.

Conclusions:

• The developed CNNeeg1-1 algorithm shows significant potential for enhancing brain-computer interfaces through improved imagined speech recognition.
• The specialized imagined vowel database (BD2) provides a valuable resource for future research in this domain.
• CNNeeg1-1 represents a promising advancement in classifying EEG signals for communication applications.