Deep learning for inner speech recognition: a pilot comparative study of EEGNet and a spectro-temporal Transformer on...

Area of Science:

• Cognitive Neuroscience
• Machine Learning
• Brain-Computer Interfaces (BCIs)

Background:

• Inner speech, the mental articulation of words, is a key cognitive process with emerging applications in BCIs.
• Non-invasive electroencephalography (EEG) offers a promising avenue for decoding inner speech.
• Evaluating deep learning models is crucial for advancing EEG-based inner speech classification.

Purpose of the Study:

• To compare the performance of deep learning models, specifically EEGNet and a spectro-temporal Transformer, for inner speech classification.
• To assess the generalizability of these models using non-invasive EEG data from a bimodal EEG-fMRI dataset.
• To identify effective model architectures and features for decoding inner speech.

Main Methods:

• Inner speech tasks involving eight target words were performed by four participants.
• EEG signals were preprocessed and segmented for classification using EEGNet and a spectro-temporal Transformer.
• Leave-one-subject-out (LOSO) cross-validation was employed, with performance evaluated using accuracy, macro-F1 score, precision, and recall.

Main Results:

• The spectro-temporal Transformer achieved the highest accuracy (82.4%) and macro-F1 score (0.70), surpassing EEGNet models.
• Wavelet-based time-frequency features and attention mechanisms significantly enhanced classification performance.
• Social word categories showed better discriminative power than number concepts, indicating differential mental processing.

Conclusions:

• Attention-based Transformer models show significant potential for decoding inner speech from EEG signals.
• These findings support the development of non-invasive, real-time BCIs for communication rehabilitation in individuals with severe disabilities.
• Future research should focus on expanding vocabulary, increasing participant diversity, and validating models in clinical settings.