Noninvasive decoding of typed sentences from human brain activity.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Modern neuroprostheses offer communication restoration but require invasive surgery with associated risks.
• There is a need for noninvasive brain-computer interfaces (BCIs) to mitigate surgical risks.

Purpose of the Study:

• To introduce and evaluate Brain2Qwerty, a novel noninvasive deep learning architecture for decoding sentence production from brain activity.
• To compare the efficacy of magnetoencephalography (MEG) and electroencephalography (EEG) in decoding sentences using the Brain2Qwerty system.

Main Methods:

• Developed Brain2Qwerty, a deep learning model trained on electroencephalography (EEG) and magnetoencephalography (MEG) data.
• Participants (n=35) typed memorized sentences on a QWERTY keyboard while their brain activity was recorded.
• Assessed model performance using character error rate (CER).

Main Results:

• Brain2Qwerty achieved an average CER of 29% with MEG, significantly outperforming EEG (65% CER).
• Top-performing participants achieved a CER of 18%, with the model decoding novel sentences accurately.
• The noninvasive approach demonstrated substantial decoding capabilities.

Conclusions:

• The Brain2Qwerty system represents a significant advancement in noninvasive BCIs for communication restoration.
• These findings bridge the gap between invasive and noninvasive neuroprosthetic methods.
• This research paves the way for developing safer BCIs for patients with communication impairments.