Neural speech decoding from segregated brain regions' neuromagnetic signature

Tsang, Brian Yinbok
Journal Title
Journal ISSN
Volume Title

Locked-in syndrome (LIS) is a condition in which patients are in full-body paralysis but retain cognitive function. Improving the ability of these patients to communicate with others may significantly increase their quality of life. Since motor control is almost completely gone for LIS patients, a brain-computer interface (BCI) using neural speech decoding may be the only method to restore communication ability. As a non-invasive neuroimaging modality, magnetoencephalography (MEG) offers several strong attributes for potential speech-BCI usage. Its high spatial and temporal resolutions make it suitable for potential real-time speech decoding, and its non-invasive quality is crucial for accessibility. Neural speech decoding using MEG is a relatively new field, and there is little research on how different regions of the brain contribute to speech decoding. Additionally, the contributions of different brainwaves from individual brain regions towards MEG speech decoding hasn’t been studied. Such information may be helpful for future MEG speech-BCI studies regarding feature selection and sensor placement. This dissertation investigates the degree in which MEG signals from the eight individual brain regions corresponding to the lobes of the brain (left- and right- frontal, parietal, temporal, and occipital lobes) contribute towards neural speech decoding, as well as how individual brainwaves from these areas influence decoding performance. Both imagined and overt speech phrases were decoded using MEG signals collected from seven healthy participants. Results indicated that almost all regions provided neuromagnetic signals helpful for speech decoding. Additionally, both temporal regions offered relatively high individual decoding performances, and when combined, speech decoding performance was almost on-par with whole-brain performance. Lastly, frequency-specific analysis showed signs that helpful MEG signals for speech decoding may be encoded similarly across brain regions, with the delta band (0.3-4Hz) consistently providing the best information for speech decoding. Possessing flexibility in sensor count and placements, optically pumped magnetometer (OPM) MEG systems may find the results of this dissertation useful in determining sensor placements and reducing overall sensor count, which would lower the cost of the system.