Optimal sensor set for MEG-based spontaneous and intended speech decoding toward practical communication

Neural speech decoding has demonstrated great potential as an alternative to current brain-computer interface (BCI) spellers, enabling a faster, more natural way of communication for patients who have lost voluntary abilities, including speech. Most neural speech decoding studies have focused on decoding cued speech, in which participants are presented with visual or auditory signals before producing imagined or overt speech. Only a few recent studies have investigated the decoding of spontaneous and intended speech without external cues using neural activity acquired from a full-scalp clinical magnetoencephalography (MEG) system. With recent advances in MEG technology, including optically pumped magnetometers (OPMs) and nitrogen-vacancy (NV) center (NV-diamond) magnetometers, MEG is emerging as a next-generation wearable system for practical speech BCIs. However, it remains unclear how many sensors are needed, where to place them, and what sampling frequency is optimal for wearable MEG devices that require cost-effective, computationally efficient operation.

In this study, we addressed these gaps by using a forward selection algorithm to find an optimal subset of sensors for spontaneous intended and overt speech decoding and by evaluating decoding performance across different frequency bands. Support vector machine and random forest were used as the classifiers in these experiments.

Our results indicate that only a small number of sensors (7-8 on average, with a maximum of fewer than 24) are sufficient to achieve decoding performance comparable to that obtained using the whole scalp (more than 200) sensors. In addition, we observed that the Lower High-Gamma (62-125 Hz) band yielded the best performance in some settings, although not statistically significantly higher than other bands in all settings.

These findings suggest that future wearable MEG devices may not need the whole-scalp sensor set and should support a sampling rate of at least 250 Hz, while 500 Hz is also encouraged, if possible. In summary, these findings provide practical insights for MEG-based BCIs for communication.