Distributed cortico-subcortical networks enable robust speech state detection from sparse intracranial recordings.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biomedical Engineering

Background:

• Accurate speech state transition detection is crucial for speech brain-computer interfaces (BCIs).
• The precise neural network underlying speech onset detection, whether localized to cortical language areas or distributed, remains unclear.

Purpose of the Study:

• To investigate the feasibility of decoding speech state transitions using sparse stereo-electroencephalography (SEEG).
• To determine if speech onset information is localized or distributed across cortico-subcortical networks.

Main Methods:

• Four epilepsy patients undergoing SEEG monitoring performed a sentence-reading task.
• A convolutional neural network classified speech states using broadband or high-gamma features from cortical and subcortical channels.
• Evaluated continuous decoding, robustness to channel dropout, and regional contributions.

Main Results:

• Speech state decoding accuracy exceeded 50% in all participants, reaching over 90% in some.
• Models integrating both cortical and subcortical signals outperformed those using single domains.
• Decoding remained robust to significant channel reduction and exclusion of canonical speech areas.
• Broadband signals showed higher accuracy than high-gamma features.

Conclusions:

• Speech state information is represented in a distributed cortico-subcortical network.
• Sparsely sampled SEEG data can effectively drive speech detection for BCIs.
• Deep brain recordings offer a path toward stable and generalized speech BCI systems.