Shared latent representations of speech production for cross-patient speech decoding

Area of Science:

• Neuroscience
• Biomedical Engineering
• Rehabilitation Technology

Background:

• Speech brain-computer interfaces (BCIs) offer communication restoration for individuals with neuromotor disorders.
• Current BCIs require extensive patient-specific data, limiting usability and deployment speed.
• Challenges in combining data across patients include neuroanatomical differences and varied electrode placement.

Purpose of the Study:

• To develop a method for training speech BCIs using combined data from multiple patients.
• To overcome limitations of patient-specific data requirements and accelerate BCI deployment.
• To improve the accuracy and usability of speech BCIs for individuals with communication impairments.

Main Methods:

• Utilized canonical correlation analysis (CCA) to align patient-specific neural data into a shared latent space.
• Employed high-density micro-electrocorticography (μECoG) for detailed neural signal capture.
• Developed cross-patient decoding models trained on aggregated and aligned neural data.

Main Results:

• Successfully uncovered shared neural latent dynamics across patients using CCA.
• Demonstrated that aligned neural data preserved micro-scale speech information.
• Achieved improved decoding accuracies with cross-patient models compared to patient-specific models.
• Highlighted the benefits of μECoG's high resolution and broad coverage.

Conclusions:

• Aligning neural data to a shared latent space enables effective training of speech BCIs with multi-patient data.
• This approach significantly enhances BCI accuracy and facilitates rapid deployment.
• Future speech BCIs can be more effective, improving quality of life for those with communication disorders.