TL;DR
This paper demonstrates that the ventral motor cortex (vPCG), typically associated with speech, can also be used to control a computer cursor and perform clicks. This opens up the possibility of creating multi-modal brain-computer interfaces (BCIs) that combine speech and cursor control from a single implanted area, potentially improving communication and computer access for individuals with paralysis.
What Problem Does This Paper Address?
For individuals with severe motor impairments, like those caused by ALS, brain-computer interfaces (BCIs) offer a pathway to regain communication and control. Historically, BCIs for cursor control have relied on neural signals from the dorsal motor cortex (dPCG), which is associated with arm and hand movements. Simultaneously, BCIs for high-performance speech communication have emerged, leveraging signals from the ventral motor cortex (vPCG), which is linked to speech and orofacial movements. This creates a dilemma: should surgeons implant electrodes in dPCG for cursor control or vPCG for speech? This paper investigates whether the vPCG, known for speech, can also support cursor control, potentially unifying these two critical BCI functions within a single brain region.
Methods
The researchers developed a cursor BCI using neural data recorded from intracortical microelectrode arrays implanted in the ventral precentral gyrus (vPCG) of a single participant with ALS (T15). This participant had previously used these arrays for a high-performance speech BCI.
Key methodological aspects include:
- Neural Recording: Four 64-electrode Utah arrays were implanted in T15's vPCG.
- Neural Feature Extraction: Raw voltage signals were processed to extract neural features, including threshold crossings and spike band power, binned into 10ms intervals.
- Cursor Decoder: A linear velocity decoder with temporal smoothing was employed to predict cursor velocity from neural features. This decoder was trained and updated online during calibration tasks. A novel aspect was the gradual reduction of computer assist based on real-time decoder performance, leading to rapid calibration.
- Click Decoder: A linear classifier with temporal smoothing was used to detect attempted clicks.
- Calibration Task (Radial8): A task involving selecting targets arranged in a circle was used to train both cursor and click decoders. The system incorporated a "closed-loop" calibration where computer assistance was progressively reduced as the decoder's performance improved.
- Evaluation Task (Grid Evaluation): A standardized grid-based task was used to measure the BCI's performance in terms of bitrate (bits per second).
- Simultaneous Speech and Cursor Task: This task investigated the impact of speaking on cursor control by introducing auditory beeps that prompted speech during cursor tasks.
Datasets
The study utilized neural data collected from a single participant (T15) with ALS over multiple sessions. This data was used for both training and evaluating the BCI performance. No new benchmark datasets were introduced; rather, the study evaluated performance on established BCI tasks.
Key Insights and Ideas
- Ventral Motor Cortex for Cursor Control: The most significant insight is that the vPCG, traditionally associated with speech, can effectively support computer cursor control and clicking. This challenges the strict dichotomy of dorsal for motor and ventral for speech.
- Rapid Calibration: The BCI demonstrated remarkably fast calibration, with the participant achieving neural control of the cursor within 40 seconds of their first usage. This rapid onboarding is crucial for user adoption and practical application.
- Multi-modal Potential: The findings strongly suggest that vPCG is a promising target for implanting electrodes that can enable both speech-based communication and cursor-based computer control, leading to more versatile BCIs.
- Speech Interference: While vPCG can support cursor control, speaking simultaneously did transiently impair cursor control performance, unlike in some dPCG-based cursor BCIs. This highlights a potential trade-off that needs to be addressed in future designs.
- Distributed Representation: The study aligns with a growing body of evidence suggesting a more distributed representation of body movements across the entire precentral gyrus, rather than strict somatotopic mapping.
Key Results
- Cursor Control Performance: The vPCG-driven cursor BCI achieved a peak performance of 3.16 bits per second (bps), with an average performance of 2.90 ± 0.16 bps during optimized evaluation blocks. This is comparable to high-performing cursor BCIs previously reported using dPCG.
- Clicking Performance: All four implanted arrays in the vPCG contained electrodes that modulated to attempted clicks, and the BCI achieved high classification accuracy for click detection.
- Rapid Onboarding: The participant acquired their first target using neural control just 40 seconds into their first-ever cursor BCI session.
- Personal Computer Control: The participant successfully used the BCI to control their personal computer independently, browsing websites and playing games.
- Speech Interference: Target acquisition times were significantly longer (median 4.51s) when the participant was speaking during cursor control compared to conditions without speech (median ~3.4s).
Limitations and Future Work
- Single Participant: The study's primary limitation is that it involved only one participant with ALS. The generalizability of these findings to other individuals and different neurological conditions needs to be assessed in future studies.
- Specific Array Dominance: Cursor velocity decoding was heavily reliant on a single array located in the dorsal aspect of area 6v within the vPCG. It remains to be seen if similar performance can be achieved from other parts of the vPCG or in other individuals.
- Speech Interference Mitigation: The interference of speech on cursor control, while not preventing daily use, suggests a need for future work to develop decoders that can explicitly account for speech-related neural modulation, potentially enabling more seamless simultaneous control.
- Decoder Architectures: The study used linear decoders. Future research could explore more advanced architectures, such as recurrent neural networks, which may yield even higher cursor decoding performance by capturing more complex neural dynamics.
- Motor Imagery Interpretation: The participant's "intuitive" control and reliance on "The Force" highlight the challenge of precisely interpreting instructed motor imageries, suggesting a need for future work to better understand how the brain represents and executes movements for BCI control.
Summary updated