Brain interface allows speech decoding and computer control in ALS patient - phys.org

Schematic of the cursor BCI. Credit: Tyler Singer-Clark et al

University of California, Davis researchers have developed a brain-computer interface (BCI) that enables computer cursor control and clicking, using neural signals from the speech motor cortex. One participant with amyotrophic lateral sclerosis (ALS) used the interface for daily life activities, including independent control of a personal desktop computer and text entry.

Neurological diseases such as stroke or ALS can interrupt the pathway from the brain to the muscles, causing a loss of movement and communication. ALS progressively destroys upper and lower motor neuron pathways, leaving cognition intact but causing paralysis in all four limbs and significant speech impairment.

Brain-computer interfaces are intracortical implanted devices that bypass any disruption by reading neural signals directly from the brain and producing output on the user's behalf. Many BCIs have relied on neural activity from the dorsal motor cortex, a brain region associated with hand and arm movements. When signals are decoded, users can move a cursor by attempting or imagining limb motion.

In contrast, speech BCIs rely on the ventral precentral gyrus, where neural signals are linked to facial movements and speech articulation. Decoding neural signals from this region enables fast, speech-based communication but has not been shown to support general computer navigation or motion control.

Implantation into both dorsal and ventral areas would be ideal, yet it is considered surgically impractical or infeasible. As a result, users and clinicians must choose between cursor control and speech decoding.

In the study, "Speech motor cortex enables BCI cursor control and click," published in the Journal of Neural Engineering, researchers conducted a single-participant case study to test whether neural activity from the speech motor cortex could support both cursor control and speech decoding with a single implant site.

One participant with ALS, a 45-year-old man with paralysis in all four limbs and difficulty speaking clearly, took part in the research. All sessions were run at the participant's home.

Four 64-electrode arrays were surgically implanted in the ventral precentral gyrus of the participant. Electrode targeting was guided by preoperative MRI and cortical alignment with the Human Connectome Project.

Neural signals were acquired at a sampling rate of 30 kHz and bandpass filtered between 250 and 5,000 Hz. Threshold crossings and spike band power were calculated every millisecond from each electrode. These features were then grouped into 10-millisecond bins, producing a stream of 512-dimensional feature vectors that served as input to the decoding systems.

Three task paradigms were used to evaluate system performance: Radial8 Calibration, Grid Evaluation, and Simultaneous Speech and Cursor. A linear velocity decoder controlled cursor movement, while a separate linear classifier decoded click events.

Decoder parameters were continuously recalibrated using linear regression for velocity and logistic regression for click classification with weights updated every few seconds during active control.

Calibration occurred quickly as the participant acquired his first target using neural control within 40 seconds after initiating the system.

During later sessions with optimized settings, the participant used the system to control the cursor with high efficiency, averaging 2.90 bits per second. Earlier sessions showed lower performance, averaging 1.67 bits per second. The highest rate recorded in any single session was 3.16 bits per second. One bit per second corresponds to the ability to make several accurate choices per minute, with higher values indicating faster and more precise control.

Across 1,263 total trials, 1,175 targets were correctly selected, corresponding to 93% accuracy. Eighty-eight incorrect selections occurred, and no trials ended due to timeout. Six clicks were registered on temporarily disabled targets, and 23 clicks occurred outside of any target boundary.

Click classification performance exceeded chance across all four electrode arrays. One well-placed array contributed the most to cursor decoding and closely matched the performance of the full-array decoder.

In sessions involving simultaneous speech and cursor control, median target acquisition time increased to 4.51 seconds. Conditions without speech ranged from 3.37 to 3.51 seconds, illustrating that speech production interfered with the participant's ability to control the cursor, yet did not cause delays in sequential actions. Improvements in decoder design could mitigate interference and enhance future usability.

A single implant site supported both communication and computing functions in an independent home setting, providing a proof of concept for the feasibility of multi-modal BCI systems.

For patients cognitively intact but unable to use their limbs or speak, a neural interface that provides both computer-cursor control and speech decoding can restore crucial channels of communication, independence, and substantially improve quality of life. 

by Justin Jackson , Medical Xpress

Source: Brain interface allows speech decoding and computer control in ALS patient