Intracortical brain-computer interfaces (iBCIs) restore motor function to people with paralysis by translating brain activity into control signals for external devices. In current iBCIs, instabilities at the neural interface result in a degradation of decoding performance, which necessitates frequent supervised recalibration using new labeled data. One potential solution is to use the latent manifold structure that underlies neural population activity to facilitate a stable mapping between brain activity and behavior. Recent efforts using unsupervised approaches have improved iBCI stability using this principle; however, existing methods treat each time step as an independent sample and do not account for latent dynamics. Dynamics have been used to enable high performance prediction of movement intention, and may also help improve stabilization. Here, we present a platform for Nonlinear Manifold Alignment with Dynamics (NoMAD), which stabilizes iBCI decoding using recurrent neural network models of dynamics. NoMAD uses unsupervised distribution alignment to update the mapping of nonstationary neural data to a consistent set of neural dynamics, thereby providing stable input to the iBCI decoder. In applications to data from monkey motor cortex collected during motor tasks, NoMAD enables accurate behavioral decoding with unparalleled stability over weeks-to months-long timescales without any supervised recalibration.

Human Participant Details 

Participant T11 had two 96-channel intracortical microelectrode arrays (Blackrock Microsystems, Salt Lake City, Utah) placed chronically into the left precentral gyrus (PCG) as part of the BrainGate pilot clinical trial (www.ClinicalTrials.gov; Identifier: NCT00912041). Permission for this study was granted by the U.S. FDA (Investigational Device Exemption #G090003) and the IRBs of Massachusetts General Hospital, Providence VA Medical Center, and Brown University. The participant gave informed consent to the study and resulting publications. This data has been previously published in Rubin, et al., J Neurosci 2022.

Human Data Acquisition 

During recording, neural activity was recorded at 30kHz from 192 channels and processed using a custom signal processing system. After digital downsampling, threshold crossing events were extracted in real time using 20 ms time-steps. 

Human iBCI Experimental Paradigm 

The participant first completed a center-out calibration task to optimize a Kalman filter to be used for subsequent tasks. Kalman filter weights were obtained using the rapid calibration procedure described previously. The participant then rested for a 30 minute period during which baseline neural activity was recorded.

The participant then completed a number of blocks of a “Simon” memory matching task. During each trial, the researcher displayed a sequential illumination of one of four colored targets on the screen. Each illumination was 400 ms in duration and accompanied by a distinct tone. The participant was asked to imagine moving his right hand to drive the cursor from the center of the screen out to each target in the order presented. After dwelling on the correct target for 300 ms, the acquisition was deemed successful and the cursor was automatically recentered at the origin. If the participant completed the sequence of four target acquisitions successfully, a tone indicated success and the next trial’s sequence was presented. A failed trial occurred if the cursor was dwelled on an incorrect target or if the participant failed to move the cursor to any target within 5s.

Each trial’s sequence always used all four polygons exactly once. Each block contained 16 trials, with one sequence provided per trial. In any given block, 12 of the 16 total trials were a fixed “target” sequence. The target sequence was different for each recording session but was consistent across blocks within a given session. The remaining 4 trials were a randomly chosen distractor sequence. The distractor sequences were made up on any of the other possible 23 four-target sequences. The participant was not made aware that there was a target sequence. Between each block the participant was given the opportunity to pause or take a short break if needed. 

For purposes of this study, each individual movement in the sequence of 4 movements was considered separately. Only successful movements were considered, and any movement that was part of a failed trial was discarded. We defined the intended cursor position as the straight line distance from the center of the screen to the center of the target, which progresses linearly in time for 500ms, and the velocity as a constant speed over this time period.