Volitional control of single-electrode high gamma local field potentials by people with paralysis

Abstract

Intracortical brain-computer interfaces (BCIs) can enable individuals to control effectors, such as a computer cursor, by directly decoding the user's movement intentions from action potentials and local field potentials (LFPs) recorded within the motor cortex. However, the accuracy and complexity of effector control achieved with such "biomimetic" BCIs will depend on the degree to which the intended movements used to elicit control modulate the neural activity. In particular, channels that do not record distinguishable action potentials and only record LFP modulations may be of limited use for BCI control. In contrast, a biofeedback approach may surpass these limitations by letting the participants generate new control signals and learn strategies that improve the volitional control of signals used for effector control. Here, we show that, by using a biofeedback paradigm, three individuals with tetraplegia achieved volitional control of gamma LFPs (40-400 Hz) recorded by a single microelectrode implanted in the precentral gyrus. Control was improved over a pair of consecutive sessions up to 3 days apart. In all but one session, the channel used to achieve control lacked distinguishable action potentials. Our results indicate that biofeedback LFP-based BCIs may potentially contribute to the neural modulation necessary to obtain reliable and useful control of effectors. NEW & NOTEWORTHY Our study demonstrates that people with tetraplegia can volitionally control individual high-gamma local-field potential (LFP) channels recorded from the motor cortex, and that this control can be improved using biofeedback. Motor cortical LFP signals are thought to be both informative and stable intracortical signals and, thus, of importance for future brain-computer interfaces.

Keywords: biofeedback; brain-computer interface; local field potentials; motor cortex; people with tetraplegia.

Conflict of interest statement

No conflicts of interest, financial or otherwise, are declared by the authors.

Figures

Fig. 1.

Experimental setting and task design. A: participants were implanted with one (T2, T6) or two (T7) intracortical microelectrode arrays in the arm/hand area of the precentral sulcus of their left (dominant) hemispheres. An array covered an area of 4×4 mm and consisted of 96 1.0-mm- or 1.5-mm-long electrodes. Neural recordings were transmitted by a wire bundle to a pedestal and then through a cable to an amplifier positioned on the participant’s wheelchair. B: sessions took place at each participant’s home, where the participant was comfortably seated in front of the computer monitor that showed the task. The technician attended the sessions, started and stopped tasks, and managed the session workflow. C: session design. Participants first observed the task without attempting any actions. The technician started the task in which she moved the cursor by moving the computer mouse while explaining the task to the participant (calibration block). After the calibration block, the technician calibrated the spectral amplitude normalization and loaded the initial scaling from the data file. The following two exploration blocks were each followed by a recalibration of the scaling. For long test block sessions, the second exploration block was followed by four consecutive long test blocks (50 targets). In short test block sessions, the second exploration block was followed by a dummy short test block (20 targets), another exploration block, and 10 consecutive short test blocks. During the exploration blocks (D), the participant would only see the current cursor position and its history (blue line). During the test blocks (E and F), the participant’s task was to touch the target with the cursor. When the cursor touched the target, it turned green (F); it was yellow otherwise (E). In D, E, and F, the black screen background seen by the participant is depicted gray to better illustrate the cursor trail.

Fig. 2.

Signal processing and the measure of participants’ volitional high-frequency selected band of local field potential (HFSB-LFP) control. A: relationship of the neural activity to the cursor position on the screen. Voltage recordings from a single electrode were first common average rereferenced and then transformed into spectral amplitudes. Amplitudes in different frequency bins were normalized and averaged over the selected frequency band. The natural logarithm of the value was taken and filtered using a 0.8-s-long (blue line) or a 0.4-s-long (yellow line) linear weighted moving average (LWMA) filter to obtain the neural feature. The neural feature was then scaled to obtain the current cursor position. B: example of cursor and target positions in one of the short test blocks (left). Participant’s ability to control the cursor was measured by the mean distance from the edge of the cursor to the edge of the target (right). If the cursor touched the target, the target was acquired and the distance for that time point was 0. C: the baseline performance was measured by calculating mean cursor edge-to-target edge distance for 10,000 random shuffles of the target positions (example of one shuffle shown at left). The participant’s volitional control in relation to baseline was measured by calculating relative improvement in control (RIC). D: example of participant T7’s volitional HFSB-LFP control performance for one of the sessions. Left shows mean cursor-edge-to-target-edge distance during the session (online, green histograms) and for shuffled targets (magenta histograms) for each test block (B1–B10) and overall. Right: RIC for each test block and overall. For this session, the participant had statistically significant control (P < 0.01, t-test) in every test block. *Statistically significant RIC values.

Fig. 3.

Frequency band signal-to-noise ratio (SNRFB) for each session used to identify frequency bands used for cursor control. For session T7 day 100, we used a single-electrode SNRFB to identify the frequency band. For all other sessions, we used an electrode-average SNRFB. Black circles and arrows show the identified local maximum of SNRFB above with bottom band frequency above 40 Hz. Top and bottom frequencies of the SNRFB local maximum are written in white.

Fig. 4.

Examples of blocks with targets for each of the participants in the study. In this figure, target color was changed from green to orange to improve visibility.

Fig. 5.

Cursor control performance of participants in all completed sessions. A: overall relative improvement in control (RIC) for all completed sessions. For three participants, horizontal axis shows days after microelectrode array implantation. Most sessions evaluated cursor control using a long feedback delay and a local field potential (LFP) signal from a different electrode each day (solid blue bars). For T6 and T7, we further investigated the effect of feedback delay on volitional control by changing the signal processing parameters to switch from “long” feedback delays (bars containing blue) to “short” feedback delays (four bars containing orange). For T6 and T7, we also examined volitional LFP control when the same electrode and the same frequency band was used for cursor control across consecutive sessions (brown-striped bars). *Statistically significant RIC values (P < 0.01, rank sum test). B: bar plot shows normalized path to target measured from testing blocks (online) and after randomly shuffling the targets in all testing blocks 10,000 times (shuffled). Bars and error bars for online condition show mean and standard errors over all targets. Bars, lower and upper error bars for shuffled condition show the median, 5th and 95th percentile over 10,000 shuffles. *Statistically significant differences between online and shuffled conditions (P < 0.05, Monte Carlo test). C: bar plot shows normalized time to target measured from testing blocks (online) and after randomly shuffling the targets in all testing blocks 10,000 times (shuffled). Bars and error bars for online condition show mean and standard errors over all targets. Bars, lower, and upper error bars for shuffled condition show the mean, 5th, and 95th percentile over 10,000 shuffles. *Statistically significant differences between online and shuffled conditions (P < 0.05, Monte Carlo test).

Fig. 6.

Scatter plot showing relative improvement in control (RIC) from the 1st and 2nd halves of each session. Individual sessions where 2nd half RIC was significantly higher than 1st half RIC (P < 0.05, t-test) are marked with an asterisk. Over all sessions, 1st half RIC was not significantly different from 2nd half RIC (P = 0.64, signed rank sum test). T2, T6, T7, individual participants.

Fig. 7.

Evaluating the possible contribution of action potentials of neurons in the neighborhood of the electrode used for to cursor control. A: identification of action potentials of neurons in the neighborhood of the electrode used for cursor control. For each cursor control session, a pair of panels show the results of the neuron identification procedure. Top (heat maps): density of threshold crossings in the space spanned by the first two principal components (PCs). Density has been normalized to the maximum density. Bottom: mean of all threshold crossings contained within the 95% of local maximum density contour (full line), and the limit spike shape amplitude aLIM originating from the spike shape maximum (line with arrows) and as a threshold (broken line). Out of all cursor control sessions, a single active neuron was identified for session T2 day 502 only (first row, first column, red spike shape). B: cursor control contribution of actions potentials of neurons in the neighborhood of the electrode used for control in session T2 day 502. Mean action potential shape (blue line) and standard error of the mean (transparent blue tube, not easily seen due to small values) of the action potential identified for session T2 day 502. The shape from the first zero crossing (left black triangle) to the last zero crossing (right black triangle) was subtracted from the electrode recordings used for control in session T2 day 502 to get the recordings with spikes removed. C: We then used the recordings with spikes removed to recalculate the cursor positions and used those to calculate the relative improvement in control (RIC) for each test block and overall. RIC calculated from signals with spikes removed was compared with the RIC from the cursor control session. The difference was not significant.

Fig. 8.

Correlation between the local field potential (LFP) used for volitional control [high-frequency selected band of LFPs (HFSB-LFP) on the electrode used for volition control] and the HFSB-LFP on other electrodes of the intracortical array without filtering the HFSB-LFP using the linear weighted moving average [no linear weighted moving average (LWMA)], filtered using 0.4-s-long LWMA (0.4 s LWMA), and filtered using 0.8-s-long LWMA (0.8 s LWMA). A: color plots show the Pearson’s correlation coefficient between the HFSB-LFP on the electrode used for control and all other electrodes for the frequency band used for control (CCel) calculated over all test blocks in the session. For each session, top, middle, and bottom rows show the correlation obtained without filtering the HFSB-LFP, when filtering the HFSB-LFP using 0.4 s LWMA, and when filtering the HFSB-LFP using 0.8 s LWMA. A brown dot denotes the electrode used for cursor control. B: plots show CCel for different distances from the electrode used for cursor control for each participant (T2, left; T6, middle; T7, right) for three temporal filters of the HFSB-LFP (blue: no LWMA; magenta: 0.4 s LWMA; red: 0.8 LWMA).

Fig. 9.

Identification of frequency bands that correlated or anticorrelated with the neural control signal. Color plots show absolute value of the mean of electrode- and frequency band-dependent Pearson's linear correlation coefficient CC(el,fb,ft,blk) across all blocks, electrodes, and sessions for each participant (T2, left; T6, middle; T7, right). The analysis uncovered one low-frequency band that correlated with the control signal and one intermediate-frequency band that anticorrelated with the control signal.

Fig. 10.

Correlation between the local field potential (LFP) used for volitional control [high-frequency selected band of LFPs (HFSB-LFP) on the electrode used for volition control] and low-frequency selected band (LFSB-LFP) identified through the correlation analysis (Fig. 9). A: color plots show the mean absolute value of the Pearson’s correlation coefficient between the HFSB-LFP on the electrode used for control and LFSB-LFP on all electrodes of the microelectrode array calculated over all the test blocks in the session, LFSB-CCel. A brown square shows the electrode used for control. B: plots show LFSB-CCel for different distances from the electrode used for cursor control for each participant (T2, left; T6, middle; T7, right).

Fig. 11.

Correlation between the local field potential (LFP) used for volitional control [high-frequency selected band of LFPs (HFSB-LFP) on the electrode used for volition control] and intermediate-frequency selected band (IFSB-LFP) identified through the correlation analysis (Fig. 9). A: color plots show the mean absolute value of the Pearson’s correlation coefficient between the HFSB-LFP on the electrode used for control and IFSB-LFP on all electrodes of the microelectrode array calculated over all the test blocks in the session, IFSB-CCel. A brown square shows the electrode used for control. B: plots show IFSB-CCel for different distances from the electrode used for cursor control for each participant (T2, left; T6, middle; T7, right).