Learned Motor Patterns Are Replayed in Human Motor Cortex during Sleep

Abstract

Consolidation of memory is believed to involve offline replay of neural activity. While amply demonstrated in rodents, evidence for replay in humans, particularly regarding motor memory, is less compelling. To determine whether replay occurs after motor learning, we sought to record from motor cortex during a novel motor task and subsequent overnight sleep. A 36-year-old man with tetraplegia secondary to cervical spinal cord injury enrolled in the ongoing BrainGate brain-computer interface pilot clinical trial had two 96-channel intracortical microelectrode arrays placed chronically into left precentral gyrus. Single- and multi-unit activity was recorded while he played a color/sound sequence matching memory game. Intended movements were decoded from motor cortical neuronal activity by a real-time steady-state Kalman filter that allowed the participant to control a neurally driven cursor on the screen. Intracortical neural activity from precentral gyrus and 2-lead scalp EEG were recorded overnight as he slept. When decoded using the same steady-state Kalman filter parameters, intracortical neural signals recorded overnight replayed the target sequence from the memory game at intervals throughout at a frequency significantly greater than expected by chance. Replay events occurred at speeds ranging from 1 to 4 times as fast as initial task execution and were most frequently observed during slow-wave sleep. These results demonstrate that recent visuomotor skill acquisition in humans may be accompanied by replay of the corresponding motor cortex neural activity during sleep.SIGNIFICANCE STATEMENT Within cortex, the acquisition of information is often followed by the offline recapitulation of specific sequences of neural firing. Replay of recent activity is enriched during sleep and may support the consolidation of learning and memory. Using an intracortical brain-computer interface, we recorded and decoded activity from motor cortex as a human research participant performed a novel motor task. By decoding neural activity throughout subsequent sleep, we find that neural sequences underlying the recently practiced motor task are repeated throughout the night, providing direct evidence of replay in human motor cortex during sleep. This approach, using an optimized brain-computer interface decoder to characterize neural activity during sleep, provides a framework for future studies exploring replay, learning, and memory.

Keywords: brain computer interface; learning; memory; replay; sleep.

Copyright © 2022 Rubin et al.

Figures

Figure 1.

Overview of experimental paradigm. A, Timeline of the recording sessions. On the second of the two recording dates, we recorded intracortical neural data and surface EEG during the night before the experimental session. Both recording sessions then otherwise began with a brief intracortical/EEG sync routine, followed by the standard calibration routine. B, During the calibration routine, the participant completes a radial-eight center-out task, during which he is cued to move the cursor to 1 of 8 targets surrounding the center starting point. Early in the calibration routine, the error attenuation (EA) (on the accompanying graph) is set close to 1, such that most of the cursor movement is performed automatically. The internal weights of the steady-state Kalman filter are iteratively updated as the EA is reduced in a stepwise fashion and the participant is given full control of the neurally driven cursor. After calibration, the participant then took a 30 min rest during which we recorded both intracortical and surface signals. Following the rest period, the participant engaged in the motor sequence memory task (detailed in C). After 10 blocks of the motor task, the participant took another 30 min rest. We then recorded neural signals overnight as the participant slept. C, The motor sequence memory task. Every trial begins with a sequence demonstration, during which 1 of the 4 colored shapes is sequentially illuminated as a brief auditory tone is played. This sequence is played over 3 s. After the sequence demonstration is complete, a small neurally controlled cursor appears in the middle of the screen. The participant's task is to move the cursor to each of the 4 shapes in the same sequence as was just demonstrated. The participant completed 10 blocks of 16 trials on each recording session day. On each recording session day, there was one “target” sequence (which was different between the two sessions). In each block, 12 of the 16 trials were of the target sequence, and the other 4 trials were randomly generated “distractor” sequences.

Figure 2.

Neural signals for this study were acquired from two microelectrode arrays chronically implanted in left precentral gyrus. A, A three-dimensional reconstruction of the research participants' brain. Blue represents left precentral gyrus. Green and blue squares represent the precise location of the two Utah arrays. B, Spike sorting reveals single-unit and multiunit activity on each array (medial array on the left, lateral array on the right) during the first (top) and second (bottom) recording session. All amplitude plots demonstrate mean ± SD of activity for each threshold crossing event, with color representing event type based on automated sorting algorithm (Vargas-Irwin and Donoghue, 2007). Gray amplitude plots represent multiunit activity. Blue, orange, and yellow amplitude plots represent isolated single-unit activity. There were 78 and 51 single units isolated from the medial array and 12 and 6 single units isolated from the lateral array from Sessions 1 and 2, respectively. The smaller number of single units recorded from the lateral array comports with the finding that activity on the lateral array is typically sparser and less closely associated with task performance.

Figure 3.

Examples of neural data and processed Kalman filter output during awake task performance and sleep. A, Data from a single experimental block of the motor sequence memory task. Top row, Spectrogram of the surface EEG signal recorded during the block. Second row, Spike raster from the medial PCG multielectrode array, showing increased neural firing during intended movement. Displayed activity is a thresholded, z-scored spike raster (with marks indicating times when spiking activity z > 1). Third row, Cursor position during each trial of the matching task. Blue line indicates X position. Red line indicates Y position. The cursor is reset to the origin after each successful target acquisition. Green shading represents successful trials. Red shading represents unsuccessful trials. A successful target trial (e.g., Trial 1, at 15:08:20) is the following sequence: up, down, right, left. Target trials indicated by light gray shading on the fourth row; distractor trials are indicated by dark gray shading on the fourth row. Fourth row, The intracortical neural data from the recording session are projected through the Kalman filter to generate a continuous model of intended cursor position (before online filtering/scaling/midpoint correction). Blue line indicates X position. Red line indicates Y position. The intended direction/magnitude of movement is evident even without the online postprocessing applied during active task participation. We used the mean output of the Kalman filter model during successful target trials to generate templates of trial-specific activity. Bottom row, Instantaneous correlation between Kalman filter output and successful target trial template. Blue line indicates X position. Red line indicates Y position. Dashed horizontal lines (blue: 0.71, red: 0.69) indicate the 99th percentile values of the correlation coefficients, which we define as the threshold for a template “match.” Vertical green dashed lines indicate the start time of STCEs, which, though indicated by a discrete event marker on this plot, encapsulate a subsequent trajectory of neural firing. In this example, there are STCEs at 8 of the 9 successful target sequence trials, but at none of the unsuccessful target trials nor during any of the successful or unsuccessful distractor sequence trials. B, Template target trajectories for the Kalman filter output in the X (top) and Y (bottom) dimensions. Templates were generated by aligning and averaging the offline-generated predicted cursor position for each of the successful target trials. The mean values (solid line) are used for subsequent template matching calculations. Shading around the solid line represents the SEM curve. C, Two examples of putative replay events occurring during overnight sleep following the motor task completion demonstrated above from experimental Session 2. Top row, EEG spectrogram showing predominately δ power, indicating a period of SWS. Second row, Relative power in low γ (40-125 Hz) power band from the medial PCG array, demonstrating brief bursts (or ripples) of activity lasting ∼200 ms during sleep. Third row, Spike raster during sleep; as above displaying thresholded, z-scored activity from the medial PCG array. Fourth row, Kalman filter output demonstrating replay of 2D trajectory associated with the target task during sleep. Bottom row, Same as in A, the instantaneous correlation between Kalman filter output and successful target trial template. Blue line indicates X position. Red line indicates Y position. Dashed horizontal lines indicate the 99th percentile values of the correlation coefficients which are used to define a threshold “match.” Vertical green dashed lines indicate the time of the STCEs. As above, though indicated by a discrete event marker on this plot, the STCEs encapsulate a subsequent trajectory of neural firing, which is why the stereotyped Kalman filter output trajectory follows the STCE by ∼2 s.

Figure 4.

Examples of preserved spiking order during replay events. A, One task completion event from an awake task completion block from the second recording session, serving as a template for the matching indices Im demonstrated in B. Top row, EEG spectrogram showing a normal awake power spectrum with predominant alpha activity. Second row, Thresholded, z-scored spike raster (showing as spikes all indices with spiking activity z > 1) from the medial PCG array. Channels are ordered from top to bottom in order of timing of peak firing rate during the task completion. Bottom row, Kalman filter output demonstrating 2D trajectory associated with target task completion. B, Four examples of putative replay events occurring during overnight sleep, demonstrating preservation of spiking order during putative replay events. Top row, EEG spectrogram showing predominately δ power, indicating a period of SWS. Second row, z-scored spike raster, with channel order as in A, demonstrating preservation of spiking order. Matching index Im and p value are indicated in the plot title (compared with control distribution). Green line indicates the best-fit linear regression line between channel order from A and the timing of the peak firing rate from the replay event being displayed. Bottom row, Kalman filter output demonstrating 2D trajectory associated with the target task completion during sleep. C, Histograms demonstrating the distributions of matching indices during each session; on each plot, the black dashed line indicates 0 and the red dashed line indicates the mean of the distribution. Leftmost panel, Matching indices generated from 100,000 pairs of randomly generated 96 element sequences, which serves as a control distribution. The second histogram represents the distribution of Im across all pairs of successful target trials during Session 1; the mean is significantly greater than the control distribution (p < 1 × 10−100, Student's t test). The third histogram represents the distribution of Im across all pairs of successful target trials and overnight STCEs during Session 1; the mean of this distribution is also significantly greater than the control (p = 7.19 × 10−82, Student's t test). The fourth distribution shows the distribution of Im across all pairs of distractor trials and overnight STCEs during Session 1; the mean of this distribution is not significantly different from the control (p = 0.10118, Student's t test). The fifth through seventh histograms represent analogous distributions generated during Session 2; the distributions of Im across all pairs of successful target trials and all pairs of successful target trials with overnight STCEs are both significantly greater than the control distribution (target trials vs target trials: p < 1 × 10−100, Student's t test; target trials vs overnight STCEs: p = 1.17 × 10−77, Student's t test). The distribution of Im across all pairs of distractor trials and overnight STCEs is again not significantly different from the control (p = 0.96093, Student's t test).

Figure 5.

Controls demonstrating expected number of STCEs by chance versus actual number of putative replay events observed. Top left, Histogram represents the number of Session 1 overnight STCEs observed when matched on 100 pseudo-templates generated from randomly selected epochs of neural data recorded during the afternoon pretask rest session. These data are meant to represent the distribution of replay events that would be expected by chance. The number of putative replay events of the target templates observed overnight (66) is greater than would be expected by chance (p = 0.019802, Mann–Whitney U test). Top right, Histogram represents the number of Session 2 STCEs observed (again matched on 100 pseudo-templates generated from randomly selected epochs of neural data recorded during the afternoon pretask rest session) during the night before (blue bars) and after (orange bars) performing the motor sequence task. The number of STCEs observed the night before task performance (4) is not statistically different from would be expected by chance (p = 0.079208, Mann–Whitney U test); the number of STCEs observed the night after task performance (85) is significantly different from would be expected by chance (p = 0.019802, Mann–Whitney U test). The number of STCEs observed by chance is not different between the two nights (p = 0.61606, Mann–Whitney U test). Bottom left, Histogram represents the number of Session 1 overnight STCEs observed using the target template matched against Kalman filter output that has been randomized in phase in 5 min segments. This phase-randomized Kalman filter output effectively preserves all underlying neural spike statistics while shuffling the relative sequences of firing. The number of overnight putative replay events is significantly greater than the number of STCEs observed in the phase-randomized data set (p = 0.041577, t test). Bottom right, Histogram represents the number of Session 2 STCEs observed using the target template matched against the phase-randomized Kalman filter output during the night before (blue bars) and after (orange bars) performing the motor sequence task. The number of STCEs observed the night before task performance is not significantly different from the number expected by chance (p = 0.57529, t test); the number of STCEs observed the night after task performance is significantly greater than the number expected by chance (p = 3.2698 × 10−36,t test). The number of STCEs observed by chance the night after task performance is significantly greater than the number of STCEs observed by chance the night before task performance (p = 9.7723 × 10−49,t test).

Figure 6.

Number of putative replay events observed overnight versus the speed of replay. Bold dark blue represents the number of replay events at each replay speed (Night 2, recording Session 2). At each replay speed, a control distribution of number of STCEs was generated using the phase-randomized control bootstrapping procedure described above. All data points are demonstrated, with median, 25%/75%, and 5%/95% limits of the control distribution indicated on each plot. Although the number of STCEs observed by chance also increased with increasing τ (because shorter patterns of activity are more likely to occur by chance), the number of replay events was nonetheless significantly greater than expected by chance over a wide range of replay speeds.

Figure 7.

Timing of replay events over the course of the night: Session 2, Night 2. A, Scalp EEG spectrogram overnight demonstrates epochs of predominately δ activity, initially 90-120 min in duration but shortening over the course of the evening, with intervening epochs lasting 15-30 min of higher-frequency predominant activity, consistent with a typical night of human sleep. B, Red line indicates the 95% FPL of the EEG spectrogram shown in A; the value of this metric used as a threshold for SWS is varied systematically over a range of 2-6 Hz. Individual putative replay events are demonstrated as raster marks on the same time axis for replay at 1× (blue), 2× (red), 3× (yellow), and 4× (purple). C, The proportion of the night spent in SWS (dashed line) or proportion of putative replay events that occurred during SWS as a function of SWS-defining cutoff threshold. As the threshold defining SWS increases, both the proportion of the night defined as SWS as well as the number of replay events categorized as occurring during SWS tautologically increase. Over the range of threshold FPLs used to define SWS, the number of putative replay events that occur during SWS is significantly greater than the number of events that would be expected if replay events occurred randomly throughout the night (and were thus simply proportional to the fraction of the night spent in SWS). Shaded blue region represents a 95% CI (based on a binomial distribution) for the proportion of putative 1x replay events expected to be observed during SWS if simply proportional to the fraction of the night spent in SWS (95% CIs for 2×, 3×, and 4× replay events all fall within this shaded region as there are more overall replay events at these faster speeds).

Figure 8.

Relationship between replay events and ripple activity. A, Example of neural activity during an epoch of sleep. Top row, Scalp EEG spectrogram represents power predominately in the δ band. Second row, Average power in the ripple band (40-125 Hz) of the LFP from the medial intracortical array (z-scored within a 1 min sliding window) to identify peaks in the ripple-band activity. Third row, Intracortical spike raster from the medial array across all 96 channels. Fourth row, Kalman filter output represents frequent deflections in hypothetical cursor position most prominently during bursts of activity aligned with peaks in the ripple-band LFP. B, Replay events (STCEs) are aligned in time with ripples. Five different ripple magnitudes were examined, each quantified by the number of channels on which a ripple was simultaneously observed (of the 96 channels on the medial array). For each threshold, for each of four replay event speeds (1×, 2×, 3×, 4×), we examined what proportion of the overall number of STCEs (recorded on Night 2 of Session 2) occurred during a period classified as a ripple. Because each STCE is point process representing a trajectory of activity, the peak in neural activity lags the STCE by fixed period. To account for this lag, we varied the timestamp of all STCEs in 20 ms time-steps from −7 to 7 s and recalculated the proportion occurring during ripple on each iteration. On each plot, dashed horizontal lines indicate the baseline proportion of overnight sleep classified as ripple based on each threshold [0.49 for ripple on ≥ 1 channel (blue); 0.38 for ripple on ≥ 2 channels (red); 0.25 for ripple on ≥ 4 channels (yellow); 0.13 for ripple on ≥ 8 channels (purple); 0.05 for ripple on ≥ 16 channels (green)]. For each replay speed, regardless of ripple channel threshold, there is a clear peak in the proportion of replay events that occur during an epoch defined as a ripple that occurs at an offset equal to approximately half the template duration. As expected, the peak offset decreases monotonically with increasing replay speed. In a window around each peak, there is a refractory period during which a replay event is less likely than baseline to be observed. C, Cross-correlation between the Kalman filter template matching signal and the number of channels observed to have ripple activity in a ±5 s window around each putative replay event. Mean cross-correlation for the X dimension (red) and Y dimension (blue) ±95% CI is shown for each replay speed. On each plot, for each putative replay event, the time of the peak in the sum of the cross-correlograms for both the X and Y dimensions is shown as gray circles along the x axis; position along the y axis is randomly scattered to allow for visualization of each data point. The median time to the peak is 2.2 s for each replay speed examined.