Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
Comparative Study
. 2008 Jan 30;28(5):1163-78.
doi: 10.1523/JNEUROSCI.4415-07.2008.

Primary Motor Cortex Tuning to Intended Movement Kinematics in Humans With Tetraplegia

Affiliations
Free PMC article
Comparative Study

Primary Motor Cortex Tuning to Intended Movement Kinematics in Humans With Tetraplegia

Wilson Truccolo et al. J Neurosci. .
Free PMC article

Abstract

The relationship between spiking activities in motor cortex and movement kinematics has been well studied in neurologically intact nonhuman primates. We examined the relationship between spiking activities in primary motor cortex (M1) and intended movement kinematics (position and velocity) using 96-microelectrode arrays chronically implanted in two humans with tetraplegia. Study participants were asked to perform two different tasks: imagined pursuit tracking of a cursor moving on a computer screen and a "neural cursor center-out" task in which cursor position was controlled by the participant's neural activity. In the pursuit tracking task, the majority of neurons were significantly tuned: 90% were tuned to velocity and 86% were tuned to position in one participant; 95% and 84%, respectively, in the other. Additionally, velocity and position of the tracked cursor could be decoded from the ensemble of neurons. In the neural cursor center-out task, tuning to direction of the intended target was well captured by a log-linear cosine function. Neural spiking soon after target appearance could be used to classify the intended target with an accuracy of 95% in one participant, and 80% in the other. It was also possible to extract information about the direction of the difference vector between the target position and the instantaneous neural cursor position. Our results indicate that correlations between spiking activity and intended movement velocity and position are present in human M1 after the loss of descending motor pathways, and that M1 spiking activities share many kinematic tuning features whether movement is imagined by humans with tetraplegia, or is performed as shown previously in able-bodied nonhuman primates.

Figures

Figure 1.
Figure 1.
Spiking activity. A, Average spike waveforms of sorted units in the corresponding microelectrode (labeled C1, C2, etc., of 96 recording channels) for S1. Examples are from sessions 1 and 2. In some cases, more than one unit was identified in the same electrode and session (for example, C10). The vertical bar on the left of each of the waveforms corresponds to 100 μV (i.e., −50 to 50 μV). Waveforms span 1.6 ms. B, Average spike waveforms for S3. Examples from session 1 are shown. Similar examples were recorded in session 2. C, D, Two examples of recorded spike waveforms and the corresponding ISI distribution are shown in C and D for S1 and S3, respectively. The thick darker curve corresponds to the average waveform in the session. Three hundred randomly chosen waveforms are shown.
Figure 2.
Figure 2.
Pursuit tracking task: kinematics of the TC. A, Examples of the paths of the cursor tracked by S1 and S3. Each plot corresponds to a 1 min tracking block. A technician moved the cursor (TC) to targets (data not shown) that appeared sequentially in random locations on the screen. B, Joint distribution of direction (in radians) and magnitude (speed) of the velocity vectors from the pursuit tracking blocks (two sessions) for S1 and S3. Speed is given in visual angle per second (degree/second). Speeds of the presented TC were generally higher for S1 than S3. The two plots on the right show the marginal distributions for direction and speed, respectively. C, Correlation functions. Autocorrelation functions, C(·, τ), for position and velocity in both horizontal (x) and vertical (y) coordinates are shown in the top plots. The bottom plots show the cross-correlation functions, denoted by C(·, ·, τ). Blue and red curves correspond to correlation functions for S1 and S3, respectively.
Figure 3.
Figure 3.
M1 velocity tuning during pursuit tracking. A, B, Optimal time lag analysis. Examples of velocity tuning functions at different time lags for a neuron (C10a, S1, session 1) are shown in A. Velocity is represented in polar coordinates. The color coding represents spiking rates with units of spikes per second. The corresponding log-likelihood under the velocity model as a function of different time lags for these examples is shown in B. The stars on the curve indicate that the velocity model was statistically significant for that time lag. This neuron seems to be maximally tuned to velocity at τ = 0.5 s (i.e., to future values of the TC's velocity), suggesting anticipation effects. C, Distribution of the optimal time lags over the population of tuned neurons in S1. D, Distribution of the optimal time lags over the population of tuned neurons in S3. Positive time lags suggest anticipation effects. Such effects could happen because a given target was present on the screen until it was acquired by the TC. E, F, Summary of PDs and spike rate modulation for the population of velocity tuned neurons in S1 (E) and S3 (F). The direction of each vector gives the preferred direction (in degrees) and the length of the vector denotes the tuning depth or spike rate modulation caused by variations in velocity. Distribution for the same preferred directions and tuning depths over the tuned population in S1 and S3 are shown in G and H, respectively. Summaries and histograms include neurons from the two recording sessions.
Figure 4.
Figure 4.
Off-line Decoding of TC velocity during pursuit tracking. Spike trains of velocity 2-tuned neurons were used, together with the fitted velocity models (tuning functions as shown in Fig. 3A), to generate a neural ensemble prediction of the velocity of the tracked cursor. In the examples shown, the models were fitted first to data from three 1 min pursuit tracking blocks (training data) and then used to decode velocity on test data consisting of a different 1 min block. A stochastic state point process filter was used for decoding (Eq. 7) (Eqs. S1–S7 in the supplemental material, available at www.jneurosci.org). At time 0, a random velocity vector was assigned to the initial state and then at every millisecond, during the 60 s of tracking, the TC velocity vector was updated according to the stochastic state point process filter output. For plotting purposes, the decoded velocity vector is shown in polar coordinates, with speed given in visual angle (degree) per second. A, C, The neural decoded movement direction (red) and the actual direction (black) are shown for S1 (A) and S3 (C). The dashed portion of the black curve represents segments where we considered the cursor to be moving too slowly to have a clearly defined or perceivable movement direction (speed, <2.5°/s). The decoded speed of the tracked cursor is shown in B for S1 and D for S3. For computational simplicity, a single time lag was used for all of the neurons in the ensemble.
Figure 5.
Figure 5.
M1 position tuning during pursuit tracking. A, B, An example of a position-tuned neuron (C18a, S1, session 1) is shown. The empirical mean spike rate conditioned on x,y position of the tracked cursor at τ = 1 s is shown in A. To compute the conditional mean rate, we used spike counts in 100 ms bins. The conditional mean rate is given in spikes per second (pseudocolor coding). Blank bins refer to states in the position space where we considered there were not enough samples to estimate the conditional mean (i.e., bins with <10 samples during the four 1 min pursuit tracking blocks). Position is given in centimeters. The corresponding position tuning function based on position model (Eq. 1) fitted to the actual point process data (1 ms time resolution spike trains, not the conditional mean) is shown in B. Summaries of the PD in position space and of the tuning depth modulation over the population of tuned neurons are shown in C for S1 and E for S3. These summaries include neurons from the two recording sessions. The corresponding distributions of optimal time lags (OTL) are shown in D for S1 and F for S3. Optimal time lags were determined based on the same log-likelihood analysis used for the velocity model (see Fig. 3A,B).
Figure 6.
Figure 6.
Velocity and position spike prediction power during pursuit tracking: ROC analysis. A, ROC curve example for one neuron. The curve was constructed by computing the true positive rate (hit probability) and false positive rate in spike prediction based on a specific model (position model in this example) and different threshold values. The probability of a spike at a given discrete time is derived from the instantaneous spiking intensity given a model. If this probability is higher than a specified threshold, a spike is predicted. The shaded region corresponds to the AUC. B, The position and velocity models' power to predict a spike are shown for S1. Each point value corresponds to predictive power for a particular tuned neuron from either session 1 or 2. The diagonal line corresponds to equal prediction power for position and velocity. The prediction power was defined as 2 × AUC − 1, ranging from 0 to 1 (maximum power). For details, see Materials and Methods (Eq. 6). C, Same as in B, but for S3. Tuned neurons from session 1 and 2 were included. Over all, differences between predictive power of position and velocity models were small in this pursuit tracking task.
Figure 7.
Figure 7.
M1 spiking activity and kinematics during pursuit tracking: correlation analysis. A, B, A model independent assessment of the relative importance of different kinematic covariates [position (pos), velocity (vel), speed, and direction] to M1 spiking activity was obtained by computing cross-correlation functions between each covariate and the sequence of spike counts for each neuron (see Materials and Methods for details). The plots summarize the distribution of the maximum absolute cross-correlation function values between spike count sequences and the time series of a specific kinematic covariate over the recorded population of neurons, including data from both sessions 1 and 2. The bottom and top boundary of each box corresponds to the 2.5 and 97.5 percentiles, respectively; the interior line shows the median and the crosses correspond to the computed maximum absolute correlation values for each neuron. A sign test, applied to data from each session separately, revealed no mean differences (p > 0.05) between cross-correlation values for position and velocity in each coordinate. Highly significant differences were detected when comparing correlation means for direction and speed (sign test, p < 0.0001) in both S1 and S3.
Figure 8.
Figure 8.
A, B, Target onset PETHs during neural cursor center-out task. Examples of PETHs for nine different neurons from S1 (A) and S3 (B) from either session 1 or 2 are shown. Each row in the figure corresponds to a particular neuron (e.g., C18a), and the four columns correspond to the four target directions (i.e., right, up, left, and down). The PETH for the top neuron includes the corresponding raster plot. Target onset is at time 0. A time bin width of 50 ms was used to compute the spike counts and the corresponding spike rate. The red curve gives the smoothed histogram. Each PETH was computed from 20 center-out trials.
Figure 9.
Figure 9.
M1 tuning to target direction during neural cursor center-out task. Examples of direction tuning functions are shown in A and C for S1 and S3, respectively. Direction tuning functions were fitted to single neuron spike count data collected over a 400 ms time window covering the time interval 200–600 ms after target onset for S1 and either 150–550 ms or 400–800 ms, depending on the neuron, for S3. The four red stars represent the observed mean spiking rate conditioned on the four intended movement directions. Models were fitted to single trial data, not to the computed conditional mean rates. The fitted model (black curve) allowed interpolation to directions in [0, 360] degrees. The red curves give the 95% confidence interval estimated via bootstrap resampling (5000 samples). Different neurons from either session 1 or 2 were used in these examples. The polar plots in B and D summarize the estimated preferred directions for the tuned neurons in both sessions for S1 and S3, respectively. The length of the vector represents the depth in modulation in the tuning function.
Figure 10.
Figure 10.
Decoding of difference vector direction during neural cursor center-out task. A, The difference vector is illustrated. The red and blue circles represent the target and the neural cursor, respectively, and the red and blue arrows represent the corresponding target and instantaneous neural cursor position vectors. The green arrow represents the difference vector between the target and neural cursor position vectors, and φ(t) denotes the instantaneous angle or direction of this difference vector. The dashed curve represents the ongoing trajectory of the neural cursor during a center-out trial. B, The normalized histogram of the error between the true direction and the decoded direction, |φ(t) − φ(t)| (in degrees), obtained by decoding the difference vector direction during the two center-out sessions for both participants. The horizontal line represents the uniform distribution expected from a “random” decoder. The fact that error concentrated at angle errors below 45° suggests that information about the direction of the difference vector was available in the M1 neural ensemble during the neural cursor center-out task. C, Estimated predictive power of the difference vector direction based on the AUC measure (for details, see Materials and Methods, Eq. 6, Fig. 6). The crosses correspond to the estimated predictive power values for each neuron; each box corresponds to the 95% data interval (over the neuronal population), and the interior line represents the median predictive power value.
Figure 11.
Figure 11.
Off-line classification of intended target during neural cursor center-out task. A, B, The distribution of times required for S1 and S3, respectively, to acquire the target with the neural cursor in all of the successful trials (both sessions). C and D plot the proportion of correct classification as a function of time, based on counts computed from windows of different lengths (100, 500, and 700 ms) for S1 and S3, respectively. The intended target for each of the center-out trials was decoded based on spike counts. Each window was shifted by 25 ms to obtain the time course of the classification performance. Black dots represent the correct classification rate computed separately for each of the two sessions. The red curves represent a smooth fit to these points. The black line at 0.25 represents the chance level of correct classification given four targets. Time 0 corresponds to target onset time. Only successful trials that lasted >2 s from target appearance to target acquisition were used in this analysis (for details, see Materials and Methods). In E and G, the maximum achieved correct classification and the corresponding peak time based on all of the explored window lengths (100, …, 700 ms) are shown for S1. F, H, Similarly for S3. The two points for each window length correspond to the values for sessions 1 and 2.

Similar articles

See all similar articles

Cited by 84 articles

See all "Cited by" articles

Publication types

LinkOut - more resources

Feedback