Preparatory encoding of intended movement in the human motor cortex and implications for brain-computer interfaces

Mattia Rigotti-Thompson; Samuel R Nason-Tomaszewski; Payton H Bechefsky; Alexander Acosta; Nick V Hahn; Donald T Avansino; Brice A Richards; Claire Nicolas; Yahia H Ali; Jaimie M Henderson; Leigh R Hochberg; Nicholas AuYong; Chethan Pandarinath

doi:10.1016/j.cub.2026.07.040

. Author manuscript; available in PMC: 2026 Oct 14.

Published in final edited form as: Curr Biol. 2026 Aug 11;36(17):4283–4296.e7. doi: 10.1016/j.cub.2026.07.040

Preparatory encoding of intended movement in the human motor cortex and implications for brain-computer interfaces

Mattia Rigotti-Thompson ¹, Samuel R Nason-Tomaszewski ¹, Payton H Bechefsky ¹, Alexander Acosta ⁴, Nick V Hahn ⁸, Donald T Avansino ^8,⁹, Brice A Richards ¹, Claire Nicolas ^4,⁵, Yahia H Ali ¹, Jaimie M Henderson ^8,¹⁰, Leigh R Hochberg ^4,^5,^6,⁷, Nicholas AuYong ^2,^1,^3,^†, Chethan Pandarinath ^1,^2,^†,^*

PMCID: PMC13660358 NIHMSID: NIHMS2204566 PMID: 42580347

Summary

Over the course of a voluntary movement, motor cortical activity exhibits a transition from preparation to execution, with markedly different activity across these phases. Preparatory activity in particular might be used to improve brain-computer interfaces (BCIs) that harness brain activity to control external assistive devices, for example by anticipating a user’s intended movement trajectory for quick and fluid performance. However, to leverage preparatory activity for clinical BCIs, we must first understand which features of upcoming movements are encoded by preparatory activity in humans. In this work, we collected intracortical recordings from 3 research participants in the BrainGate2 clinical trial to investigate whether diverse features of movement, such as direction, curvature, and distance, are encoded by preparatory activity in the human motor cortex. We first show that preparatory activity is tuned to the direction of upcoming movements, and this tuning is largely preserved across movements with different effectors. Further investigation demonstrated this preparatory activity is also informative of initial and target directions of curved movement trajectories, and encodes, to a weaker extent, either movement distance or speed. Finally, we present an online control paradigm that leverages preparatory activity to predict movements towards intended directions in advance, yielding rapid, user-initiated control of a computer cursor by human participants. Altogether, these results demonstrate that preparatory activity in the human motor cortex encodes rich features of upcoming movement, highlighting its potential use for high performance brain-computer interface applications.

eTOC blurb

Rigotti-Thompson et al. investigate preparatory activity in human motor cortex in three research participants, finding encoding of diverse, multi-level features of intended movement: direction, effector, curvature, and, less prominently, distance and speed. They demonstrate that preparatory activity can drive rapid, fluid cursor control, highlighting its potential for improving brain-computer interfaces.

Graphical Abstract

graphic file with name nihms-2204566-f0006.webp

Introduction

Voluntary movement execution is often preceded by a period of movement preparation¹. This preparation optimizes the upcoming movement for a particular goal, as evidenced by behavior becoming more precise with longer preparation times^2,3. Neurophysiological studies support a neural basis for motor preparation, as spiking activity that precedes execution is an inherent property of neuron populations within motor cortex^4–6. This “preparatory” activity is distinct from the neural activity observed during movement execution and can arise several seconds before any motor output^7–9. Interestingly, this preparatory activity also arises without any explicitly allotted preparation time^10–13, highlighting it as a critical component of movement generation. In this regard, its role is often linked to a priming mechanism that sets an optimal neural state for the upcoming movement^1,14,15.

Preparatory activity encodes several aspects of upcoming movement, as identified by extensive research in non-human primates. Using instructed-delay paradigms that explicitly separate preparatory and execution-related activity, studies have demonstrated that preparatory activity is informative of both temporal properties of upcoming movement, such as reaction time and speed^16–20, and spatial properties, such as direction, distance, and curvature^4,21–24.

Given its richness, preparatory activity might be used to improve brain-computer interfaces (BCIs) which aim to predict a user’s intended movement from neural recordings to control assistive devices^25–29. However, standard BCIs do not capitalize on preparatory activity, instead mapping execution-related neural activity onto instantaneous movement commands for an external effector. Such paradigms can produce inaccurate trajectories due to the characteristic moment-to-moment variability in neural activity. An alternative BCI paradigm might leverage preparatory information to predict intended movement features in advance and precompute trajectories for quick, fluid control. Previous non-human primate work has shown that BCIs predicting movement goals in advance can yield fast and accurate control during simple motor tasks^30–32.

Nonetheless, few studies have examined preparatory activity from the human motor cortex, particularly in people unable to perform overt movements as is typical for users of implantable BCIs. Other brain areas associated with motor planning and sensory integration (posterior parietal cortex) exhibit prominent pre-movement activity that is tuned to features of upcoming movement such as direction or selected action^33,34. Human motor cortex, in turn, shows preparatory activity that is distinct between simple movements towards different targets³⁵ or involving different finger groupings³⁴. However, building BCI control paradigms around precomputing intended trajectories in advance would require a deeper understanding of how a broad repertoire of movement features is represented in human preparatory activity. For example, movement curvature and distance are critical features whose preparatory encoding remains underexplored in humans.

In this work, we studied preparatory activity in the human motor cortex through intracortical recordings from three research participants enrolled in the BrainGate2 clinical trial. In all participants, the dorsal precentral gyrus exhibited preparatory activity strongly tuned to upcoming movement direction, with a substantial component preserved across movement effectors. To characterize preparatory tuning to a richer assortment of features, we designed tasks to probe movements with different curvatures, distances, and speeds. We found that preparatory activity was also tuned to both the initial and target directions of curved trajectories, and had slight, participant-specific tuning to either distance or speed. Lastly, as a proof of concept for practical BCI use, we designed an online control paradigm that leveraged preparatory activity to predict movements towards intended directions in advance, allowing rapid, user-initiated cursor control for two participants.

Results

The human motor cortex exhibits preparatory activity strongly tuned to upcoming movement direction

We collected multi-unit intracortical recordings from the motor cortex of participants T16, T11, and T5 enrolled in the BrainGate2 clinical trial. Specifically, we analysed the hand-knob area of the precentral gyrus³⁶ (putative Brodmann area 6 in Human Connectome Project cortical parcellations³⁷). To identify preparatory activity preceding movement initiation, we asked participants to perform an instructed-delay task using attempted wrist movements (Figure 1A). Of note is that attempted movements produced, at most, slight wrist deviations given participants’ limited upper-extremity movement capabilities. On each trial, one of eight possible targets (radial-8) was highlighted (“target cue”), followed by a randomized delay before a “go cue” to attempt a quick movement to the target (Figure 1B). For T16 and T11, a random subset of trials omitted the go cue, instructing participants to skip the movement attempt (no-go trials).

Figure 1. — A. Intracortical microelectrodes recorded from the hand-knob area of precentral gyrus in human participants preparing and attempting directional wrist movements. B. Trial structure and timing for the instructed-delay radial-8 task with example trial visuals. C. Peri-stimulus time histograms for example neural channels from each participant, for different movement directions. Left: delay period activity aligned to the target cue; right: movement and non-movement activity aligned to go/no-go cues (green and red lines, respectively). For T11 and T16, solid traces are movement-cued trials and dashed traces are non-movement trials. Shaded regions: +/− standard error of the mean. D. Percentage of active channels per participant with preparatory activity tuned to movement direction (p-value < 0.01, ANOVA). E. Average delay-period population activity for each direction (same color scheme as C) visualized by demixed principal component analysis; projections onto the top condition-independent (“CI 1”) and condition-dependent (“CD 1”, “CD 2”) components are shown. Trajectories span the time from the target cue (circle) to 1000 ms (T11, T16) or 800 ms (T5) after the target cue (triangle). F. Confusion matrices for single-trial direction predictions decoded from delay-period activity using SVM classifiers (leave-one-out cross-validation). G. Single-trial accuracies for SVM classifiers trained and tested across epochs (delay, movement, and no-movement). Dashed lines: chance; error bars: 95% confidence intervals for decoders trained and tested on the same epoch (leave-one-out cross-validation). See also Figures S1–S4.

In all participants, neural channels exhibited sustained delay-period activity tuned to upcoming direction and distinct from execution-period activity (Figure 1C). This preparatory tuning was prominent: across the neural population, >50% of active channels per participant were tuned to direction during the delay period (p-value < 0.01, ANOVA; Figure 1D). On no-go trials, delay-period activity decayed back to baseline after the cue to skip movement (Figure 1C), confirming it was tied to preparation and independent from execution itself.

To confirm that pre-execution neural activity reflected preparation rather than early movement attempts, we recorded residual muscle activity from participant T16’s hand and forearm using surface electromyography (EMG). EMG modulation only appeared during the execution period and was absent on no-go trials, confirming movements were attempted only when instructed (Figure S1). Additionally, to test whether preparatory activity was driven by visual stimuli or eye movements resulting from the cued targets, we ran a reduced-saliency variant of the task in which direction was cued by an arrow at the screen center while the participant held their gaze fixed (Figure S2). Even without pronounced eye movements, preparatory activity persisted and was consistent with that from the standard task, confirming it was intrinsically tied to movement intention.

At the neural population level, demixed principal component analysis (dPCA)³⁸ showed that preparatory activity for each direction became distinct from baseline over time, with cross-condition structure that matched the targets’ spatial layout (Figure 1E). Furthermore, we trained support vector machine (SVM) classifiers to predict movement direction from single-trial preparatory population activity. Cross-validated accuracies ranged from 48% to 88% (chance: 14%; Figure 1F) – notable given only pre-movement activity was used. Time-resolved decoding further revealed that directional information emerged early after target onset and remained stable up to the go cue (Figure S3A). The variation in classification accuracy across participants mirrored differences in preparatory signal-to-noise (Figure S4).

When we extended the single-trial decoding analysis beyond the delay period, accuracy approached 100% during movement execution, as expected, and fell near chance when no movement was attempted (Figure 1G). Prior non-human primate work suggests that movement-related activity occupies a subspace of neural population activity orthogonal to preparatory activity^8,9. To test whether this orthogonality also holds in humans, we trained decoders on activity from one epoch and tested them on the other. Contrary to expectations, cross-epoch decoders generalized well above chance, using either the whole epoch trajectory (accuracies: 45–80% for T11, 24–42% for T16, 26–29% for T5; Figure 1G) or time-resolved neural features (Figure S3B). Similarly, principal component analysis (PCA) revealed partially overlapping low-dimensional subspaces occupied by preparatory and movement activity (alignment indices of 33–59%; Figure S3C), with preparatory activity engaging dimensions that are subsequently expanded or reshaped during execution (Figure S3D). Together, these results confirm preparatory activity is a characteristic feature of human motor cortex, with strong directional tuning that partially overlaps with movement-related activity rather than being fully distinct as prior work predicted.

Preparatory activity exhibits effector-independent directional tuning

We next asked whether preparatory directional tuning is tied to the specific effector used, or instead reflects effector-agnostic intention. Both would be useful for BCIs: the first would enable prediction of which body part a user plans to move; the latter, robust direction decoding that generalizes across effectors. During movement execution, hand-knob activity represents movement compositionally, with some components tied to specific effectors and others preserved across different effectors³⁹. However, whether this compositionality holds for movement preparation, a distinct process that precedes effector-specific muscle output, is uncertain.

To evaluate effector dependence of preparatory tuning, we asked participants T16 and T11 to perform the radial-8 task using attempted finger movements (small finger for T16, index finger for T11) or wrist movements (Figure 2A). As before, attempted movements produced at most slight deviations of participants’ wrist and fingers. Nevertheless, EMG recordings from participant T16 confirmed that muscle activation patterns differed across effectors (Figure 2B).

For individual neural channels, preparatory activity ranged from effector-invariant to strongly effector-dependent (Figure 2C). At the population level, delay-period activity for matched directions was positively correlated across effectors (mean off-diagonal correlations of 0.94 for T16 and 0.42 for T11; Figure 2D). This correlation demonstrates that some component of preparatory activity is effector-independent – related to abstract task features rather than the specific attempted physical movement.

To quantify the degree of effector dependence, we used dPCA to linearly decompose preparatory activity into effector-dependent, effector-independent, and condition-independent components. For both participants, effector-independent components were consistently tuned to direction, alongside distinct effector-dependent components separating wrist and finger movements (Figure 2E). Their relative magnitudes differed across participants: for T16, effector-independent components accounted for more variance than effector-dependent components, whereas the opposite was found for T11. Accordingly, PCA subspaces for each effector overlapped partially, with greater alignment and more consistent cross-subspace trajectories for T16 than T11 (alignment indices of 45% vs 35%; Figure S5). These across-participant differences may reflect variation in the prepared movements (e.g., more wrist involvement in T16’s finger imagery), or in other factors such as the particular recording location within the hand-knob area.

Given the prominent effector-independent preparatory activity, it was unclear whether direction and effector identity could simultaneously be distinguished from preparatory activity, which would be particularly important for BCI control that requires both. In this regard, single-trial joint decoding accuracies ranged from 36% to 78% (chance: 7%; Figure 2F), indicating that preparatory activity is unique to each effector-direction combination. This specificity aligns with the theorized role of preparatory activity setting a movement-specific initial neural state^14,15.

Preparatory activity is tuned to both the initial and target directions of curved trajectories

Given that the attempted movements were all straight in the previous sections, we could not distinguish whether directional tuning was tied to the initial movement direction or the trajectory endpoint direction (“target” direction). To address this ambiguity, we asked participants T16 and T11 to perform a new task that cued movements with various straight and curved trajectories. Trajectories were cued via on-screen barriers and targets, with initial and target directions diverging by up to 90° in 45° curvature increments (Figure 3A). This design yielded 40 movement conditions (8 targets × 5 trajectories towards each target), allowing initial direction, target direction, and curvature to be dissociated – though, by construction, any two of these determine the third. As before, this task followed an instructed-delay paradigm where a random subset of trials skipped the execution period (Figure 3B).

Figure 3. — A. Example layouts cueing movements with different directions and curvatures. B. Trial structure and timing for the instructed-delay curved-movement task. C. Peri-stimulus time histograms for delay period activity from example channels, computed across initial directions (top), target directions (middle), or curvatures (bottom). Shaded regions: +/− standard error of the mean. D. Percentage of active channels per participant tuned to either the initial direction, target direction, or curvature (p-value < 0.01, 2-way ANOVA main effects, Benjamini-Hochberg corrected). E. Delay-period population activity (averaged across trials and time) for all conditions (8 directions × 5 curvatures), visualized via PCA (top 3 components); markers denote curvature. Top: projections for an example initial direction and all corresponding curvatures; middle: same but for an example target direction; bottom: one example curvature (0°) and all corresponding initial directions. Dotted lines connect conditions, with solid markers indicating condition averages. **F-H**. Same as E for both participants, but with all conditions colored by initial direction (F), target direction (G) or curvature (H). I. Single-trial preparatory decoding accuracy for initial movement direction, target direction, and curvature, using SVM classifiers (leave-one-out cross-validation). Error bars: 95% confidence interval; dashed lines: chance. See also Figure S6.

Individual channels showed distinct preparatory tuning to the initial and target directions; for some, activity was only discernible across different initial directions, whereas others were strongly modulated by target direction (Figure 3C). A subset of channels also showed tuning to movement curvature. Across the population, 49–66% of active channels per participant were tuned to initial movement direction, versus 10–28% for target direction and 8–36% for curvature (p-value < 0.01, 2-way ANOVA main effects, Benjamini-Hochberg corrected; Figure 3D), indicating joint encoding of multiple trajectory features, strongest for initial direction. At the population level, we projected activity from all 40 conditions onto the top delay-period principal component dimensions (Figure 3E). For both participants, activity was organized by initial direction even across different curvatures (Figure 3F), and was less structured when grouped by target direction across curvatures or curvature across initial directions (Figure 3G, H). Although initial direction tuning dominated the top preparatory dimensions, movements with the same initial directions but different curvatures recruited neural dimensions that were partly distinct from one another – more so for T16 than T11 (Figure S6).

Using SVM classifiers, all three features were decodable above chance (chance: 14% for direction, 22% for curvature), with significantly higher accuracy for initial direction than target direction and curvature (T16: 51% > 37% and 39%, p-value = 4.4 × 10⁻⁷ and 1.2 × 10⁻⁵, respectively; T11: 51% > 22% and 34%, p-value = 2.1 × 10⁻²² and 1.4 × 10⁻⁸; two-proportion z-test; Figure 3I). Target direction and curvature showed comparable decoding accuracy despite reflecting distinct features of the upcoming trajectories. We note that intended directions were decoded with similar accuracy for T16 and T11 in this task, in contrast to the earlier straight-movement task where decoding from T11’s activity was more accurate. The basis for these differences is unclear; it may reflect participant differences in preparatory activity consistency between curved and straight movements, though day-to-day signal variations across sessions may also have contributed. Collectively, these analyses confirm preparatory activity encodes multiple features of curved trajectories, with stronger tuning to initial direction, and some target-direction and curvature-specific structure.

Preparatory activity is slightly tuned to upcoming movement distance and speed

Decoding movement extent or distance is critical for practical BCI applications; for example, to enable control that predicts a user’s intended target in advance. However, an important consideration is that distance is typically correlated with speed^40,41. Thus distance and speed tuning have been difficult to tease apart in non-human primate studies²², with only limited evidence of independent tuning²⁰.

To dissociate preparatory tuning to distance and speed, we developed a new task (inspired by prior work⁴²) that explicitly cued spatiotemporal trajectories. In each trial, a dotted path indicated the upcoming position of a “Pac-Man” icon through time (horizontal and vertical axes respectively), with independently defined direction, distance, and speed (Figure 4A). Consistent with an instructed-delay paradigm, dots were stationary at first and then scrolled downwards to cue movement initiation (Figure 4B). Movements were drawn from 18 conditions: 2 directions (right or left), 3 distances (short, medium, long), and 3 speeds (slow, medium, fast).

Individual neural channels were selectively tuned to each feature (Figure 4C). As before, a substantial percentage of active channels – 40–43% across participants – were tuned to direction (p-value < 0.05, 3-way ANOVA, Benjamini-Hochberg corrected; Figure 4D). The proportion was lower for distance or speed: <4% for main effects, or 8–12% including interactions across features (e.g. distance-direction and speed-direction interactions). dPCA mirrored this pattern: direction and condition-independent components dominated preparatory variance (9–18% and 8–29% respectively), while distance and speed components, though modulated, only accounted for ~1% each (Figure 4E, F; Figure S7). Additionally, 6–7% of the population activity variance was explained by direction-distance-speed interactions, suggesting integrated preparatory encoding of these features.

To test single-trial encoding robustness, we fit SVM classifiers to separate preparatory activity for short vs long distances, slow vs fast speeds, and both directions (chance: 51% for all features). As expected, decoders reliably separated direction (85–95% accuracy; Figure 4G), but performance for distance and speed varied across participants: T16 decoding was above-chance for distance (65% accuracy) but not speed, whereas T11 showed the opposite (62% accuracy for speed). Low preparatory decoding accuracy was not due to the subsequent movement attempts being indistinguishable within each condition: execution period decoding was well above chance for all conditions in both participants (Figure 4H).

Time-resolved decoding revealed a richer picture: unlike the direction-only task (Figure S3A), decodability was sometimes restricted to specific windows during preparation and occasionally exceeded chance where the full-trajectory decoder could not (Figure S8). This result suggests a degree of transient preparatory encoding is obscured by the simple decoding architecture we used that concatenates activity across the full delay window. Altogether, human preparatory activity carries information about upcoming distance and speed in addition to direction, though its reliability varies by participant – possibly reflecting variability in electrode array placement or participants’ movement preparation strategy.

Online decoding from preparatory activity enables rapid and fluid brain-computer interface control

Since human preparatory activity is informative about upcoming movement, an online BCI should be able to decode a user’s intended movement trajectory from preparatory activity alone. Unlike standard paradigms that map neural activity during movement onto instantaneous control commands (Figure 5A), a preparatory control paradigm could precompute and execute trajectories as soon as movement intention is detected. Along these lines, previous non-human primates work shows this type of control can enable fast and accurate BCIs^30–32.

As a proof-of-concept, we built a preparatory control paradigm with three components (Figure 5B; see STAR Methods). A movement initiation decoder continuously read incoming neural features and decoded the probability of a “movement” state, with initiation taken as when this probability crossed a predetermined threshold. Threshold crossings triggered a movement direction decoder that predicted the intended direction from a window of preparatory activity preceding the detected onset. An autonomous movement generator then executed a quick cursor movement to the decoded goal position at a fixed distance in the decoded direction. Notably, unlike some prior non-human primate preparatory BCIs^30,31, this paradigm is user-initiated: movements were only triggered when intended by the user – allowing variable timing – and the decoded direction could take any continuous value rather than being restricted to discrete target positions^30–32.

To evaluate online performance, participants performed a modified radial-8 task in which movement initiation could be triggered before or after the go cue. Thus, the window of preparatory activity used by the movement direction decoder was independent of task timing. Additionally, movement direction decoding could predict endpoints outside of the cued targets. For T16, we tested movement initiation decoding accuracy by omitting the go cue on some trials, and trials failed if initiation was decoded outside the cued execution period. T11 performed a more flexible task variant: he was asked to match movement initiation timing to a delay-period timer, but moving early did not fail trials. Because early movements were not penalized, he could effectively skip the delay period, as would be the case in typical cursor control applications. Accordingly, T11 had a short 200-ms preparatory window used for decoding, versus 800 ms for T16.

Participants successfully used the preparatory control paradigm, achieving responsive control with the ability to perform arbitrarily quick cursor movements at rates comparable to able-bodied movement (400–600 ms; Figure 5C)⁴⁰. Decoded onset times were ~400 ms after the go cue for T16 and ~600 ms after the target cue for T11 (Figure 5D), reasonable reaction time values (plus extra preparation time for T11 given his instruction to match the delayed go cue). Residual EMG from T16’s hand confirmed that the inferred movement onset timing was synchronized with actual intended movement (Figure S9). In terms of endpoint accuracy, success rates – the percentages of trials ending within a radius of 30% of the start-to-target distance – were 45% for T16 and 56% for T11 (Figure 5E). Nonetheless, decoded directions clustered around the cued targets, with median angular errors of 19° for T16 and 16° for T11 (Figure 5F). Although success rates were lower than the near-100% performance of standard moment-by-moment decoders, these results are promising for a simple implementation that relies exclusively on pre-movement predictions and cannot correct errors on the fly.

We compared performance against optimal linear estimation (OLE), a common BCI approach that continuously predicts intended movement velocity^43–45. For successful trials, the preparatory control paradigm reached targets substantially faster than the OLE (mean time-to-target of 0.90 < 1.28 s and 0.97 < 1.38 s for T16 and T11, respectively; p-values = 9.9 × 10⁻⁹ and 3.3 × 10⁻⁷, Mann-Whitney U test; Figure 5G). Moreover, the cursor-to-target distance decreased much more consistently for the preparatory control paradigm (Figure 5H): by design, the auto-generated trajectories approached the decoded target monotonically, whereas under OLE control the distance often increased before the target was reached. These advantages held for T11 even with his encouraged pre-movement delay. This proof of concept highlights the speed and consistency benefits of control based on anticipated predictions from preparatory activity, at the expense of a lower (though still substantially above chance) success rate given the inability to correct movements once initiated.

Finally, because preparatory and execution-related activity have partially non-overlapping tuning that may provide complementary information (Figure S3), we explored the potential benefits of combining both windows within the presented paradigm. To this end, we quantified how performance changed as progressively longer windows of post-onset execution activity were incorporated (Figure 5I). Unsurprisingly, performance improved – though at the cost of latency, since predictions become delayed by the duration of the accumulated window – peaking ~600 ms after intended onset. Importantly, including preparatory activity still improved accuracy relative to decoders relying on execution activity alone. Overall, these results demonstrate the benefits of leveraging preparatory activity for BCI control, both on its own and in combination with execution-related activity.

Discussion

Preparatory activity in the human motor cortex is a rich, underexplored source of information for neurally controlled assistive devices. In three clinical trial participants with tetraplegia, we demonstrate that preparatory activity encodes many features of upcoming movements: direction, effector, trajectory curvature, and, to some extent, distance and speed. Furthermore, it was robust enough to enable online BCI control, here yielding substantially faster cursor control than standard algorithms.

Delimiting subregions in human motor cortex tied to motor preparation

In non-human primates, subdivisions of motor cortex are relatively well established functionally and cytoarchitectonically^46–48, with dorsal premotor cortex (PMd) most prominently associated with preparation^5,21,7 and primary motor cortex (M1) with movement execution^49,50,7. These boundaries are harder to define in humans: the putative M1/PMd transition varies considerably across individuals^51–53 and has only recently been assigned subject-specific labels through individualized Human Connectome Project (HCP) parcellation⁵⁴. In this work, we recorded from the hand-knob area of dorsal precentral gyrus³⁶, labeled as Area 6d by HCP parcellation^37,54. This region is often equated with caudal PMd in non-human primates^47,46,55,39 with which it shares functional properties such as preparatory and cross-effector tuning^56,39. However, its pronounced execution-related and relatively weaker preparatory activity suggest that human Area 6d may not be an exact analog of non-human primate PMd. Future work could study whether other regions (e.g., rostral to the hand-knob) show stronger preparatory tuning in line with PMd in non-human primates.

Comparison with non-human primate studies

The human preparatory activity we observed shared several key features with that of non-human primates (NHPs). As in NHPs^4,21,7–9, it was distinct from execution activity and strongly direction-tuned (Figure 1). It was also tuned to movement magnitude (distance or speed; Figure 4), consistent with prior findings^22,20. These correspondences indicate that core properties of preparatory activity established in NHPs are largely preserved in humans. Our findings also diverged in notable ways. Most prominently, the neural dimensions engaged by human preparatory and execution-related activity were partially overlapping (Figure S3), unlike the orthogonality consistently demonstrated in NHPs^8,9,6. This discrepancy might stem from a functional mismatch between the subregion we targeted (area 6d) and its PMd analog in NHPs (see above). Alternatively, it could reflect the lack of overt movement: a substantial component of execution-period activity may depend on sensory feedback⁵⁷, absent for our participants attempting non-overt movements. We also observed comparatively weaker preparatory encoding of distance and speed, which we discuss below.

Strength of preparatory activity tuning to distance and speed compared to direction

Our results show that preparatory activity in human participants encodes distance and speed, but these features accounted for only a small fraction of the total preparatory modulation (Figure 4). This weak tuning is at odds with non-human primate recordings which show stronger encoding of distance and speed^20,22. One key consideration is that our participants received no proprioceptive or visual feedback given their lack of overt movement. Without feedback, greater variability is expected during movement execution and its associated preparation. Separately, movements of different distances and speeds are both tied to grading in muscle activation intensity, which may be perceived similarly during preparation. In contrast, movements towards different directions recruit distinct sets of muscles, so their motor plans may be easier to distinguish. Lastly, and related to the functional subregion question raised earlier, preparatory distance and speed tuning might be more prominent in other cortical regions outside the hand-knob area.

Preparatory activity reflects multiple levels of movement abstraction

Our results indicate preparatory activity in the human motor cortex encodes movement features at multiple levels of abstraction. At a higher level, directional tuning was preserved across effectors (Figure 2) – consistent with the motor equivalence long reported in behavioral and neuroimaging studies^58,59, here extended to preparation – and target direction tuning was preserved across curved trajectories (Figure 3). Preparatory activity was also comparable across tasks differing in visual saliency (Figure S2), indicating this abstract, task-related tuning was not merely a representation of visual features. At the same time, a substantial part of preparatory activity was linked to lower-level features closely related to motor execution, such as effector identity and initial movement direction. Together, these findings extend non-human primate evidence that preparatory activity encodes abstract task features such as target position independent of physical movement direction⁶⁰. The role of these multi-level preparatory representations is an open question, but could potentially serve to anticipate neural computations that transform abstract features from upstream brain areas into muscle-like motor commands³⁹.

Rich preparatory encoding as a mechanism for optimal movement generation

The diverse encoding we observed is likely a limited view of a more complex, less interpretable neural state arising during preparation. Within the framework of computation through neural population dynamics⁶¹, preparatory activity can be seen as an optimal initial condition for the generation of upcoming movement^14,15. Our findings are consistent with this view: the distinct preparatory activity for diverse movement features can be interpreted as the corresponding initial neural state for each condition, such that each differently prepared trajectory has its own preparatory state. For example, because curved trajectories depend on their initial and target directions, both are encoded in preparatory activity (Figure 3). Further investigation requires scaling task complexity to an even greater diversity of movements and, ideally, increasing recording channels to capture the information encompassed by an expanding set of movement conditions.

Implications for brain-computer interface control

In this work we also showed, for the first time with clinical trial participants, that anticipated predictions of intended movement decoded from preparatory activity can effectively drive BCI cursor control (Figure 5). Even with a deliberately simple implementation – which ignored all information from execution-period activity – our preparatory paradigm enabled substantially faster control than standard methods (Figure 5G, H). The preparatory decoding paradigm was also unaffected by moment-by-moment variability in neural activity during execution, and thus may prove to be less cognitively demanding for users.

How best to combine preparatory and execution-related activity remains an open question. Extending the decoding window into the movement period would reduce directional errors at the expense of increased latency (Figure 5I); alternatively, given that preparatory and execution activity evolve in partially distinct subspaces with complementary decoding performance (Figure S3, Figure 5I), our control paradigm could be extended by seeding trajectories with preparatory predictions and then correcting them on the fly using execution-specific activity. Notably, our paradigm relied on an explicit, fixed-length preparatory period, but naturalistic computer use does not impose one. In this regard, non-human primate work shows preparatory activity arises even without an explicit delay (if only for a couple hundred milliseconds)¹¹ and, accordingly, T11’s effective control using a 200 ms preparatory window (Figure 5) suggests the presented approach should extend to unconstrained settings.

Our characterization of preparatory activity across different tasks also highlights future challenges for preparatory control paradigms. The low signal-to-noise ratio in preparatory activity (Figure S4) can yield inaccurate predictions; though these could be mitigated by online error detection and correction^62,63, or by incorporating execution-related signals (e.g., pairing preparatory control with OLE). Additionally, our analyses suggest that decoding errors from preparatory activity will decrease substantially at higher channel counts (Figure S10), implying that higher-resolution recordings should improve the viability of this preparatory control. Separately, movement distance encoding might not prove robust enough for pre-programming trajectories to arbitrary points in space. As an alternative, a paradigm that generates fast movements towards a decoded direction and detects the transient intention to stop at a given time^64,65 may prove more feasible. Further exploration of these control schemes could be crucial to closing the gap between current BCI control and able-bodied movement.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Chethan Pandarinath (chethan.pandarinath@emory.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

All data required to reproduce the findings in this study are publicly on Dryad at https://doi.org/10.5061/dryad.brv15dvr9 as of the date of publication.
All original code has been deposited at https://github.com/snel-repo/human-preparatory-activity-paper and is publicly available as of the date of publication.
Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

STAR★Methods

Experimental model and study participant details

This study includes data from three participants, referred to as T16, T11, and T5, who each gave informed consent and were enrolled in the BrainGate2 Neural Interface System clinical trial (ClinicalTrials.gov Identifier: NCT00912041, registered June 3, 2009). This pilot clinical trial was approved under an Investigational Device Exemption (IDE) by the US Food and Drug Administration (Investigational Device Exemption #G090003; CAUTION: Investigational device. Limited by federal law to investigational use), as well as granted approval by the Emory University IRB (protocol #00003070), the Mass General Brigham IRB (protocol #2009P000505), the Providence VA Medical Center, and the Stanford University IRB (protocol #20804). All research was performed in accordance with relevant guidelines and protocols.

T16 is a right-handed woman, 52 years of age at the time of data collection for this study, with tetraplegia and dysarthria due to a pontine stroke which occurred approximately 19 years prior to enrollment in the BrainGate2 clinical trial. T16 had four 64-channel intracortical microelectrode arrays (1.5 mm electrode length; NeuroPort Electrode; Blackrock Neurotech, Salt Lake City, Utah, USA) placed in her left precentral gyrus, guided by individualized Human Connectome Project (HCP) cortical parcellation: two in HCP-identified hand knob area (6d), one in HCP-identified ventral premotor cortex (6v), and one on the border of the HCP-identified premotor eye fields (PEF) and speech-related area 55b. T16 is able to speak slowly and quietly, but enunciation is restricted by limited face and mouth movement. She has limited voluntary control of her upper extremities, with some shoulder motion and some slow and contractured wrist and finger movements. She has limited to no voluntary control of her lower extremities. T16’s sensation is fully intact. Data for this study was collected with T16 between post-implant days 253–543.

T11 is a right-handed man, 38 years of age at the time of data collection for this study, with tetraplegia due to a cervical spinal cord injury (classified as C4 AIS-B) which occurred approximately 11 years prior to enrollment in the BrainGate2 clinical trial. T11 had two 96-channel intracortical microelectrode arrays (1.5 mm electrode length; NeuroPort Electrode) placed in his left dorsal precentral gyrus, targeting the hand knob area (6d) as identified anatomically by preoperative magnetic resonance imaging (MRI). T11 has full movement of his face and head with limited voluntary motion of the arms. Data for this study was collected with T11 between post-implant days 1321–1961.

T5 was a right-handed man, 70 years of age at the time of data collection for this study, with tetraplegia due to cervical spinal cord injury (classified as C4 AIS-C) which occurred approximately 9 years prior to enrollment in the BrainGate2 clinical trial. T5 had two 96-channel intracortical microelectrode arrays (1.5 mm electrode length; NeuroPort Electrode) placed in his left dorsal precentral gyrus, targeting the hand knob area (6d) as identified anatomically by preoperative MRI. T5 had full movement of his face and head and the ability to shrug his shoulders. Below the level of spinal cord injury, T5 had limited voluntary motion of the legs and arms. Data for this study was collected with T5 during post-implant day 2623.

Method details

Research sessions and data collection

For each participant, data was collected during research “sessions” that took place across different days. All research sessions were performed at the participant’s place of residence. During each session, the participant was positioned at an idle position in front of a computer monitor (screen size: between 24 and 24.5 inches diagonal, 1920 × 1080 pixels), and asked to perform different motor tasks in a series of “blocks” lasting 3–6 minutes each. Table S1 lists all research sessions and number of blocks during which data was collected with each participant.

Data collection was run on a dedicated computer rig. Execution and communication of modular Python and C++ processes for running tasks and synching with neural recordings were handled using the Backend for Realtime Asynchronous Neural Decoding system (BRAND)⁶⁶.

Intracortical neural signal acquisition and processing

Intracortical electrical voltages were recorded from microelectrode arrays using NeuroPort Neural Signal Processor systems (Blackrock Neurotech). Recorded signals were first analog band-pass filtered (0.3 Hz to 7.5 kHz, 4th order Butterworth filter) and digitized at 30 kHz (250 nV resolution). Linear regression referencing (LRR)⁶⁷ was then applied, with electrode-specific LRR coefficients determined for each block using the filtered 30 kHz data either from the same block for offline analyses (Figures 1–4, S2–S8), or from data collected during the previous block for online decoding results (Figures 5, S9, S10). For offline analyses, the resulting voltage time series were then digitally band-pass filtered (250 Hz to 5000 Hz, acausal 4th order Butterworth filter). For online decoding results, voltage time series were digitally high-pass filtered (>250 Hz, acausal 4th order Butterworth filter with 4 ms delay).

After rereferencing and filtering, threshold crossings were computed for each neural channel. Thresholds were set to −4.5 times the root-mean-square of the voltage signal for each channel (except for the 2023–05-07 and 2023–06-18 sessions with participant T11, where the multiplier was −3.5), computed on the same block for offline analyses or from a previous block for online decoding. Threshold crossings were then grouped into bins with a duration of 10 ms (2023–05-07 and 2023–06-18 sessions with participant T11, Figure 5) or 20 ms (all other results). Spike-band power (SBP) was also computed for each channel, by taking the mean of squared values of the rereferenced and filtered 30 kHz samples in each time bin. Only channels from arrays in the hand-knob area of participants’ precentral gyrus (128 channels for T16, 192 channels for T11 and T5) with a mean firing rate of at least 2 Hz across all time points and blocks for each particular task (denoted “active” channels) were considered for analysis or online decoding. Binned threshold crossings and spike-band power for active channels were then appended to form a vector of “neural features” for each time bin. Finally, neural features were smoothed acausally using a Gaussian kernel with σ = 100 ms for offline analyses or causally using exponential smoothing for online decoding (smoothing parameter varied across sessions, see below).

For neural population analyses (all analyses except peri-stimulus time histograms and single-channel tuning), binned neural features were z-scored (each channel’s activity was baseline-centered and divided by its standard deviation) to prevent channels with high magnitude values from dominating the variance of the data. For offline analyses, the baseline subtracted from each channel’s values was computed on a per-block basis (to account for baseline shifts across the session) and the standard deviation was computed across all blocks for a particular task (after baseline-centering). For each block of online decoding, the baseline and standard deviation for z-scoring was computed on the previous block.

Electromyography activity acquisition

Electromyography activity for participant T16 was recorded using a Trigno Duo sensor (Delsys Inc., Natick, Massachusetts, USA). Voltage was recorded from electrodes placed on the ulnar surface of her right hand and dorsal surface of her right forearm (2 channels total). Recorded signals were first analog band-pass filtered (100 Hz to 850 Hz, Butterworth filter with 2-pole high-pass corner and 4-pole low-pass corner) and digitized at 2148 Hz (168 nV resolution). The resulting voltage time series were then digitally band-pass filtered (100 Hz to 500 Hz, acausal 4th order Butterworth filter), squared to compute the signal power, and digitally low-pass filtered (<5 Hz, acausal 4th order Butterworth filter). To align with the neural spiking data, the processed EMG activity samples were subsequently resampled to 100 Hz (through linear interpolation) and grouped into 20 ms bins (by averaging across samples).

For plotting EMG activity (Figures 2, S1, S9), each channel’s samples were baseline-centered and divided by their standard deviation. The baseline subtracted from each channel’s values was computed on a per-block basis and the standard deviation was computed across all blocks for a particular task (after baseline-centering).

Eye tracking data collection

Gaze position data for participant T16 was collected using a Tobii Pro Spark eye tracker (Tobii AB, Stockholm, Sweden). Data was collected at a sampling rate of 60 Hz. To align with the neural spiking data, eye tracking samples were resampled to 100 Hz (through linear interpolation) and grouped into 20 ms bins (by averaging across samples).

Instructed-delay tasks

During research sessions, participants performed tasks with an instructed-delay paradigm, structured as a sequence of “trials”. Each trial began with a “target cue” marking the start of a delay period, during which participants were instructed not to move while being visually cued on the target position and features of the upcoming movement. Some tasks also showed a circular timer during the delay period to indicate the remaining delay time (‘instructed-cued-delay’). The delay duration and cued movement condition were selected through pseudorandom sampling from a distribution promoting a balanced number of trials for each combination of these values within each block. The delay period was followed by an execution period indicated by a “go cue” where participants attempted to match quick on-screen cursor movements consistent with what was previously cued and then remained on the target position for a specified hold time. After the execution period, the go cue turned off to indicate an inter-trial period where participants returned to a rest state. For some tasks, a random subset of trials skipped the go cue (33% probability), such that the delay period was immediately followed by the inter-trial period preceding the next trial. Participants performed different instructed-delay tasks designed to cue different movement features (Table S1): “radial-8” and “radial-4” tasks to cue direction, “curved radial-8” task to cue separate initial and target directions, and “Pac-Man” task to cue direction, distance, and speed. Tasks could have the cursor move automatically during the execution period to perform the cued movement – in which case participants were instructed to track the cursor without visual feedback of their own attempted movement (“open-loop”) – or have cursor movements be controlled online from recorded neural activity (“closed-loop”). Possible values for the delay period duration, hold times, and inter-trial times for each task are indicated in Table S2.

For the open-loop instructed-delay radial-8 and radial-4 tasks, participants attempted movements to match on-screen cursor movements towards radially distributed targets. 8 or 4 targets were shown at a fixed distance (400 pixels) from the center of the screen with equal angular spacing. All targets were initially dark during the inter-trial period with the cursor centered on the screen. For each trial, a target was selected during the delay period and was highlighted in yellow. The target remained highlighted for the duration of the delay period for participants T16 and T11; for T5 the target turned dark again 200 ms into the delay period. During the execution period, the highlighted target turned red (or green in the case of T5) to signal the go cue. The cursor then moved automatically towards the target following a truncated Gaussian velocity profile, with movement durations of 1.1 s for T16, 0.8 s for T11, or 0.5 s for T5. Participants were instructed to match cursor movements using either their wrists (Figures 1, 2, S1–5) or fingers (small finger for participant T16 or index finger for participant T11; T5 did not perform the task using his finger; Figures 2, S5); since movement capabilities amounted to at most slight deviations of the wrist and fingers, verification that the correct movements were attempted was through coaching and verbal confirmation from the participants. Once the cursor reached the target, the target turned pink (or red in the case of T5). The trial then ended and the cursor returned to the center, followed by an inter-trial period that preceded the next trial.

An alternate version of the open-loop instructed-delay radial-4 task with suppressed visuals was also run with participant T16 (Figure S2). For this version of the task, movement direction was cued during the delay period by an arrow within a yellow circle at the center of the screen, and this circle turned red to indicate the go cue. No cursor was shown on screen during this task. T16 was also given the additional instruction to try to maintain her gaze fixated on the center of the screen as best as possible while performing this task, and an additional center target was shown during the inter-trial period to provide a gaze fixation point.

For the open-loop instructed-delay curved radial-8 task, participants used attempted movements of their wrists to match on-screen cursor movements following curved trajectories (Figure 3). As in the base radial-8 task, 8 targets were shown at a fixed distance (400 pixels) from the center of the screen with equal angular spacing. Targets were initially dark during the inter-trial period. For each trial, one of 8 possible targets and one of 5 possible trajectories to that target (where the initial trajectory direction diverged by up to 90 degrees in 45-degree increments) were selected. During the delay period, the target was highlighted in yellow and barriers were displayed halfway between the center of the screen and the target, with an opening in the direction of the initial trajectory direction. To signal the go cue, the highlighted target turned red and the cursor moved automatically towards the target following a curved trajectory that evaded the barriers, with a truncated Gaussian speed profile and duration that scaled with trajectory length (ranging from 1.2 s to 2.1 s for T16 or 0.8 s to 1.4 s for T11). The highlighted target turned pink when the cursor came in contact with it, after which the cursor returned to the center followed by an inter-trial period.

For the open-loop instructed-delay Pac-Man task, participants used attempted movements of their wrists to match side-to-side on-screen movements of a Pac-Man icon (Figure 4). The Pac-Man began each trial centered with no targets shown. During the delay period, the direction (right or left), target distance from the center (200, 400, or 600 pixels), and average speed (200, 400, or 600 pixels/s) of the upcoming movement were cued through a stationary trajectory of dots displayed on screen (which followed a truncated Gaussian velocity profile): position and time were represented by the horizontal and vertical axes, respectively. The target position was also highlighted by a gray vertical line and a gray circle in the horizontal Pac-Man axis. To signal the go cue, the dotted trajectory started scrolling downwards at a constant rate during the execution period and the Pac-Man automatically moved horizontally towards the target following the scrolling dots. After reaching the target and remaining in its position for a specified time, the trial ended and Pac-Man returned to the center position followed by an inter-trial period.

Participants also performed closed-loop variations of the radial-8 task where cursor movements were decoded online from neural activity either using a preparatory control paradigm or an optimal linear estimator (OLE) (Figures 5, S9, S10). For the preparatory control paradigm blocks with participant T16, the task followed the instructed-delay paradigm and trials could be failed if movement was initiated before the go cue or after a timeout of 3 s when the decoded movement ended off-target, in which case the task skipped to the inter-trial period. For the preparatory control paradigm blocks with participant T11, the task followed the instructed-cued-delay paradigm and early movements before the go cue were not failed. OLE blocks with participant T16 also followed the instructed-delay paradigm. OLE blocks with participant T11 did not have a delay period, so he was instructed to move the cursor towards the target immediately after it was cued. For both participants’ OLE blocks, the cursor position was not automatically recentered at the end of each trial, so trials alternated between outwards and inwards movements (only outwards trials were considered for analyses).

Peri-stimulus time histograms

Peri-stimulus histograms (PSTHs) for each channel during the different tasks were computed by averaging smoothed threshold crossings for all trials from each condition, aligned either to their respective target cue for delay period activity (Figures 1–4, S2) or go cue for execution period activity (Figure 1). PSTHs of the delay period activity only considered trials with a delay duration of at least 1000 ms, meaning that trials from participant T5 with delay durations of 200 ms and 600 ms were excluded from analysis.

Single-channel tuning analyses

For each task, channels with preparatory activity tuned to different movement features were identified using ANOVA. Specifically, separate ANOVA tests were performed on the delay period activity from each channel during each task. For all ANOVA tests, the dependent variable was the average single-channel firing rate during the delay period: between 400 and 1000 ms after the target cue over all trials for T16 and T11 and between 400 and 800 ms after the target cue for trials with a delay duration of at least 1000 ms for T5. For the radial-8 task data, 1-way ANOVAs were performed to identify channels where preparatory activity was affected by the upcoming movement direction (p-values < 0.01). For the curved radial-8 task data, 2-way ANOVAs were performed to identify channels where preparatory activity was affected by either the initial direction, target direction, or curvature of the upcoming curved movement (main effect p-values < 0.01). For the Pac-Man task data, 3-way ANOVAs were performed to identify channels where preparatory activity was affected by either the direction, distance, or speed of upcoming movement (separately reported main effect and interactions p-values < 0.05). For the 2-way and 3-way ANOVAs, p-values were corrected for multiple comparisons using the Benjamini–Hochberg procedure⁶⁸.

Demixed principal component analysis

Demixed principal component analysis (dPCA)³⁸ was used to decompose preparatory population activity from the different tasks onto distinct components specific to each of the different movement features (or independent of all features). dPCA was performed using publicly available code: https://github.com/machenslab/dPCA. The built-in procedure to determine the best regularization term to prevent overfitting (10-fold cross-validation) was used for the data from each task. For the single-effector radial-8 task data (Figure 1), the features/marginalizations of interest were time (condition-independent) and direction (condition-dependent). For the multi-effector radial-8 task data (Figure 2), the features/marginalizations of interest were time (condition-independent), direction (effector-independent), effector identity (effector-dependent), and direction-effector interactions (also effector-dependent). For the Pac-Man task data (Figure 4), the features/marginalizations of interest were time (condition-independent), direction, distance, speed, and direction-distance-speed interactions. For the data from each task and participant, dPCA projection matrices were fit on activity in the window from 0 ms to 1000 ms after the target cue, except for radial-8 data with T5, where dPCA was fit over the window from 0 ms to 800 ms after the target cue for trials with a delay duration of at least 1000 ms.

Cross-validated movement feature decoding

Single-trial movement feature decoding from preparatory activity during the open-loop tasks (Figures 1–4, S2) was computed using cross-validated support vector machine (SVM) classifiers with a linear kernel. The input to the SVM for each trial was a neural feature vector constructed by concatenating the neural feature values (threshold crossings and SBP) across all time bins in the window from 400 ms to 1000 ms after the target cue (30 bins). For participant T5, only trials with a delay duration of at least 1000 ms were used for this decoding analysis, and only neural features up to 800 ms after the target cue (20 bins) were concatenated to construct the input feature vector. Classifiers were trained through leave-one-out cross-validation: for each fold, all trials except one were included for training and performance was evaluated on the held-out trial. To choose the SVM regularization parameters for each fold, a nested 10-fold cross-validation was performed within the training data. Decoding performance was then computed as the mean accuracy on held-out trials across all folds and chance was computed as the frequency of the most frequent class label in the dataset from which test trials were sampled. Single-trial movement feature decoding from execution period activity (Figure 4) followed the same procedure, but with the input to the SVM for each trial being a neural feature vector constructed by concatenating the neural feature values across all time bins in the window from 400 to 1000 ms after the go cue (30 bins).

For the cross-epoch decoding generalization analysis (Figure 1), linear SVMs were again used but each trial’s neural feature vector was computed as the mean over time bins within the corresponding epoch window, rather than the concatenation used above. Three epochs were considered: delay period (same window as for the main preparatory activity decoding analysis mentioned above), movement period (400 to 1000 ms after the go cue for T16 and T11, 200 to 600 ms after the go cue for T5; go trials only), and no-movement period (same windows as for the movement period, but no-go trials only). For within-epoch decoder training and evaluation, the same leave-one-out cross-validation described above was used. For cross-epoch entries, a single classifier was trained on all trials from the source epoch and evaluated once on all trials from the target epoch, with the same nested 10-fold cross-validation used within the training data to select the regularization parameter; source and target epochs are non-overlapping so no held-out trial splits are required.

For the time-resolved same-sample and cross-sample decoding analyses (Figures S3, S8), linear SVMs were trained on neural features averaged over 100 ms time windows aligned to either the target cue or the go cue, using the same nested 10-fold cross-validation to select the regularization parameter. Classifiers were evaluated using 10-fold cross-validated trial splits, either on the same time window (to test tuning at each time point) or across all time windows (to test tuning generalization across time).

Decoding analyses were performed separately on data from the different tasks. All available trials were used for decoding analyses on data from the radial-8, radial-4, and curved radial-8 tasks (Figures 1–3, S2). For data from the Pac-Man task (Figure 4), direction decoding used all available trials. For distance decoding, only trials with values corresponding to either short or long distances were considered (trials with a distance of 400 pixels were excluded). For speed decoding, only trials with values corresponding to slow or fast speeds were considered (trials with a speed of 400 pixels/s were excluded).

Neural subspace alignment index and visualization

The overlap between neural subspaces occupied by different subsets of neural population activity was quantified using the alignment index metric⁹. For each data subset, condition-averaged neural features were computed and the condition-independent component (mean across all conditions at each time point) was subtracted. Principal component analysis (PCA) was then applied over each resulting dataset, and the top-10 principal components defined the corresponding neural subspace. The alignment index is then defined as the fraction of the total variance in one dataset that is accounted for by its projection onto the other dataset’s corresponding subspace. An alignment index of 1 indicates fully overlapping subspaces, and an index of 0 indicates orthogonal subspaces.

The alignment index was computed between delay period and movement period activity for movements in different directions (Figure S3), for delay period activity between wrist and finger movements (Figure S5), and pairwise across the delay period activity for all curvature conditions from the curved movement trials (Figure S6). Delay period subspaces were computed over activity spanning 400 to 1000 ms after the target cue for T16 and T11, or 400 to 800 ms after the target cue for trials with a delay of at least 800 ms for T5. Movement period subspaces were computed over activity spanning 400 to 1000 ms after the go cue for T16 and T11, or 200 to 600 ms after the go cue for T5.

The neural subspace overlap was also visualized by projecting the condition-averaged population trajectories onto the top-2 principal components of subspaces identified above (Figures S3, S5).

Cross-validated pairwise neural Mahalanobis distances

Pairwise discriminability between preparatory population activity for different conditions (Figure S4) was quantified using the cross-validated Mahalanobis (crossnobis) distance estimator^69,70, which yields unbiased estimates with an expected value of zero when two conditions are identical. Preparatory activity vectors used to compute pairwise crossnobis distances were obtained by averaging the normalized neural features (threshold crossings and SBP) across time in the window from 400 ms to 1000 ms (T11 and T16) or 400 ms to 800 ms (T5) after the target cue, with the mean activity across conditions subtracted prior to time-averaging. For execution activity vectors, windows from 400 ms to 1000 ms (T11 and T16) or 200 ms to 600 ms (T5) after the go cue were used instead.

Cross-validated pairwise neural correlations

Pairwise correlations between preparatory activity for movements with different directions and/or effectors (Figure 2) were computed using a cross-validated estimator that reduces bias³⁹. This estimator can result in values with magnitude greater than 1, which should be interpreted as the true correlation magnitude being close to 1. Preparatory activity vectors used to compute pairwise correlations were obtained by averaging the normalized neural features (threshold crossings and SBP) across time in the window from 400 ms to 1000 ms after the target cue. The average activity across directions within each effector condition was subtracted prior to time-averaging and computing correlations, to compare directional tuning within the same effector-agnostic reference frame.

Visualization of population activity across curvature conditions using principal component analysis

Preparatory population activity during the curved radial-8 task was visualized by using principal component analysis to compute projections onto the top-3 dimensions of activity that explained the highest amount of variance during the delay period across all conditions (Figure 3). For each of the 40 conditions (all valid combinations of initial directions and curved movement trajectories), neural features (threshold crossings and SBP) were trial-averaged and had the mean activity across conditions subtracted. The resulting trial-averaged neural features were then averaged across time for the window from 400 ms to 1000 ms after the target cue, and concatenated to create a matrix with rows equal to the number of neural features and columns equal to the number of conditions. PCA was then applied on this matrix to compute an orthogonal linear transformation that projects the high-dimensional neural features onto a 3-dimensional space.

Preparatory control paradigm for online cursor control

The first component of the preparatory cursor control paradigm was a movement initiation decoder. This decoder continuously read a stream of incoming normalized neural features (threshold crossings and SBP for participant T16, only threshold crossings for participant T11) and exponentially smoothed them (time constant of 170 ms for T16, 144 ms for T11). A linear discriminant analysis (LDA) classifier then computed the probability that these neural features corresponded to either a “movement” or “no movement” state. During online control, if the probability output by this LDA classifier crossed a threshold of 0.9 for T16 or 0.98 for T11, the controller state was switched to the movement state, which triggered the other components of the control paradigm.

The second component of this control paradigm was a movement direction decoder. This decoder continuously buffered samples (40 × 20 ms bins for T16, 20 × 10 ms bins for T11) from a stream of incoming normalized neural features and exponentially smoothed them (same time constants as the movement initiation decoder for both participants). In the case of participant T16, a linear transformation (computed using PCA) was then applied to reduce the smoothed neural features to a lower-dimensional vector. Once a switch to the movement state was detected by the movement initiation decoder, a linear decoder predicted the relative position of the upcoming target (in Cartesian coordinates) from the processed buffered preparatory activity preceding the detected movement initiation time. The direction of intended movement was computed as the angle towards these predicted coordinates.

The last component of this control paradigm was an autonomous movement generator. When movement was detected by the movement initiation decoder and an intended movement direction was decoded from preparatory activity, this movement generator executed a quick cursor movement to the decoded movement goal position (at a fixed distance in the decoded direction). This autonomous movement followed a truncated Gaussian velocity profile, with a duration of 600 ms for T16 or 400 ms for T11. After reaching the decoded goal position, the cursor held in place until the next trial.

To train the preparatory control paradigm presented in this work, we first collected calibration blocks with each participant during the evaluation sessions: 3 × 3 min open-loop instructed-cued-delay radial-8 blocks with participant T11 and 5 × 4 min open-loop instructed-delay radial-8 blocks with T16. To fit the LDA classifier for the movement initiation decoder with T16, training samples in the window from 400 ms to 0 ms before the target cue and 400 ms to 1000 ms after the target cue were labeled as “no movement” and samples in the window from 400 ms to 1000 ms after the go cue were labeled as “movement”. For trials that skipped the go cue, samples in the window from 400 ms to 1000 ms after the end of the delay period were also labeled as “no movement”. To fit this LDA classifier with participant T11, training samples in the window from 400 ms to 200 ms before the go cue were labeled as “no movement” and samples in the window from 200 ms to 400 ms after the go cue were labeled as “movement”. The movement direction decoder was fit using ridge regression: the inputs to the model were the time-flattened samples of the preparatory activity buffer preceding the decoded movement initiation for each calibration trial and the outputs were the relative Cartesian coordinates of the target position. A cross-validated hyperparameter sweep was run on the training data to determine the regularization value for the ridge regression. For participant T16, the linear transformation to reduce neural data dimensionality prior to movement direction decoding was computed by using PCA to find the components of activity that explained ≥90% of the preparatory activity variance during the calibration blocks.

Optimal linear estimator for online cursor control

Online cursor control using optimal linear estimators (OLEs) was used as a performance comparison to the preparatory control paradigm. The OLE decoders continuously predicted a participant’s intended cursor velocity from a linear readout of an incoming stream of normalized neural features (threshold crossings and SBP). For participant T16, neural features were exponentially smoothed (time constant of 90 ms) before being input to the decoder. For both participants, the velocity predictions from the decoder were also exponentially smoothed (time constant between 28 ms and 45 ms for T11, 70 ms for T16). The OLE implementation used with participant T16 included a softplus nonlinearity in its speed output (to suppress movement at low intended velocities). The OLE control data with participant T11 has been partially released in previous work⁶⁶.

OLE decoders were trained on calibration blocks that preceded evaluation. For participant T16, calibration blocks consisted of 6 min of the radial-8 task: 1 min of open-loop control where T16 attempted to match on-screen cursor movements to different targets followed by 5 min of closed-loop control where cursor movements were driven by an OLE decoder. An initial decoder to map neural features onto intended cursor velocities was fit using ridge regression on data from the open-loop portion. New decoders were then fit every 5 s (similar to procedures for decoder calibration from previous studies^71,72) using up to the previous 5 min of open-loop and closed-loop data. For data from the closed-loop portion used for training, intended cursor velocities at every timepoint were assumed to be the unit vector from the cursor to the target scaled by the magnitude of the original decoded velocity, or zero if the cursor was on the target. To ease control during the start of the closed-loop portion of the task, components of decoded cursor velocities that were perpendicular to a straight line from the cursor to the target were attenuated, with attenuation decreasing from 100% to 0% over the course of the block. For participant T11, the calibration block consisted of 3 min of the open-loop radial-8 task, and the OLE decoder was fit using ridge regression on data from the whole block.

Preparatory control paradigm decoding window sweep

Direction decoding performance for the preparatory control paradigm was evaluated as a function of the start and end times of the window of neural activity used by the decoder, relative to decoded movement onset (Figure 5). Window start times were swept from −800 to 0 ms relative to decoded movement onset (preparatory activity length) and end times from 0 to +800 ms (execution activity length), in steps of 100 ms. For each window, a ridge regression decoder was trained (following the same procedure as for the preparatory control paradigm detailed above) using data from held-out calibration trials – the same trials used to train the decoders used in-session. Decoders were then evaluated on the same trials presented for the online decoding results for participant T16. Performance was quantified as the median angular error across trials between the predicted and true target directions.

Channel count scaling analysis for direction decoding for the preparatory control paradigm

Direction decoding performance for the preparatory control paradigm as a function of the number of active channels (Figure S10) was quantified through random subsampling of available channels. Specifically, channel counts ranging up to the maximum number of channels available for each participant were swept; for each channel count, 20 random subsets were drawn from the pool of total channels, with threshold crossings and spike-band power retained as the neural features for each selected channel. For each subset, a ridge regression decoder was trained (following the same procedure as for the preparatory control paradigm detailed above) on the preparatory activity window preceding decoded movement onset (800 ms for T16, 200 ms for T11) from held-out calibration trials – as for the window sweep, the same trials used to train the decoders used in-session. Decoders were then evaluated on the preparatory windows from the same trials presented for the online decoding results. Performance was quantified as the median angular error between the predicted and true target directions, averaged across the 20 random subsets at each channel count. To predict decoding performance at channel counts higher than those available in the data, a power-law function was fit and used for extrapolation.

Quantification and statistical analysis

Statistical details – analyses, tests used, significance thresholds – are reported in the Results and corresponding figure legends, with the underlying procedures described in the STAR Methods. All analyses were performed using publicly available custom code (see Data and code availability).

For statistical analyses, the sample size n corresponds to the number of trials for the corresponding task or, in some decoding analyses, the number of cross-validation trial splits. Per-participant trial counts for each task dataset are presented in Table S1. At the study level, n = 3 human participants (T16, T11, and T5), though for most figures analyses were only run on data collected from T16 and T11. All results are reported per participant rather than pooled across participants.

Measures of center and dispersion were computed as follows. For single-channel peri-stimulus time histograms (Figures 1C, 2C, 3C, 4C, S2C), EMG traces (Figures 2B, S1), and demixed principal component analysis projections (Figures 2E, 4E, S7), means and standard errors of the mean (traces and shaded regions in the corresponding plots) were computed across trials grouped by condition. For cross-validated decoding analyses, means and 95% confidence intervals (denoted by error bars or shaded regions in plots) were computed across leave-one-out (Figures 1F, 1G, 2F, 3I, 4G, 4H, S2D) or 10-fold (Figures S3A, S3B, S8) cross-validated trial splits, using a Student’s t-distribution. For online decoding results, angular error and time-to-target are summarized by the median and the mean, respectively (Figure 5).

Single-channel tuning was assessed by one-way, two-way, or three-way ANOVA, with significance defined as p < 0.01 (Figures 1D, 3D) or p < 0.05 (Figure 4D); p-values for the multi-way analyses were corrected using the Benjamini-Hochberg procedure⁶⁸. Differences in single-trial decoding accuracy between features (Figure 3I) were assessed using a two-proportion z-test; differences in the time-to-target between online control paradigms (Figure 5G) were assessed using a Mann-Whitney U test. For both tests, significance was defined as p < 0.05 and p-values are reported in the Results.

Only channels meeting the active-channel criterion (mean firing rate ≥ 2 Hz) were included in analyses and online decoding. Trials were excluded from analysis if distractions were logged during data collection with the participant; these trials are not included in the trial counts in Table S1.

Additional resources

Clinical trial registry number: NCT00912041 (https://clinicaltrials.gov/study/NCT00912041?id=NCT0091204).

Supplementary Material

Supplemental PDF

NIHMS2204566-supplement-Supplemental_PDF.pdf^{(3.3MB, pdf)}

Supplemental information

Document S1. Figures S1–S10 and Tables S1–S2.

Key resources table.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Neural and task data used to produce all main text and supplemental figures are publicly available	Dryad	https://doi.org/10.5061/dryad.brv15dvr9
Software and algorithms
BRAND	Ali et al.⁶⁶	https://github.com/brandbci/brand
MATLAB R2023a	MathWorks Inc. https://www.mathworks.com/products/matlab.html	RRID: SCR_001622
Python 3.9	python.org	RRID: SCR_008394
NumPy 1.23.1	numpy.org	RRID: SCR_008633
Scikit-learn 1.4.1	scikit-learn.org	RRID: SCR_002577
Matplotlib 3.8.3	matplotlib.org	RRID: SCR_008624
Seaborn 0.13.2	seaborn.pydata.org	RRID: SCR_018132
SciPy 1.12.0	scipy.org	RRID: SCR_008058
Custom analysis code	https://github.com/snel-repo/human-preparatory-activity-paper	N/A
Other
NeuroPort Neural Signal Processor	Blackrock Neurotech	https://blackrockneurotech.com/products/neuroport/
NeuroPlex E	Blackrock Neurotech	https://blackrockneurotech.com/products/neuroplex-e/
64-channel NeuroPort Electrode	Blackrock Neurotech	https://blackrockneurotech.com/products/neuroport-electrode/
96-channel NeuroPort Electrode	Blackrock Neurotech	https://blackrockneurotech.com/products/neuroport-electrode/

Open in a new tab

Highlights.

Human motor cortex has task-tuned preparatory activity distinct from execution
Preparatory activity is strongly tuned to direction, largely independent of effector
Curvature, distance, and speed are encoded during preparation but less prominently
Online preparatory decoding enables rapid and fluid brain-computer interface control

Acknowledgments

The authors thank participants T16, T11, T5, and their families and care partners for their contributions to this research. The authors also thank Yvan Bamps, Haris Rashid, Colleen Spellen, Maryam Masood, and Dave Rosler for administrative support and clinical research support. This work was supported by NIH-NINDS/OD DP2NS127291, Simons Foundation as part of the Simons-Emory International Consortium on Motor Control (CP), NIH F32HD112173 (SRN), NIH-NIDCD U01DC017844, Department of Veterans Affairs Rehabilitation Research, Development, and Translation A2295R, A4820R (LRH), NIH-NIDCD R01DC014034, Simons Foundation Collaboration on the Global Brain 543045, Larry and Pamela Garlick, Wu Tsai Neurosciences Institute (JMH), and Howard Hughes Medical Institute (DTA). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health, or the Department of Veterans Affairs, or the United States Government.

Footnotes

Declaration of interests

C.P. has served as a consultant and research scientist for Meta (Reality Labs). The MGH Translational Research Center has a clinical research support agreement (CRSA) with Ability Neuro, Axoft, Neuralink, Neurobionics, Paradromics, Precision Neuro, Synchron, and Reach Neuro, for which L.R.H. provides consultative input. L.R.H. is a non-compensated member of the Board of Directors of a nonprofit assistive communication device technology foundation (Speak Your Mind Foundation). Mass General Brigham (MGB) is convening the Implantable Brain-Computer Interface Collaborative Community (iBCI-CC); charitable gift agreements to MGB, including those received to date from Paradromics, Synchron, Precision Neuro, Neuralink, and Blackrock Neurotech, support the iBCI-CC, for which L.R.H. provides effort. J.M.H. is a consultant for Paradromics, serves on the Medical Advisory Board of Enspire DBS and is a shareholder in Maplight Therapeutics. He is also the co-founder of Re-EmergeDBS. He is also an inventor on intellectual property licensed by Stanford University to Blackrock Neurotech and Neuralink Corp. All other authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work the authors used ChatGPT (GPT-4o/GPT-5) and Claude (Sonnet/Opus 4.x) for assistance in generating plotting code for some figures as well as to improve readability and language of the manuscript. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

References

1.Haith AM, and Bestmann S (2020). Preparation of Movement. In The Cognitive Neurosciences, Poeppel D, Mangun GR, and Gazzaniga MS, eds. (The MIT Press; ), pp. 541–548. [Google Scholar]
2.Sutter K, Wijdenes LO, Beers RJ van, and Medendorp, W.P. (2021). Movement preparation time determines movement variability. J. Neurophysiol. 125, 2375–2383. 10.1152/jn.00087.2020. [DOI] [PubMed] [Google Scholar]
3.Vyas S, O’Shea DJ, Ryu SI, and Shenoy KV (2020). Causal Role of Motor Preparation during Error-Driven Learning. Neuron 106, 329–339.e4. 10.1016/j.neuron.2020.01.019. [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Tanji J, and Evarts EV (1976). Anticipatory activity of motor cortex neurons in relation to direction of an intended movement. J. Neurophysiol. 39, 1062–1068. 10.1152/jn.1976.39.5.1062. [DOI] [PubMed] [Google Scholar]
5.Weinrich M, and Wise SP (1982). The premotor cortex of the monkey. J. Neurosci. 2, 1329–1345. 10.1523/JNEUROSCI.02-09-01329.1982. [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Churchland MM, and Shenoy KV (2024). Preparatory activity and the expansive null-space. Nat. Rev. Neurosci. 25, 213–236. 10.1038/s41583-024-00796-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
7.Crammond DJ, and Kalaska JF (2000). Prior Information in Motor and Premotor Cortex: Activity During the Delay Period and Effect on Pre-Movement Activity. J. Neurophysiol. 84, 986–1005. 10.1152/jn.2000.84.2.986. [DOI] [PubMed] [Google Scholar]
8.Kaufman MT, Churchland MM, Ryu SI, and Shenoy KV (2014). Cortical activity in the null space: permitting preparation without movement. Nat. Neurosci. 17, 440–448. 10.1038/nn.3643. [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Elsayed GF, Lara AH, Kaufman MT, Churchland MM, and Cunningham JP (2016). Reorganization between preparatory and movement population responses in motor cortex. Nat. Commun. 7, 13239. 10.1038/ncomms13239. [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Ames KC, Ryu SI, and Shenoy KV (2014). Neural Dynamics of Reaching following Incorrect or Absent Motor Preparation. Neuron 81, 438–451. 10.1016/j.neuron.2013.11.003. [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Lara AH, Elsayed GF, Zimnik AJ, Cunningham JP, and Churchland MM (2018). Conservation of preparatory neural events in monkey motor cortex regardless of how movement is initiated. eLife 7, e31826. 10.7554/eLife.31826. [DOI] [PMC free article] [PubMed] [Google Scholar]
12.Ames KC, Ryu SI, and Shenoy KV (2019). Simultaneous motor preparation and execution in a last-moment reach correction task. Nat. Commun. 10, 2718. 10.1038/s41467-019-10772-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Zimnik AJ, and Churchland MM (2021). Independent generation of sequence elements by motor cortex. Nat. Neurosci. 24, 412–424. 10.1038/s41593-021-00798-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Churchland MM, Cunningham JP, Kaufman MT, Ryu SI, and Shenoy KV (2010). Cortical Preparatory Activity: Representation of Movement or First Cog in a Dynamical Machine? Neuron 68, 387–400. 10.1016/j.neuron.2010.09.015. [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Kao T-C, Sadabadi MS, and Hennequin G (2021). Optimal anticipatory control as a theory of motor preparation: A thalamo-cortical circuit model. Neuron 109, 1567–1581.e12. 10.1016/j.neuron.2021.03.009. [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Riehle A, and Requin J (1993). The predictive value for performance speed of preparatory changes in neuronal activity of the monkey motor and premotor cortex. Behav. Brain Res. 53, 35–49. 10.1016/S0166-4328(05)80264-5. [DOI] [PubMed] [Google Scholar]
17.Churchland MM, Yu BM, Ryu SI, Santhanam G, and Shenoy KV (2006). Neural Variability in Premotor Cortex Provides a Signature of Motor Preparation. J. Neurosci. 26, 3697–3712. 10.1523/JNEUROSCI.3762-05.2006. [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Churchland MM, and Shenoy KV (2007). Delay of Movement Caused by Disruption of Cortical Preparatory Activity. J. Neurophysiol. 97, 348–359. 10.1152/jn.00808.2006. [DOI] [PubMed] [Google Scholar]
19.Michaels JA, Dann B, Intveld RW, and Scherberger H (2015). Predicting Reaction Time from the Neural State Space of the Premotor and Parietal Grasping Network. J. Neurosci. 35, 11415–11432. 10.1523/JNEUROSCI.1714-15.2015. [DOI] [PMC free article] [PubMed] [Google Scholar]
20.Churchland MM, Santhanam G, and Shenoy KV (2006). Preparatory Activity in Premotor and Motor Cortex Reflects the Speed of the Upcoming Reach. J. Neurophysiol. 96, 3130–3146. 10.1152/jn.00307.2006. [DOI] [PubMed] [Google Scholar]
21.Riehle A, and Requin J (1989). Monkey primary motor and premotor cortex: single-cell activity related to prior information about direction and extent of an intended movement. J. Neurophysiol. 61, 534–549. 10.1152/jn.1989.61.3.534. [DOI] [PubMed] [Google Scholar]
22.Even-Chen N, Sheffer B, Vyas S, Ryu SI, and Shenoy KV (2019). Structure and variability of delay activity in premotor cortex. PLOS Comput. Biol. 15, e1006808. 10.1371/journal.pcbi.1006808. [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Churchland MM, Cunningham JP, Kaufman MT, Foster JD, Nuyujukian P, Ryu SI, and Shenoy KV (2012). Neural population dynamics during reaching. Nature 487, 51–56. 10.1038/nature11129. [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Pandarinath C, O’Shea DJ, Collins J, Jozefowicz R, Stavisky SD, Kao JC, Trautmann EM, Kaufman MT, Ryu SI, Hochberg LR, et al. (2018). Inferring single-trial neural population dynamics using sequential auto-encoders. Nat. Methods 15, 805–815. 10.1038/s41592-018-0109-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
25.Hochberg LR, Bacher D, Jarosiewicz B, Masse NY, Simeral JD, Vogel J, Haddadin S, Liu J, Cash SS, van der Smagt P, et al. (2012). Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. Nature 485, 372–375. 10.1038/nature11076. [DOI] [PMC free article] [PubMed] [Google Scholar]
26.Collinger JL, Wodlinger B, Downey JE, Wang W, Tyler-Kabara EC, Weber DJ, McMorland AJ, Velliste M, Boninger ML, and Schwartz AB (2013). High-performance neuroprosthetic control by an individual with tetraplegia. The Lancet 381, 557–564. 10.1016/S0140-6736(12)61816-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Gilja V, Pandarinath C, Blabe CH, Nuyujukian P, Simeral JD, Sarma AA, Sorice BL, Perge JA, Jarosiewicz B, Hochberg LR, et al. (2015). Clinical translation of a high-performance neural prosthesis. Nat. Med. 21, 1142–1145. 10.1038/nm.3953. [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Pandarinath C, Nuyujukian P, Blabe CH, Sorice BL, Saab J, Willett FR, Hochberg LR, Shenoy KV, and Henderson JM (2017). High performance communication by people with paralysis using an intracortical brain-computer interface. eLife 6, e18554. 10.7554/eLife.18554. [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Ajiboye AB, Willett FR, Young DR, Memberg WD, Murphy BA, Miller JP, Walter BL, Sweet JA, Hoyen HA, Keith MW, et al. (2017). Restoration of reaching and grasping movements through brain-controlled muscle stimulation in a person with tetraplegia: a proof-of-concept demonstration. The Lancet 389, 1821–1830. 10.1016/S0140-6736(17)30601-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
30.Musallam S, Corneil BD, Greger B, Scherberger H, and Andersen RA (2004). Cognitive Control Signals for Neural Prosthetics. Science 305, 258–262. 10.1126/science.1097938. [DOI] [PubMed] [Google Scholar]
31.Santhanam G, Ryu SI, Yu BM, Afshar A, and Shenoy KV (2006). A high-performance brain–computer interface. Nature 442, 195–198. 10.1038/nature04968. [DOI] [PubMed] [Google Scholar]
32.Achtman N, Afshar A, Santhanam G, Yu BM, Ryu SI, and Shenoy KV (2007). Free-paced high-performance brain–computer interfaces. J. Neural Eng. 4, 336. 10.1088/1741-2560/4/3/018. [DOI] [PubMed] [Google Scholar]
33.Aflalo T, Kellis S, Klaes C, Lee B, Shi Y, Pejsa K, Shanfield K, Hayes-Jackson S, Aisen M, Heck C, et al. (2015). Decoding motor imagery from the posterior parietal cortex of a tetraplegic human. Science 348, 906–910. 10.1126/science.aaa5417. [DOI] [PMC free article] [PubMed] [Google Scholar]
34.Shahbazi M, Ariani G, Kashefi M, Pruszynski JA, and Diedrichsen J (2024). Neural Correlates of Online Action Preparation. J. Neurosci. 44, e1880232024. 10.1523/JNEUROSCI.1880-23.2024. [DOI] [PMC free article] [PubMed] [Google Scholar]
35.Hosman T, Hynes JB, Saab J, Wilcoxen KG, Buchbinder BR, Schmansky N, Cash SS, Eskandar EN, Simeral JD, Franco B, et al. (2021). Auditory cues reveal intended movement information in middle frontal gyrus neuronal ensemble activity of a person with tetraplegia. Sci. Rep. 11, 98. 10.1038/s41598-020-77616-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
36.Yousry TA, Schmid UD, Alkadhi H, Schmidt D, Peraud A, Buettner A, and Winkler P (1997). Localization of the motor hand area to a knob on the precentral gyrus. A new landmark. Brain 120, 141–157. 10.1093/brain/120.1.141. [DOI] [PubMed] [Google Scholar]
37.Deo DR, Okorokova EV, Pritchard AL, Hahn NV, Card NS, Nason-Tomaszewski SR, Jude J, Hosman T, Choi EY, Qiu D, et al. (2024). A mosaic of whole-body representations in human motor cortex. Preprint at bioRxiv, 10.1101/2024.09.14.613041 https://doi.org/10.1101/2024.09.14.613041. [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Kobak D, Brendel W, Constantinidis C, Feierstein CE, Kepecs A, Mainen ZF, Qi X-L, Romo R, Uchida N, and Machens CK (2016). Demixed principal component analysis of neural population data. eLife 5, e10989. 10.7554/eLife.10989. [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Willett FR, Deo DR, Avansino DT, Rezaii P, Hochberg LR, Henderson JM, and Shenoy KV (2020). Hand Knob Area of Premotor Cortex Represents the Whole Body in a Compositional Way. Cell 181, 396–409.e26. 10.1016/j.cell.2020.02.043. [DOI] [PMC free article] [PubMed] [Google Scholar]
40.Gordon J, Ghilardi MF, and Ghez C (1994). Accuracy of planar reaching movements. Exp. Brain Res. 99, 97–111. 10.1007/BF00241415. [DOI] [PubMed] [Google Scholar]
41.Colins Rodriguez A, Perich MG, Miller LE, and Humphries MD (2024). Motor Cortex Latent Dynamics Encode Spatial and Temporal Arm Movement Parameters Independently. J. Neurosci. 44, e1777232024. 10.1523/JNEUROSCI.1777-23.2024. [DOI] [PMC free article] [PubMed] [Google Scholar]
42.Marshall NJ, Glaser JI, Trautmann EM, Amematsro EA, Perkins SM, Shadlen MN, Abbott LF, Cunningham JP, and Churchland MM (2022). Flexible neural control of motor units. Nat. Neurosci. 25, 1492–1504. 10.1038/s41593-022-01165-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
43.Salinas E, and Abbott LF (1994). Vector reconstruction from firing rates. J. Comput. Neurosci. 1, 89–107. 10.1007/BF00962720. [DOI] [PubMed] [Google Scholar]
44.Willett FR, Young DR, Murphy BA, Memberg WD, Blabe CH, Pandarinath C, Stavisky SD, Rezaii P, Saab J, Walter BL, et al. (2019). Principled BCI decoder design and parameter selection using a feedback control model. Sci. Rep. 10.1038/s41598-019-44166-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
45.Singer-Clark T, Hou X, Card NS, Wairagkar M, Iacobacci C, Peracha H, Hochberg LR, Stavisky SD, and Brandman DM (2025). Speech motor cortex enables BCI cursor control and click. J. Neural Eng. 22, 036015. 10.1088/1741-2552/add0e5. [DOI] [PMC free article] [PubMed] [Google Scholar]
46.Rizzolatti G, and Luppino G (2001). The Cortical Motor System. Neuron 31, 889–901. 10.1016/S0896-6273(01)00423-8. [DOI] [PubMed] [Google Scholar]
47.Geyer S, Matelli M, Luppino G, and Zilles K (2000). Functional neuroanatomy of the primate isocortical motor system. Anat. Embryol. (Berl.) 202, 443–474. 10.1007/s004290000127. [DOI] [PubMed] [Google Scholar]
48.Matelli M, Luppino G, and Rizzolatti G (1991). Architecture of superior and mesial area 6 and the adjacent cingulate cortex in the macaque monkey. J. Comp. Neurol. 311, 445–462. 10.1002/cne.903110402. [DOI] [PubMed] [Google Scholar]
49.Evarts EV (1968). Relation of pyramidal tract activity to force exerted during voluntary movement. J. Neurophysiol. 31, 14–27. 10.1152/jn.1968.31.1.14. [DOI] [PubMed] [Google Scholar]
50.Georgopoulos AP, Kalaska JF, Caminiti R, and Massey JT (1982). On the relations between the direction of two-dimensional arm movements and cell discharge in primate motor cortex. J. Neurosci. 2, 1527–1537. 10.1523/JNEUROSCI.02-11-01527.1982. [DOI] [PMC free article] [PubMed] [Google Scholar]
51.Rademacher J, Caviness VS, Steinmetz H, and Galaburda AM (1993). Topographical variation of the human primary cortices: implications for neuroimaging, brain mapping, and neurobiology. Cereb. Cortex 3, 313–329. 10.1093/cercor/3.4.313. [DOI] [PubMed] [Google Scholar]
52.Rademacher J, Bürgel U, Geyer S, Schormann T, Schleicher A, Freund H-J, and Zilles K (2001). Variability and asymmetry in the human precentral motor system: A cytoarchitectonic and myeloarchitectonic brain mapping study. Brain 124, 2232–2258. 10.1093/brain/124.11.2232. [DOI] [PubMed] [Google Scholar]
53.Geyer S (2004). The microstructural border between the motor and the cognitive domain in the human cerebral cortex. Adv. Anat. Embryol. Cell Biol. 174, I–VIII, 1–89. 10.1007/978-3-642-18910-4. [DOI] [PubMed] [Google Scholar]
54.Glasser MF, Coalson TS, Robinson EC, Hacker CD, Harwell J, Yacoub E, Ugurbil K, Andersson J, Beckmann CF, Jenkinson M, et al. (2016). A multi-modal parcellation of human cerebral cortex. Nature 536, 171–178. 10.1038/nature18933. [DOI] [PMC free article] [PubMed] [Google Scholar]
55.White LE, Andrews TJ, Hulette C, Richards A, Groelle M, Paydarfar J, and Purves D (1997). Structure of the human sensorimotor system. I: Morphology and cytoarchitecture of the central sulcus. Cereb. Cortex 7, 18–30. 10.1093/cercor/7.1.18. [DOI] [PubMed] [Google Scholar]
56.Kurata K (1989). Distribution of neurons with set- and movement-related activity before hand and foot movements in the premotor cortex of rhesus monkeys. Exp. Brain Res. 77, 245–256. 10.1007/BF00274982. [DOI] [PubMed] [Google Scholar]
57.Kalidindi HT, Cross KP, Lillicrap TP, Omrani M, Falotico E, Sabes PN, and Scott SH (2021). Rotational dynamics in motor cortex are consistent with a feedback controller. eLife 10, e67256. 10.7554/eLife.67256. [DOI] [PMC free article] [PubMed] [Google Scholar]
58.Rijntjes M, Dettmers C, Büchel C, Kiebel S, Frackowiak RSJ, and Weiller C (1999). A Blueprint for Movement: Functional and Anatomical Representations in the Human Motor System. J. Neurosci. 19, 8043–8048. 10.1523/JNEUROSCI.19-18-08043.1999. [DOI] [PMC free article] [PubMed] [Google Scholar]
59.Wing AM (2000). Motor control: Mechanisms of motor equivalence in handwriting. Curr. Biol. 10, R245–R248. 10.1016/S0960-9822(00)00375-4. [DOI] [PubMed] [Google Scholar]
60.Shen L, and Alexander GE (1997). Preferential Representation of Instructed Target Location Versus Limb Trajectory in Dorsal Premotor Area. J. Neurophysiol. 77, 1195–1212. 10.1152/jn.1997.77.3.1195. [DOI] [PubMed] [Google Scholar]
61.Vyas S, Golub MD, Sussillo D, and Shenoy KV (2020). Computation Through Neural Population Dynamics. Annu. Rev. Neurosci. 43, 249–275. 10.1146/annurev-neuro-092619-094115. [DOI] [PMC free article] [PubMed] [Google Scholar]
62.Even-Chen N, Stavisky SD, Kao JC, Ryu SI, and Shenoy KV (2017). Augmenting intracortical brain-machine interface with neurally driven error detectors. J. Neural Eng. 14, 066007. 10.1088/1741-2552/aa8dc1. [DOI] [PMC free article] [PubMed] [Google Scholar]
63.Wallace DM, Benyamini M, Nason-Tomaszewski SR, Costello JT, Cubillos LH, Mender MJ, Temmar H, Willsey MS, Patil PG, Chestek CA, et al. (2023). Error detection and correction in intracortical brain–machine interfaces controlling two finger groups. J. Neural Eng. 20, 046037. 10.1088/1741-2552/acef95. [DOI] [PMC free article] [PubMed] [Google Scholar]
64.Dekleva BM, Weiss JM, Boninger ML, and Collinger JL (2021). Generalizable cursor click decoding using grasp-related neural transients. J. Neural Eng. 18, 0460e9. 10.1088/1741-2552/ac16b2. [DOI] [PMC free article] [PubMed] [Google Scholar]
65.Dekleva BM, and Collinger JL (2025). Using transient, effector-specific neural responses to gate decoding for brain–computer interfaces. J. Neural Eng. 22, 016036. 10.1088/1741-2552/adaa1f. [DOI] [PMC free article] [PubMed] [Google Scholar]
66.Ali YH, Bodkin K, Rigotti-Thompson M, Patel K, Card NS, Bhaduri B, Nason-Tomaszewski SR, Mifsud DM, Hou X, Nicolas C, et al. (2024). BRAND: a platform for closed-loop experiments with deep network models. J. Neural Eng. 21, 026046. 10.1088/1741-2552/ad3b3a. [DOI] [PMC free article] [PubMed] [Google Scholar]
67.Young D, Willett F, Memberg WD, Murphy B, Walter B, Sweet J, Miller J, Hochberg LR, Kirsch RF, and Ajiboye AB (2018). Signal processing methods for reducing artifacts in microelectrode brain recordings caused by functional electrical stimulation. J. Neural Eng. 15, 026014. 10.1088/1741-2552/aa9ee8. [DOI] [PMC free article] [PubMed] [Google Scholar]
68.Benjamini Y, and Hochberg Y (1995). Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. J. R. Stat. Soc. Ser. B Methodol. 57, 289–300. 10.1111/j.2517-6161.1995.tb02031.x. [DOI] [Google Scholar]
69.Walther A, Nili H, Ejaz N, Alink A, Kriegeskorte N, and Diedrichsen J (2016). Reliability of dissimilarity measures for multi-voxel pattern analysis. NeuroImage 137, 188–200. 10.1016/j.neuroimage.2015.12.012. [DOI] [PubMed] [Google Scholar]
70.Diedrichsen J, Provost S, and Zareamoghaddam H (2016). On the distribution of cross-validated Mahalanobis distances. Preprint at arXiv, 10.48550/arXiv.1607.01371 https://doi.org/10.48550/arXiv.1607.01371. [DOI] [Google Scholar]
71.Brandman DM, Hosman T, Saab J, Burkhart MC, Shanahan BE, Ciancibello JG, Sarma AA, Milstein DJ, Vargas-Irwin CE, Franco B, et al. (2018). Rapid calibration of an intracortical brain–computer interface for people with tetraplegia. J. Neural Eng. 15, 026007. 10.1088/1741-2552/aa9ee7. [DOI] [PMC free article] [PubMed] [Google Scholar]
72.Orsborn AL, Dangi S, Moorman HG, and Carmena JM (2012). Closed-Loop Decoder Adaptation on Intermediate Time-Scales Facilitates Rapid BMI Performance Improvements Independent of Decoder Initialization Conditions. IEEE Trans. Neural Syst. Rehabil. Eng. 20, 468–477. 10.1109/TNSRE.2012.2185066. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplemental PDF

NIHMS2204566-supplement-Supplemental_PDF.pdf^{(3.3MB, pdf)}

Data Availability Statement

All data required to reproduce the findings in this study are publicly on Dryad at https://doi.org/10.5061/dryad.brv15dvr9 as of the date of publication.
All original code has been deposited at https://github.com/snel-repo/human-preparatory-activity-paper and is publicly available as of the date of publication.
Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

[R1] 1.Haith AM, and Bestmann S (2020). Preparation of Movement. In The Cognitive Neurosciences, Poeppel D, Mangun GR, and Gazzaniga MS, eds. (The MIT Press; ), pp. 541–548. [Google Scholar]

[R2] 2.Sutter K, Wijdenes LO, Beers RJ van, and Medendorp, W.P. (2021). Movement preparation time determines movement variability. J. Neurophysiol. 125, 2375–2383. 10.1152/jn.00087.2020. [DOI] [PubMed] [Google Scholar]

[R3] 3.Vyas S, O’Shea DJ, Ryu SI, and Shenoy KV (2020). Causal Role of Motor Preparation during Error-Driven Learning. Neuron 106, 329–339.e4. 10.1016/j.neuron.2020.01.019. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R4] 4.Tanji J, and Evarts EV (1976). Anticipatory activity of motor cortex neurons in relation to direction of an intended movement. J. Neurophysiol. 39, 1062–1068. 10.1152/jn.1976.39.5.1062. [DOI] [PubMed] [Google Scholar]

[R5] 5.Weinrich M, and Wise SP (1982). The premotor cortex of the monkey. J. Neurosci. 2, 1329–1345. 10.1523/JNEUROSCI.02-09-01329.1982. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R6] 6.Churchland MM, and Shenoy KV (2024). Preparatory activity and the expansive null-space. Nat. Rev. Neurosci. 25, 213–236. 10.1038/s41583-024-00796-z. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R7] 7.Crammond DJ, and Kalaska JF (2000). Prior Information in Motor and Premotor Cortex: Activity During the Delay Period and Effect on Pre-Movement Activity. J. Neurophysiol. 84, 986–1005. 10.1152/jn.2000.84.2.986. [DOI] [PubMed] [Google Scholar]

[R8] 8.Kaufman MT, Churchland MM, Ryu SI, and Shenoy KV (2014). Cortical activity in the null space: permitting preparation without movement. Nat. Neurosci. 17, 440–448. 10.1038/nn.3643. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R9] 9.Elsayed GF, Lara AH, Kaufman MT, Churchland MM, and Cunningham JP (2016). Reorganization between preparatory and movement population responses in motor cortex. Nat. Commun. 7, 13239. 10.1038/ncomms13239. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R10] 10.Ames KC, Ryu SI, and Shenoy KV (2014). Neural Dynamics of Reaching following Incorrect or Absent Motor Preparation. Neuron 81, 438–451. 10.1016/j.neuron.2013.11.003. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R11] 11.Lara AH, Elsayed GF, Zimnik AJ, Cunningham JP, and Churchland MM (2018). Conservation of preparatory neural events in monkey motor cortex regardless of how movement is initiated. eLife 7, e31826. 10.7554/eLife.31826. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R12] 12.Ames KC, Ryu SI, and Shenoy KV (2019). Simultaneous motor preparation and execution in a last-moment reach correction task. Nat. Commun. 10, 2718. 10.1038/s41467-019-10772-2. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R13] 13.Zimnik AJ, and Churchland MM (2021). Independent generation of sequence elements by motor cortex. Nat. Neurosci. 24, 412–424. 10.1038/s41593-021-00798-5. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R14] 14.Churchland MM, Cunningham JP, Kaufman MT, Ryu SI, and Shenoy KV (2010). Cortical Preparatory Activity: Representation of Movement or First Cog in a Dynamical Machine? Neuron 68, 387–400. 10.1016/j.neuron.2010.09.015. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R15] 15.Kao T-C, Sadabadi MS, and Hennequin G (2021). Optimal anticipatory control as a theory of motor preparation: A thalamo-cortical circuit model. Neuron 109, 1567–1581.e12. 10.1016/j.neuron.2021.03.009. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R16] 16.Riehle A, and Requin J (1993). The predictive value for performance speed of preparatory changes in neuronal activity of the monkey motor and premotor cortex. Behav. Brain Res. 53, 35–49. 10.1016/S0166-4328(05)80264-5. [DOI] [PubMed] [Google Scholar]

[R17] 17.Churchland MM, Yu BM, Ryu SI, Santhanam G, and Shenoy KV (2006). Neural Variability in Premotor Cortex Provides a Signature of Motor Preparation. J. Neurosci. 26, 3697–3712. 10.1523/JNEUROSCI.3762-05.2006. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R18] 18.Churchland MM, and Shenoy KV (2007). Delay of Movement Caused by Disruption of Cortical Preparatory Activity. J. Neurophysiol. 97, 348–359. 10.1152/jn.00808.2006. [DOI] [PubMed] [Google Scholar]

[R19] 19.Michaels JA, Dann B, Intveld RW, and Scherberger H (2015). Predicting Reaction Time from the Neural State Space of the Premotor and Parietal Grasping Network. J. Neurosci. 35, 11415–11432. 10.1523/JNEUROSCI.1714-15.2015. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R20] 20.Churchland MM, Santhanam G, and Shenoy KV (2006). Preparatory Activity in Premotor and Motor Cortex Reflects the Speed of the Upcoming Reach. J. Neurophysiol. 96, 3130–3146. 10.1152/jn.00307.2006. [DOI] [PubMed] [Google Scholar]

[R21] 21.Riehle A, and Requin J (1989). Monkey primary motor and premotor cortex: single-cell activity related to prior information about direction and extent of an intended movement. J. Neurophysiol. 61, 534–549. 10.1152/jn.1989.61.3.534. [DOI] [PubMed] [Google Scholar]

[R22] 22.Even-Chen N, Sheffer B, Vyas S, Ryu SI, and Shenoy KV (2019). Structure and variability of delay activity in premotor cortex. PLOS Comput. Biol. 15, e1006808. 10.1371/journal.pcbi.1006808. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R23] 23.Churchland MM, Cunningham JP, Kaufman MT, Foster JD, Nuyujukian P, Ryu SI, and Shenoy KV (2012). Neural population dynamics during reaching. Nature 487, 51–56. 10.1038/nature11129. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R24] 24.Pandarinath C, O’Shea DJ, Collins J, Jozefowicz R, Stavisky SD, Kao JC, Trautmann EM, Kaufman MT, Ryu SI, Hochberg LR, et al. (2018). Inferring single-trial neural population dynamics using sequential auto-encoders. Nat. Methods 15, 805–815. 10.1038/s41592-018-0109-9. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R25] 25.Hochberg LR, Bacher D, Jarosiewicz B, Masse NY, Simeral JD, Vogel J, Haddadin S, Liu J, Cash SS, van der Smagt P, et al. (2012). Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. Nature 485, 372–375. 10.1038/nature11076. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R26] 26.Collinger JL, Wodlinger B, Downey JE, Wang W, Tyler-Kabara EC, Weber DJ, McMorland AJ, Velliste M, Boninger ML, and Schwartz AB (2013). High-performance neuroprosthetic control by an individual with tetraplegia. The Lancet 381, 557–564. 10.1016/S0140-6736(12)61816-9. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R27] 27.Gilja V, Pandarinath C, Blabe CH, Nuyujukian P, Simeral JD, Sarma AA, Sorice BL, Perge JA, Jarosiewicz B, Hochberg LR, et al. (2015). Clinical translation of a high-performance neural prosthesis. Nat. Med. 21, 1142–1145. 10.1038/nm.3953. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R28] 28.Pandarinath C, Nuyujukian P, Blabe CH, Sorice BL, Saab J, Willett FR, Hochberg LR, Shenoy KV, and Henderson JM (2017). High performance communication by people with paralysis using an intracortical brain-computer interface. eLife 6, e18554. 10.7554/eLife.18554. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R29] 29.Ajiboye AB, Willett FR, Young DR, Memberg WD, Murphy BA, Miller JP, Walter BL, Sweet JA, Hoyen HA, Keith MW, et al. (2017). Restoration of reaching and grasping movements through brain-controlled muscle stimulation in a person with tetraplegia: a proof-of-concept demonstration. The Lancet 389, 1821–1830. 10.1016/S0140-6736(17)30601-3. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R30] 30.Musallam S, Corneil BD, Greger B, Scherberger H, and Andersen RA (2004). Cognitive Control Signals for Neural Prosthetics. Science 305, 258–262. 10.1126/science.1097938. [DOI] [PubMed] [Google Scholar]

[R31] 31.Santhanam G, Ryu SI, Yu BM, Afshar A, and Shenoy KV (2006). A high-performance brain–computer interface. Nature 442, 195–198. 10.1038/nature04968. [DOI] [PubMed] [Google Scholar]

[R32] 32.Achtman N, Afshar A, Santhanam G, Yu BM, Ryu SI, and Shenoy KV (2007). Free-paced high-performance brain–computer interfaces. J. Neural Eng. 4, 336. 10.1088/1741-2560/4/3/018. [DOI] [PubMed] [Google Scholar]

[R33] 33.Aflalo T, Kellis S, Klaes C, Lee B, Shi Y, Pejsa K, Shanfield K, Hayes-Jackson S, Aisen M, Heck C, et al. (2015). Decoding motor imagery from the posterior parietal cortex of a tetraplegic human. Science 348, 906–910. 10.1126/science.aaa5417. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R34] 34.Shahbazi M, Ariani G, Kashefi M, Pruszynski JA, and Diedrichsen J (2024). Neural Correlates of Online Action Preparation. J. Neurosci. 44, e1880232024. 10.1523/JNEUROSCI.1880-23.2024. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R35] 35.Hosman T, Hynes JB, Saab J, Wilcoxen KG, Buchbinder BR, Schmansky N, Cash SS, Eskandar EN, Simeral JD, Franco B, et al. (2021). Auditory cues reveal intended movement information in middle frontal gyrus neuronal ensemble activity of a person with tetraplegia. Sci. Rep. 11, 98. 10.1038/s41598-020-77616-8. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R36] 36.Yousry TA, Schmid UD, Alkadhi H, Schmidt D, Peraud A, Buettner A, and Winkler P (1997). Localization of the motor hand area to a knob on the precentral gyrus. A new landmark. Brain 120, 141–157. 10.1093/brain/120.1.141. [DOI] [PubMed] [Google Scholar]

[R37] 37.Deo DR, Okorokova EV, Pritchard AL, Hahn NV, Card NS, Nason-Tomaszewski SR, Jude J, Hosman T, Choi EY, Qiu D, et al. (2024). A mosaic of whole-body representations in human motor cortex. Preprint at bioRxiv, 10.1101/2024.09.14.613041 https://doi.org/10.1101/2024.09.14.613041. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R38] 38.Kobak D, Brendel W, Constantinidis C, Feierstein CE, Kepecs A, Mainen ZF, Qi X-L, Romo R, Uchida N, and Machens CK (2016). Demixed principal component analysis of neural population data. eLife 5, e10989. 10.7554/eLife.10989. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R39] 39.Willett FR, Deo DR, Avansino DT, Rezaii P, Hochberg LR, Henderson JM, and Shenoy KV (2020). Hand Knob Area of Premotor Cortex Represents the Whole Body in a Compositional Way. Cell 181, 396–409.e26. 10.1016/j.cell.2020.02.043. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R40] 40.Gordon J, Ghilardi MF, and Ghez C (1994). Accuracy of planar reaching movements. Exp. Brain Res. 99, 97–111. 10.1007/BF00241415. [DOI] [PubMed] [Google Scholar]

[R41] 41.Colins Rodriguez A, Perich MG, Miller LE, and Humphries MD (2024). Motor Cortex Latent Dynamics Encode Spatial and Temporal Arm Movement Parameters Independently. J. Neurosci. 44, e1777232024. 10.1523/JNEUROSCI.1777-23.2024. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R42] 42.Marshall NJ, Glaser JI, Trautmann EM, Amematsro EA, Perkins SM, Shadlen MN, Abbott LF, Cunningham JP, and Churchland MM (2022). Flexible neural control of motor units. Nat. Neurosci. 25, 1492–1504. 10.1038/s41593-022-01165-8. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R43] 43.Salinas E, and Abbott LF (1994). Vector reconstruction from firing rates. J. Comput. Neurosci. 1, 89–107. 10.1007/BF00962720. [DOI] [PubMed] [Google Scholar]

[R44] 44.Willett FR, Young DR, Murphy BA, Memberg WD, Blabe CH, Pandarinath C, Stavisky SD, Rezaii P, Saab J, Walter BL, et al. (2019). Principled BCI decoder design and parameter selection using a feedback control model. Sci. Rep. 10.1038/s41598-019-44166-7. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R45] 45.Singer-Clark T, Hou X, Card NS, Wairagkar M, Iacobacci C, Peracha H, Hochberg LR, Stavisky SD, and Brandman DM (2025). Speech motor cortex enables BCI cursor control and click. J. Neural Eng. 22, 036015. 10.1088/1741-2552/add0e5. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R46] 46.Rizzolatti G, and Luppino G (2001). The Cortical Motor System. Neuron 31, 889–901. 10.1016/S0896-6273(01)00423-8. [DOI] [PubMed] [Google Scholar]

[R47] 47.Geyer S, Matelli M, Luppino G, and Zilles K (2000). Functional neuroanatomy of the primate isocortical motor system. Anat. Embryol. (Berl.) 202, 443–474. 10.1007/s004290000127. [DOI] [PubMed] [Google Scholar]

[R48] 48.Matelli M, Luppino G, and Rizzolatti G (1991). Architecture of superior and mesial area 6 and the adjacent cingulate cortex in the macaque monkey. J. Comp. Neurol. 311, 445–462. 10.1002/cne.903110402. [DOI] [PubMed] [Google Scholar]

[R49] 49.Evarts EV (1968). Relation of pyramidal tract activity to force exerted during voluntary movement. J. Neurophysiol. 31, 14–27. 10.1152/jn.1968.31.1.14. [DOI] [PubMed] [Google Scholar]

[R50] 50.Georgopoulos AP, Kalaska JF, Caminiti R, and Massey JT (1982). On the relations between the direction of two-dimensional arm movements and cell discharge in primate motor cortex. J. Neurosci. 2, 1527–1537. 10.1523/JNEUROSCI.02-11-01527.1982. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R51] 51.Rademacher J, Caviness VS, Steinmetz H, and Galaburda AM (1993). Topographical variation of the human primary cortices: implications for neuroimaging, brain mapping, and neurobiology. Cereb. Cortex 3, 313–329. 10.1093/cercor/3.4.313. [DOI] [PubMed] [Google Scholar]

[R52] 52.Rademacher J, Bürgel U, Geyer S, Schormann T, Schleicher A, Freund H-J, and Zilles K (2001). Variability and asymmetry in the human precentral motor system: A cytoarchitectonic and myeloarchitectonic brain mapping study. Brain 124, 2232–2258. 10.1093/brain/124.11.2232. [DOI] [PubMed] [Google Scholar]

[R53] 53.Geyer S (2004). The microstructural border between the motor and the cognitive domain in the human cerebral cortex. Adv. Anat. Embryol. Cell Biol. 174, I–VIII, 1–89. 10.1007/978-3-642-18910-4. [DOI] [PubMed] [Google Scholar]

[R54] 54.Glasser MF, Coalson TS, Robinson EC, Hacker CD, Harwell J, Yacoub E, Ugurbil K, Andersson J, Beckmann CF, Jenkinson M, et al. (2016). A multi-modal parcellation of human cerebral cortex. Nature 536, 171–178. 10.1038/nature18933. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R55] 55.White LE, Andrews TJ, Hulette C, Richards A, Groelle M, Paydarfar J, and Purves D (1997). Structure of the human sensorimotor system. I: Morphology and cytoarchitecture of the central sulcus. Cereb. Cortex 7, 18–30. 10.1093/cercor/7.1.18. [DOI] [PubMed] [Google Scholar]

[R56] 56.Kurata K (1989). Distribution of neurons with set- and movement-related activity before hand and foot movements in the premotor cortex of rhesus monkeys. Exp. Brain Res. 77, 245–256. 10.1007/BF00274982. [DOI] [PubMed] [Google Scholar]

[R57] 57.Kalidindi HT, Cross KP, Lillicrap TP, Omrani M, Falotico E, Sabes PN, and Scott SH (2021). Rotational dynamics in motor cortex are consistent with a feedback controller. eLife 10, e67256. 10.7554/eLife.67256. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R58] 58.Rijntjes M, Dettmers C, Büchel C, Kiebel S, Frackowiak RSJ, and Weiller C (1999). A Blueprint for Movement: Functional and Anatomical Representations in the Human Motor System. J. Neurosci. 19, 8043–8048. 10.1523/JNEUROSCI.19-18-08043.1999. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R59] 59.Wing AM (2000). Motor control: Mechanisms of motor equivalence in handwriting. Curr. Biol. 10, R245–R248. 10.1016/S0960-9822(00)00375-4. [DOI] [PubMed] [Google Scholar]

[R60] 60.Shen L, and Alexander GE (1997). Preferential Representation of Instructed Target Location Versus Limb Trajectory in Dorsal Premotor Area. J. Neurophysiol. 77, 1195–1212. 10.1152/jn.1997.77.3.1195. [DOI] [PubMed] [Google Scholar]

[R61] 61.Vyas S, Golub MD, Sussillo D, and Shenoy KV (2020). Computation Through Neural Population Dynamics. Annu. Rev. Neurosci. 43, 249–275. 10.1146/annurev-neuro-092619-094115. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R62] 62.Even-Chen N, Stavisky SD, Kao JC, Ryu SI, and Shenoy KV (2017). Augmenting intracortical brain-machine interface with neurally driven error detectors. J. Neural Eng. 14, 066007. 10.1088/1741-2552/aa8dc1. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R63] 63.Wallace DM, Benyamini M, Nason-Tomaszewski SR, Costello JT, Cubillos LH, Mender MJ, Temmar H, Willsey MS, Patil PG, Chestek CA, et al. (2023). Error detection and correction in intracortical brain–machine interfaces controlling two finger groups. J. Neural Eng. 20, 046037. 10.1088/1741-2552/acef95. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R64] 64.Dekleva BM, Weiss JM, Boninger ML, and Collinger JL (2021). Generalizable cursor click decoding using grasp-related neural transients. J. Neural Eng. 18, 0460e9. 10.1088/1741-2552/ac16b2. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R65] 65.Dekleva BM, and Collinger JL (2025). Using transient, effector-specific neural responses to gate decoding for brain–computer interfaces. J. Neural Eng. 22, 016036. 10.1088/1741-2552/adaa1f. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R66] 66.Ali YH, Bodkin K, Rigotti-Thompson M, Patel K, Card NS, Bhaduri B, Nason-Tomaszewski SR, Mifsud DM, Hou X, Nicolas C, et al. (2024). BRAND: a platform for closed-loop experiments with deep network models. J. Neural Eng. 21, 026046. 10.1088/1741-2552/ad3b3a. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R67] 67.Young D, Willett F, Memberg WD, Murphy B, Walter B, Sweet J, Miller J, Hochberg LR, Kirsch RF, and Ajiboye AB (2018). Signal processing methods for reducing artifacts in microelectrode brain recordings caused by functional electrical stimulation. J. Neural Eng. 15, 026014. 10.1088/1741-2552/aa9ee8. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R68] 68.Benjamini Y, and Hochberg Y (1995). Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. J. R. Stat. Soc. Ser. B Methodol. 57, 289–300. 10.1111/j.2517-6161.1995.tb02031.x. [DOI] [Google Scholar]

[R69] 69.Walther A, Nili H, Ejaz N, Alink A, Kriegeskorte N, and Diedrichsen J (2016). Reliability of dissimilarity measures for multi-voxel pattern analysis. NeuroImage 137, 188–200. 10.1016/j.neuroimage.2015.12.012. [DOI] [PubMed] [Google Scholar]

[R70] 70.Diedrichsen J, Provost S, and Zareamoghaddam H (2016). On the distribution of cross-validated Mahalanobis distances. Preprint at arXiv, 10.48550/arXiv.1607.01371 https://doi.org/10.48550/arXiv.1607.01371. [DOI] [Google Scholar]

[R71] 71.Brandman DM, Hosman T, Saab J, Burkhart MC, Shanahan BE, Ciancibello JG, Sarma AA, Milstein DJ, Vargas-Irwin CE, Franco B, et al. (2018). Rapid calibration of an intracortical brain–computer interface for people with tetraplegia. J. Neural Eng. 15, 026007. 10.1088/1741-2552/aa9ee7. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R72] 72.Orsborn AL, Dangi S, Moorman HG, and Carmena JM (2012). Closed-Loop Decoder Adaptation on Intermediate Time-Scales Facilitates Rapid BMI Performance Improvements Independent of Decoder Initialization Conditions. IEEE Trans. Neural Syst. Rehabil. Eng. 20, 468–477. 10.1109/TNSRE.2012.2185066. [DOI] [PubMed] [Google Scholar]

PERMALINK

Preparatory encoding of intended movement in the human motor cortex and implications for brain-computer interfaces

Mattia Rigotti-Thompson

Samuel R Nason-Tomaszewski

Payton H Bechefsky

Alexander Acosta

Nick V Hahn

Donald T Avansino

Brice A Richards

Claire Nicolas

Yahia H Ali

Jaimie M Henderson

Leigh R Hochberg

Nicholas AuYong

Chethan Pandarinath

Summary

eTOC blurb

Graphical Abstract

Introduction

Results

The human motor cortex exhibits preparatory activity strongly tuned to upcoming movement direction

Figure 1. The hand-knob area of human participants exhibits preparatory activity strongly tuned to upcoming movement direction.

Preparatory activity exhibits effector-independent directional tuning

Figure 2. Preparatory activity exhibits effector-independent directional tuning.

Preparatory activity is tuned to both the initial and target directions of curved trajectories

Figure 3. Preparatory activity is tuned to both the initial and target directions of curved trajectories.

Preparatory activity is slightly tuned to upcoming movement distance and speed

Figure 4. Preparatory activity is slightly tuned to upcoming movement distance and speed.

Online decoding from preparatory activity enables rapid and fluid brain-computer interface control

Figure 5. Online decoding from preparatory activity enables rapid and fluid brain-computer interface control.

Discussion

Delimiting subregions in human motor cortex tied to motor preparation

Comparison with non-human primate studies

Strength of preparatory activity tuning to distance and speed compared to direction

Preparatory activity reflects multiple levels of movement abstraction

Rich preparatory encoding as a mechanism for optimal movement generation

Implications for brain-computer interface control

Resource availability

Lead contact

Materials availability

Data and code availability

STAR★Methods

Experimental model and study participant details

Method details

Research sessions and data collection

Intracortical neural signal acquisition and processing

Electromyography activity acquisition

Eye tracking data collection

Instructed-delay tasks

Peri-stimulus time histograms

Single-channel tuning analyses

Demixed principal component analysis

Cross-validated movement feature decoding

Neural subspace alignment index and visualization

Cross-validated pairwise neural Mahalanobis distances

Cross-validated pairwise neural correlations

Visualization of population activity across curvature conditions using principal component analysis

Preparatory control paradigm for online cursor control

Optimal linear estimator for online cursor control

Preparatory control paradigm decoding window sweep

Channel count scaling analysis for direction decoding for the preparatory control paradigm

Quantification and statistical analysis

Additional resources

Supplementary Material

Key resources table.

Highlights.

Acknowledgments

Footnotes

References

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases