Stability of motor representations after paralysis

  1. Charles Guan
  2. Tyson Aflalo
  3. Carey Y Zhang
  4. Elena Amoruso
  5. Emily R Rosario
  6. Nader Pouratian
  7. Richard A Andersen

  • Curated by eLife

    eLife logo

    Evaluation Summary:

    Using data from a tetraplegic individual, the authors show that the neural representations for attempted single finger movements after multiple years after the injury is still organized in a way that is typical for healthy participants. They also show that the representational structure does not change during task training on a simple finger classification task, and that the representational structure - even without active motor outflow or sensory inflow - switches from a motor representation to a sensory representation during the trial. The results have important implications for the use and training of BCI devices in humans.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 agreed to share their name with the authors.)

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Neural plasticity allows us to learn skills and incorporate new experiences. What happens when our lived experiences fundamentally change, such as after a severe injury? To address this question, we analyzed intracortical population activity in the posterior parietal cortex (PPC) of a tetraplegic adult as she controlled a virtual hand through a brain–computer interface (BCI). By attempting to move her fingers, she could accurately drive the corresponding virtual fingers. Neural activity during finger movements exhibited robust representational structure similar to fMRI recordings of able-bodied individuals’ motor cortex, which is known to reflect able-bodied usage patterns. The finger representational structure was consistent throughout multiple sessions, even though the structure contributed to BCI decoding errors. Within individual BCI movements, the representational structure was dynamic, first resembling muscle activation patterns and then resembling the anticipated sensory consequences. Our results reveal that motor representations in PPC reflect able-bodied motor usage patterns even after paralysis, and BCIs can re-engage these stable representations to restore lost motor functions.

Article activity feed

  1. Version published to 10.7554/elife.74478 on eLife
  2. eLife

    Evaluation Summary:

    Using data from a tetraplegic individual, the authors show that the neural representations for attempted single finger movements after multiple years after the injury is still organized in a way that is typical for healthy participants. They also show that the representational structure does not change during task training on a simple finger classification task, and that the representational structure - even without active motor outflow or sensory inflow - switches from a motor representation to a sensory representation during the trial. The results have important implications for the use and training of BCI devices in humans.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 agreed to share their name with the authors.)

  3. eLife

    Reviewer #1 (Public Review):

    Using data from a tetraplegic individual, the authors first addressed whether the neural representations for attempted single finger movements would still be organized in a way that is typical for healthy participants. They did this by comparing the distances between attempted finger movements in the implanted area to fMRI measures in healthy participants in M1 and ROI that mostly encompassed BA5 (SPLa). The representational structure was more similar to M1 than to SPLa. One weakness in the current posted version is that a) the comparison RDM differs strongly in their reliability and b) the SPLa RDM is likely not very well matched for the implanted location.

    Secondly, they test how the representational structure would change during task training on a simple finger classification task. The authors convincingly demonstrate the stability of the representational structure of the finger movements despite ongoing training and continued confusion between middle, ring and pinkie finger.

    Finally, they demonstrate that the representational structure in the recorded area switches from a more muscle-like representation to a representation that is better explained by an orderly sensory mapping, even though the central-peripheral exchange of sensory-motor signals was completely disrupted in the tetraplegic individual.

    Together the results have potentially important practical implications for the placement of BCI implants, as well as theoretical implications for the role of the implanted region in sensory-motor control of the fingers.

  4. eLife

    Reviewer #2 (Public Review):

    Guan et al. recorded single neuron activity from intracortical microelectrode arrays implanted in the parietal cortex of a person with tetraplegia, while she imagined performing individual finger movements. Their goal was to evaluate whether the neurons in parietal cortex were selective for individual finger movements and whether the representational structure of finger-related information in parietal cortex matched that of able-bodied subjects measured in previous studies using functional magnetic resonance imaging (fMRI). Further, the authors studied whether the representational structure persisted across 10 sessions of brain-computer interface (BCI) use even when it was detrimental to performance.

    Strengths:

    1. Posterior parietal cortex is known to play a role in functional grasping behaviors, but the study of individual finger movement representations is novel. The experiments were carefully designed to isolate activity associated with attempted movements from activity related to visual stimuli or eye movements. The authors clearly demonstrate that neurons in parietal cortex are tuned to individual finger movements and that these activity patterns can be used to achieve a high level of performance in a BCI classification task.

    2. The authors demonstrated that the representational structure in the parietal cortex remained stable across 10 sessions of BCI control, consisting of approximately 4000 trials. The study design and research question is extremely interesting and the data clearly demonstrate that for the BCI paradigm tested there was not a significant change in the representational structure.

    3. Representational dissimilarity analysis quantifies the difference between neural feature patterns associated with different conditions, in this case, individual finger movements. The resulting dissimilarity estimates can be compared across recording modalities and spatial scales, which allows the authors to compare their single unit recordings to rigorous, published work in fMRI of similar tasks.

    4. One advantage of single unit recordings as compared to fMRI is the ability to conduct detailed temporal analysis of the representational structure. The authors consistently found two dominant representational structures with a muscle or usage based model appearing early in movement, and a more somatotopic representational structure later in movement. The suggestion that these components support the idea of parietal cortex as a feedforward controller is intriguing.

    Weaknesses:

    1. The claim that motor representations are preserved after paralysis is incomplete because of the complexity of the findings. Clear evidence is presented that the representational structure in single unit recordings from parietal cortex matches that recorded in motor cortex in able-bodied subjects, and this structure correlates with natural usage statistics. However, imaging studies have yet to demonstrate this structure in the same brain area (parietal lobe). Therefore, it is not clear that the results should be interpreted as a "preservation" of a pattern that existed prior to injury. Rather, the single unit recordings in parietal cortex reflect these usage patterns and perhaps (as speculated by the authors in the discussion) this is due to efferent copy information from motor cortex. Additionally, the authors demonstrate that these representations are dynamic, and that the usage-related structure that emerges early in the movement period is replaced by a more somatotopic structure towards the end of movement. The primary conclusions should reflect the more nuanced findings.

    2. The authors present evidence that the observed representational structure was preserved after many sessions of BCI use even though it was detrimental to performance. This does support the idea that BCIs should take advantage of inherent activity patterns rather than relying on plasticity. However, an important limitation is that the study was not explicitly designed to support learning. The discussion mentions a lack of learning pressure as a potential explanation for the lack of neuroplasticity. However, other aspects of the study design likely contributed including the BCI task itself, which displayed feedback only at the end of each trial rather than continuously. Similarly, the BCI decoder was recalibrated each day. This limitation is mitigated to some extent by the stability of the representational structure, but it may have limited learning when combined with the other aspects of the study. Finally, if an unstructured representation is considered optimal for the task, this would be very far away from the natural covariation patterns that were observed. The authors cite Oby et al. as an example of "off-manifold" learning, but importantly that learning was only achieved when the BCI was perturbed incrementally towards a novel neural activity pattern. A similar approach could be tested (in future studies) to explore the plasticity of finger representations.

  5. eLife

    Reviewer #3 (Public Review):

    The main goals of this study by Guan, Aflalo and colleagues were to examine the encoding scheme of populations of neurons in the posterior parietal cortex (PPC) of a person with paralysis while she attempted individual finger movements as part of a brain-computer interface task (BCI). They used these data to answer several questions:

    1. Could they decode attempted finger movements from these data (building on this group's prior work decoding a variety of movements, including arm movements, from PPC)?

    2. Is there evidence that the encoding scheme for these movements is similar to that of able-bodied individuals, which would argue that even after paralysis, this area is not reorganized and that the motor representations remain more or less stable after the injury?

    3. Related to #2: is there beneficial remapping, such that neural correlates of attempted movements change to improve BCI performance over time?

    4. Can looking at the interrelationship between different fingers' population firing rate patterns (one aspect of the encoding scheme) indicate whether the representation structure is similar to the statistics of natural finger use, a somatotopic organization (how close the fingers are to each other), or be uniformly different from one another (which would be advantageous for the BCI and connects to question #3)? Furthermore, does the best fit amongst these choices to the data change over the course of a movement, indicating a time-varying neural encoding structure or multiple overlapping processes?

    The study is well-conducted and uses sound analysis methods, and is able to contribute some new knowledge related to all of the above questions. These are rare and precious data, given the relatively few people implanted with multielectrode arrays like the Utah arrays used in this study. Even more so when considering that to this reviewer's knowledge, no other group is recording from PPC, and this manuscript thus is the first look at the attempted finger moving encoding scheme in this part of human cortex .

    An important caveat is that the representational similarity analysis (RDA) method and resulting representational dissimilarity matrix (RDM) that is the workhorse analysis/metric throughout the study is capturing a fairly specific question: which pairs of finger movements' neural correlates are more/less similar, and how does that pattern across the pairings compare to other datasets. There are other questions that one could ask with these data (and perhaps this group will in subsequent studies), which will provide additional information about the encoding; for example, how well does the population activity correlate with the kinematics, kinetics, and predicted sensory feedback that would accompany such movements in an able-bodied person?

    What this study shows is that the RDMs from these PPC Utah array data are most similar to motor cortical RDMs based on a prior fMRI study. It's innovative to compare effectors' representational similarity across different recording modalities, but this apparent similarity should be interpreted in light of several limitations: 1) the vastly different spatial scales (voxels spanning cm that average activity of millions of neurons each versus a few mm of cortex with sparse sampling of individual neurons, 2) the vastly different temporal scales (firing rates versus blood flow), 3) that dramatically different encoding schemes and dynamics could still result in the same RDMs. As currently written, the study does not adequately caveat the relatively superficial and narrow similarity being made between these data and the prior Ejaz et al (2015) sensorimotor cortex fMRI results before except for (some) exposition in the Discussion.

    Relatedly, the study would benefit from additional explanation for why the comparison is being made to able-bodied fMRI data, rather than similar intracortical neural recordings made in homologous areas of non-human primates (NHPs), which have been traditionally used as an animal model for vision-guided forelimb reaching. This group has an illustrious history of such macaque studies, which makes this omission more surprising.

    A second area in which the manuscript in its current form could better set the context for its reader is in how it introduces their motivating question of "do paralyzed BCI users need to learn a fundamentally new skillset, or can they leverage their pre-injury motor repertoire". Until the Discussion, there is almost no mention of the many previous human BCI studies where high performance movement decoding was possible based on asking participants to attempt to make arm or hand movements (to just list a small number of the many such studies: Hochberg et al 2006 and 2012, Collinger et al 2013, Gilja et al 2015, Bouton et al 2016, Ajiboye*, Willett* et al 2017; Brandman et al 2018; Willett et al 2020; Flesher et al 2021). This is important; while most of these past studies examined motor (and somatosensory) cortex and not PPC (though this group's prior Aflalo*, Kellis* et al 2015 study did!), they all did show that motor representations remain at least distinct enough between movements to allow for decoding; were qualitatively similar to the able-bodied animal studies upon which that body of work was build; and could be readily engaged by the user just by attempting/imagining a movement. Thus, there was a very strong expectation going into this present study that the result would be that there would be a resemblance to able-bodied motor representational similarity. While explicitly making this connection is a meaningful contribution to the literature by the present study (and so is comparing it to different areas' representational similarity), care should be taken not to overstate the novelty of retained motor encoding schemes in people with paralysis, given the extensive prior work.

    The final analyses in the manuscript are particularly interesting: they examine the representational structure as a function of a short sliding analysis window, which indicates that there is a more motoric representational structure at the start of the movement, followed by a more somatotopic structure. These analyses are a welcome expansion of the study scope to include the population dynamics, and provides clues as to the role of this activity / the computations this area is involved in throughout movement (e.g., the authors speculate the initial activity is an efference copy from motor cortex, and the later activity is a sensory-consequence model).

    An interesting result in this study is that the participant did not improve performance at the task (and that the neural representations of each finger did not change to become more separable by the decoder). This was despite ample room for improvement (the performance was below 90% accuracy across 5 possible choices), at least not over 4,016 trials. The authors provide several possible explanations for this in the Discussion. Another possibility is that the nature of the task impeded learning because feedback was delayed until the end of the 1.5 second attempted movement period (at which time the participant was presented with text reporting which finger's movement was decoded). This is a very different discrete-and-delayed paradigm from the continuous control used in prior NHP BCI studies that showed motor learning (e.g., Sadtler et al 2014 and follow-ups; Vyas et al 2018 and follow-up; Ganguly & Carmena 2009 and follow-ups). It is possible that having continuous visual feedback about the BCI effector is more similar to the natural motor system (where there is consistent visual, as well as proprioceptive and somatosensory feedback about movements), and thus better engages motor adaptation/learning mechanisms.

    Overall the study contributes to the state of knowledge about human PPC cortex and its neurophysiology even years after injury when a person attempts movements. The methods are sound, but are unlikely (in this reviewer's view) to be widely adopted by the community. Two specific contributions of this study are 1) that it provides an additional data point that motor representations are stable after injury, lowering the risk of BCI strategies based on PPC recording; and 2) that it starts the conversation about how to make deeper comparisons between able-bodied neural dynamics and those of people unable to make overt movements.

  6. eLife

    Author Response:

    Reviewer #3 (Public Review):

    The main goals of this study by Guan, Aflalo and colleagues were to examine the encoding scheme of populations of neurons in the posterior parietal cortex (PPC) of a person with paralysis while she attempted individual finger movements as part of a brain-computer interface task (BCI). They used these data to answer several questions:

    1. Could they decode attempted finger movements from these data (building on this group's prior work decoding a variety of movements, including arm movements, from PPC)?
    2. Is there evidence that the encoding scheme for these movements is similar to that of able-bodied individuals, which would argue that even after paralysis, this area is not reorganized and that the motor representations remain more or less stable after the injury?
    3. Related to #2: is there beneficial remapping, such that neural correlates of attempted movements change to improve BCI performance over time?
    4. Can looking at the interrelationship between different fingers' population firing rate patterns (one aspect of the encoding scheme) indicate whether the representation structure is similar to the statistics of natural finger use, a somatotopic organization (how close the fingers are to each other), or be uniformly different from one another (which would be advantageous for the BCI and connects to question #3)? Furthermore, does the best fit amongst these choices to the data change over the course of a movement, indicating a time-varying neural encoding structure or multiple overlapping processes? The study is well-conducted and uses sound analysis methods, and is able to contribute some new knowledge related to all of the above questions. These are rare and precious data, given the relatively few people implanted with multielectrode arrays like the Utah arrays used in this study. Even more so when considering that to this reviewer's knowledge, no other group is recording from PPC, and this manuscript thus is the first look at the attempted finger moving encoding scheme in this part of human cortex .

    An important caveat is that the representational similarity analysis (RDA) method and resulting representational dissimilarity matrix (RDM) that is the workhorse analysis/metric throughout the study is capturing a fairly specific question: which pairs of finger movements' neural correlates are more/less similar, and how does that pattern across the pairings compare to other datasets. There are other questions that one could ask with these data (and perhaps this group will in subsequent studies), which will provide additional information about the encoding; for example, how well does the population activity correlate with the kinematics, kinetics, and predicted sensory feedback that would accompany such movements in an able-bodied person?

    What this study shows is that the RDMs from these PPC Utah array data are most similar to motor cortical RDMs based on a prior fMRI study. It's innovative to compare effectors' representational similarity across different recording modalities, but this apparent similarity should be interpreted in light of several limitations:

    1. the vastly different spatial scales (voxels spanning cm that average activity of millions of neurons each versus a few mm of cortex with sparse sampling of individual neurons, 2) the vastly different temporal scales (firing rates versus blood flow), 3) that dramatically different encoding schemes and dynamics could still result in the same RDMs. As currently written, the study does not adequately caveat the relatively superficial and narrow similarity being made between these data and the prior Ejaz et al (2015) sensorimotor cortex fMRI results before except for (some) exposition in the Discussion.

    We agree that vastly different spatiotemporal scales (comments 1 and 2) limit the chances of finding correspondence between fMRI and single-neuron recordings. We have added motivation for our comparisons to the Results and Discussion sections.

    Revised text in the Results: “We note that our able-bodied model was recorded from human PC-IP using fMRI, which measures fundamentally different features (millimeter-scale blood oxygenation) than microelectrode arrays (sparse sampling of single neurons).”

    Revised text in the Discussion: “This match was surprising because single-neuron and fMRI recordings differ fundamentally; single-neuron recordings sparsely sample 102 neurons in a small region, while fMRI samples 104 – 106 neurons/voxel (Guest and Love, 2017; Kriegeskorte and Diedrichsen, 2016). The correspondence suggested that RSA might identify modality-invariant neural organizations (Kriegeskorte et al., 2008b), so here we used fMRI recordings of human PC-IP as an able-bodied model.” “This result does obscure a straightforward interpretation of the RSA results – why does our recording area match MC better than the corresponding implant location? Several factors might contribute, including differing neurovascular sensitivity to the early and late dynamic phases of the neural response (Figure 4e), heterogeneous neural organizations across the single-neuron and voxel spatial scales (Arbuckle et al., 2020; Guest and Love, 2017; Kriegeskorte and Diedrichsen, 2016), or mismatches in functional anatomy between participant NS and standard atlases (Eickhoff et al., 2018).”

    …3) that dramatically different encoding schemes and dynamics could still result in the same RDMs…

    Regarding point 3, we agree that RSA provides a second-order correspondence (Kriegeskorte et al., 2008a) rather than direct neuron-to-neuron comparisons. To supplement RSA, we also provide more detail on single-neuron responses for the reader in Figure 1–figure supplement 5. However, we believe that population metrics helpfully summarize the computational strategies of recorded brain regions (Cunningham and Yu, 2014; Saxena and Cunningham, 2019), so we focus on population comparisons here.

    Relatedly, the study would benefit from additional explanation for why the comparison is being made to able-bodied fMRI data, rather than similar intracortical neural recordings made in homologous areas of non-human primates (NHPs), which have been traditionally used as an animal model for vision-guided forelimb reaching. This group has an illustrious history of such macaque studies, which makes this omission more surprising.

    We agree that similar intracortical recordings from homologous areas of NHPs would be useful to construct an able-bodied model. While our lab has historically studied NHP reaching and grasping, we unfortunately did not perform any analogous experiments involving individuated finger movements. We have updated the Discussion to clarify this.

    Revised text in the Discussion: “We asked whether participant NS’s BCI finger representations resembled that of able-bodied individuals or whether her finger representations had reorganized after paralysis. Single-neuron recordings of PC-IP during individuated finger movements are not available in either able-bodied human participants or non-human primates. However, many fMRI studies have characterized finger representations (Ejaz et al., 2015; Kikkert et al., 2021, 2016; Yousry et al., 1997), and representational similarity analysis (RSA) has previously shown RDM correspondence between fMRI and single-neuron recordings of another cortical region (inferior temporal cortex) (Kriegeskorte et al., 2008b).”

    A second area in which the manuscript in its current form could better set the context for its reader is in how it introduces their motivating question of "do paralyzed BCI users need to learn a fundamentally new skillset, or can they leverage their pre-injury motor repertoire". Until the Discussion, there is almost no mention of the many previous human BCI studies where high performance movement decoding was possible based on asking participants to attempt to make arm or hand movements (to just list a small number of the many such studies: Hochberg et al 2006 and 2012, Collinger et al 2013, Gilja et al 2015, Bouton et al 2016, Ajiboye, Willett et al 2017; Brandman et al 2018; Willett et al 2020; Flesher et al 2021). This is important; while most of these past studies examined motor (and somatosensory) cortex and not PPC (though this group's prior Aflalo, Kellis et al 2015 study did!), they all did show that motor representations remain at least distinct enough between movements to allow for decoding; were qualitatively similar to the able-bodied animal studies upon which that body of work was build; and could be readily engaged by the user just by attempting/imagining a movement. Thus, there was a very strong expectation going into this present study that the result would be that there would be a resemblance to able-bodied motor representational similarity. While explicitly making this connection is a meaningful contribution to the literature by the present study (and so is comparing it to different areas' representational similarity), care should be taken not to overstate the novelty of retained motor encoding schemes in people with paralysis, given the extensive prior work.

    We agree that multiple previous BCI studies instruct participants to attempt arm/hand movements and that these studies are important to discuss. We have updated the Introduction/Discussion to include these references.

    Our work does fill in two important gaps in the existing literature. First, prior BCI studies had shown general resemblance between able-bodied and BCI movement, but previous human BCI studies had not shown whether the details of pre-injury representations are preserved. We have also updated the manuscript to describe a second motivation: that outside of the BCI community, neuroscientists do not agree on whether BCI studies of tetraplegic humans generalize to able-bodied movement, given the potential for reorganization after severe injury. In the Discussion sections of several recent BCI studies (Armenta Salas et al., 2018; Fifer et al., 2021; Flesher et al., 2016; Stavisky et al., 2019; Willett et al., 2020), the authors addressed whether the newly discovered phenomena were simply artifacts of reorganization (we believe not).

    Revised text in the Introduction: Understanding plasticity is necessary to develop brain-computer interfaces (BCIs) that can restore sensorimotor function to paralyzed individuals(Orsborn et al., 2014). First, paralysis disrupts movement and blocks somatosensory inputs to motor areas, which could cause neural reorganization (Jain et al., 2008; Kambi et al., 2014; Pons et al., 1991). Second, BCIs bypass supporting cortical, subcortical, and spinal circuits, fundamentally altering how the cortex affects movement. Do these changes require paralyzed BCI users to learn fundamentally new motor skills (Sadtler et al., 2014), or do paralyzed participants use a preserved, pre-injury motor repertoire (Hwang et al., 2013)? Several paralyzed participants have been able to control BCI cursors by attempting arm or hand movements (Ajiboye et al., 2017; Bouton et al., 2016; Brandman et al., 2018; Collinger et al., 2013; Gilja et al., 2015; Hochberg et al., 2012, 2006), hinting that motor representations could remain stable after paralysis. However, the nervous system’s capacity for reorganization (Jain et al., 2008; Kambi et al., 2014; Kikkert et al., 2021; Pons et al., 1991) still leaves many BCI studies speculating whether their findings in tetraplegic individuals also generalize to able-bodied individuals (Armenta Salas et al., 2018; Fifer et al., 2021; Flesher et al., 2016; Stavisky et al., 2019; Willett et al., 2020). A direct comparison, between BCI control and able-bodied neural control of movement, would help address questions about generalization.

    In the revised Discussion, we further contextualize our study in the prior work. In particular, as BCI studies have made fundamental neuroscience discoveries, they have had to address whether their results generalize to able-bodied individuals. Direct comparisons between able-bodied movement and tetraplegic BCI movement, like our study, help to bridge this gap.

    Revised text in the Discussion: Early human BCI studies (Collinger et al., 2013; Hochberg et al., 2006) recorded from the motor cortex and found that single-neuron directional tuning is qualitatively similar to that of able-bodied non-human primates (NHPs) (Georgopoulos et al., 1982; Hochberg et al., 2006). Many subsequent human BCI studies have also successfully replicated results from other classical NHP neurophysiology studies (Aflalo et al., 2015; Ajiboye et al., 2017; Bouton et al., 2016; Brandman et al., 2018; Collinger et al., 2013; Gilja et al., 2015; Hochberg et al., 2012), leading to the general heuristic that the sensorimotor cortex retains its major properties after spinal cord injury (Andersen and Aflalo, 2022). This heuristic further suggests that BCI studies of tetraplegic individuals should generalize to able-bodied individuals. However, this generalization hypothesis has so far lacked direct, quantitative comparisons between tetraplegic and able-bodied individuals. Thus, as human BCI studies expand beyond replicating results and begin to challenge conventional wisdom, neuroscientists have questioned whether cortical reorganization could influence these novel phenomena (see Discussions of (Andersen and Aflalo, 2022; Armenta Salas et al., 2018; Chivukula et al., 2021; Fifer et al., 2021; Flesher et al., 2016; Stavisky et al., 2019; Willett et al., 2020)). As an example of a novel discovery, a recent BCI study found that the hand knob of tetraplegic individuals is directionally tuned to movements of the entire body (Willett et al., 2020), challenging the traditional notion that primary somatosensory and motor subregions respond selectively to individual body parts (Penfield and Boldrey, 1937). Given the brain’s capacity for reorganization (Jain et al., 2008; Kambi et al., 2014), could these BCI results be specific to cortical remapping? Detailed comparisons with able-bodied individuals, as shown here, may help shed light on this question.

    The final analyses in the manuscript are particularly interesting: they examine the representational structure as a function of a short sliding analysis window, which indicates that there is a more motoric representational structure at the start of the movement, followed by a more somatotopic structure. These analyses are a welcome expansion of the study scope to include the population dynamics, and provides clues as to the role of this activity / the computations this area is involved in throughout movement (e.g., the authors speculate the initial activity is an efference copy from motor cortex, and the later activity is a sensory-consequence model).

    An interesting result in this study is that the participant did not improve performance at the task (and that the neural representations of each finger did not change to become more separable by the decoder). This was despite ample room for improvement (the performance was below 90% accuracy across 5 possible choices), at least not over 4,016 trials. The authors provide several possible explanations for this in the Discussion. Another possibility is that the nature of the task impeded learning because feedback was delayed until the end of the 1.5 second attempted movement period (at which time the participant was presented with text reporting which finger's movement was decoded). This is a very different discrete-and-delayed paradigm from the continuous control used in prior NHP BCI studies that showed motor learning (e.g., Sadtler et al 2014 and follow-ups; Vyas et al 2018 and follow-up; Ganguly & Carmena 2009 and follow-ups). It is possible that having continuous visual feedback about the BCI effector is more similar to the natural motor system (where there is consistent visual, as well as proprioceptive and somatosensory feedback about movements), and thus better engages motor adaptation/learning mechanisms.

    We agree that different BCI paradigms could better engage motor adaptation and learning, although it is interesting that participant NSS did not improve her performance simply by attempting “natural” finger movements. To better caveat our findings, we have revised our manuscript as suggested.

    Revised text in the Discussion: “The stability of finger representations here suggests that BCIs can benefit from the pre-existing, natural repertoire (Hwang et al., 2013), although learning can play an important role under different experimental constraints. In our study, the participant received only a delayed, discrete feedback signal after classification (Figure 1a). Because we were interested in understanding participant NS’s natural finger representation, we did not artificially perturb the BCI mapping. When given continuous feedback, however, participants in previous BCI studies could learn to adapt to within-manifold perturbations to the BCI mapping (Ganguly and Carmena, 2009; Sadtler et al., 2014; Sakellaridi et al., 2019; Vyas et al., 2018). BCI users can even slowly learn to generate off-manifold neural activity patterns when the BCI decoder perturbations were incremental (Oby et al., 2019). Notably, learning was inconsistent when perturbations were sudden, indicating that learning is sensitive to specific training procedures. So far, most BCI learning studies have focused on two-dimensional cursor control. To further understand how much finger representations can be actively modified, future studies could benefit from perturbations (Kieliba et al., 2021; Oby et al., 2019), continuous neurofeedback (Ganguly and Carmena, 2009; Oby et al., 2019; Vyas et al., 2018), and additional participants.”

    Overall the study contributes to the state of knowledge about human PPC cortex and its neurophysiology even years after injury when a person attempts movements. The methods are sound, but are unlikely (in this reviewer's view) to be widely adopted by the community. Two specific contributions of this study are 1) that it provides an additional data point that motor representations are stable after injury, lowering the risk of BCI strategies based on PPC recording; and 2) that it starts the conversation about how to make deeper comparisons between able-bodied neural dynamics and those of people unable to make overt movements.

  14. Version published to 10.1101/2021.10.07.463105v1 on bioRxiv