Neural stimulation can alleviate or even reverse paralysis and sensory deficits. Rapid technological advancements bring the possibility to develop complex and refined patterns of neurostimulation. However, multipronged interventions with high-density neural interfaces will require algorithmic frameworks to handle optimization in large parameter spaces. Here, we used an algorithmic class, Gaussian-Process (GP)-based Bayesian Optimization (BO), to solve this online problem. We show that GP-BO can efficiently explore the neurostimulation parameters’ space, exceeding extensive search performance after testing only a fraction of the possible combinations. It can quickly optimize multi-channel neurostimulation across diverse biological targets (brain and spinal cord), animal models (rats and non-human primates), in healthy and injured subjects. Moreover, since BO can embed and improve ‘prior’ expert/clinical knowledge, the performance can be dramatically enhanced even further. These results support broad establishment of learning agents as a structural part of neuroprosthetic design, enabling therapeutic personalization and maximization of intervention effectiveness.
In rats, forelimb movements can be evoked from two distinct cortical regions, the rostral (RFA) and the caudal (CFA) forelimb areas. RFA and CFA have numerous reciprocal connections, and their projections reach several common targets, which allows them to interact at multiple levels of the motor axis. Lesions affecting these areas result in profound and persistent deficits, supporting their essential role for the production of arm and hand movements. Whereas rats are widely used to study motor control and recovery following lesions, little is known as to how cortical motor areas in this model interact to generate movements. To study interactions between RFA and CFA, we used paired-pulse protocols with intracortical microstimulation techniques (ICMS). A conditioning stimulus (C) in RFA was applied simultaneously, or before a test stimulus (T) in CFA. The impact of RFA conditioning on CFA outputs was quantified by recording electromyographic signals (EMG) signals from the contralateral arm muscles. We found that stimulation of RFA substantially modulates the intensity of CFA outputs while only mildly affecting the latency. In general, the effect of RFA conditioning changed from predominantly facilitatory to inhibitory with increasing delays between the C and the T stimulus. However, inspection of individual cortical sites revealed that RFA has a wide range of influence on CFA outputs with each interstimulation delay we used. Our results show that RFA has powerful and complex modulatory effects on CFA outputs that can allow it to play a major role in the cortical control of forelimb movements.
Monkeys were trained in a serial reaction time task to produce hand movements according to changing locations of visual targets. In most trials, targets followed the same sequence repeatedly, whereas in other trials targets were presented in random locations or switched unpredictably between two alternative sequences. Single-unit activity was recorded from the caudal supplementary motor area (SMA-proper). Based on the activity associated with random movement sequences, effects of hand position and movement direction were evaluated. Activity was influenced by the hand position in ∼60% of the neurons, and the movement direction influenced the activity of 51% of the neurons. In addition, 37 and 71% of SMA neurons displayed nonstationarity in their activity across successive movements within a given trial and across trials, respectively. Such nonstationarity in the ongoing neural activity and the effects of performance-related variables were evaluated using a regression model and separated from learning-related activity changes. About a third of SMA neurons displayed gradual changes in neural activity related to experience with a movement sequence across trials. Furthermore, about a quarter of SMA neurons showed similar changes within individual trials. When the individual movements included in the frequently repeated movement sequences were introduced unexpectedly, learning-related changes in neural activity were reduced, indicating that many SMA neurons changed their activity in relation to the learning of particular movement sequences. These results suggest that the pattern of neural activity in the cortical network involved in the control of movement sequences can be modified continuously by experience.
It is well established that superior colliculus (SC) multisensory neurons integrate cues from different senses; however, the mechanisms responsible for producing multisensory responses are poorly understood. Previous studies have shown that spatially congruent cues from different modalities (e.g., auditory and visual) yield enhanced responses and that the greatest relative enhancements occur for combinations of the least effective modality-specific stimuli. Although these phenomena are well documented, little is known about the mechanisms that underlie them, because no study has systematically examined the operation that multisensory neurons perform on their modality-specific inputs. The goal of this study was to evaluate the computations that multisensory neurons perform in combining the influences of stimuli from two modalities. The extracellular activities of single neurons in the SC of the cat were recorded in response to visual, auditory, and bimodal visual-auditory stimulation. Each neuron was tested across a range of stimulus intensities and multisensory responses evaluated against the null hypothesis of simple summation of unisensory influences. We found that the multisensory response could be superadditive, additive, or subadditive but that the computation was strongly dictated by the efficacies of the modality-specific stimulus components. Superadditivity was most common within a restricted range of near-threshold stimulus efficacies, whereas for the majority of stimuli, response magnitudes were consistent with the linear summation of modality-specific influences. In addition to providing a constraint for developing models of multisensory integration, the relationship between response mode and stimulus efficacy emphasizes the importance of considering stimulus parameters when inducing or interpreting multisensory phenomena.
A rapidly growing number of studies using inhibition of the contralesional hemisphere after stroke are reporting improvement in motor performance of the paretic hand. These studies have used different treatment onset time, duration and non-invasive methods of inhibition. Whereas these results are encouraging, several questions regarding the mechanisms of inhibition and the most effective treatment parameters are currently unanswered. In the present study, we used a rat model of cortical lesion to study the effects of GABA-mediated inactivation on motor recovery. In particular, we were interested in understanding better the effect of inactivation duration when it is initiated within hours following a cortical lesion. Cortical lesions were induced with endothelin-1 microinjections. The contralesional hemisphere was inactivated with continuous infusion of the GABA-A agonist Muscimol for 3, 7 or 14days in three different groups of animals. In a fourth group, Muscimol was infused at slower rate for 14days to provide additional insights on the relation between the effects of inactivation on the non-paretic forelimb behavior and the recovery of the paretic forelimb. In spontaneously recovered animals, the lesion caused a sustained bias to use the non-paretic forelimb and long-lasting grasping deficits with the paretic forelimb. Contralesional inactivation produced a general decrease of behavioral activity, affected the spontaneous use of the forelimbs and caused a specific reduction of the non-paretic forelimb function. The intensity and the duration of these behavioral effects varied in the different experimental groups. For the paretic forelimb, increasing inactivation duration accelerated the recovery of grasping function. Both groups with 14days of inactivation had similar recovery profiles and performed better than animals that spontaneously recovered. Whereas the plateau performance of the paretic forelimb correlated with the duration of contralesional inactivation, it was not correlated with the spontaneous use of the forelimbs or with grasping performance of the non-paretic hand. Our results support that contralesional inactivation initiated within hours after a cortical lesion can improve recovery of the paretic forelimb. In our model, increasing the duration of the inactivation improved motor outcomes but the spontaneous use and motor performance of the non-paretic forelimb had no impact on recovery of the paretic forelimb.
Neural stimulation can alleviate paralysis and sensory deficits. Novel high-density neural interfaces can enable refined and multipronged neurostimulation interventions. To achieve this, it is essential to develop algorithmic frameworks capable of handling optimization in large parameter spaces. Here, we leveraged an algorithmic class, Gaussian-process (GP)-based Bayesian optimization (BO), to solve this problem. We show that GP-BO efficiently explores the neurostimulation space, outperforming other search strategies after testing only a fraction of the possible combinations. Through a series of real-time multi-dimensional neurostimulation experiments, we demonstrate optimization across diverse biological targets (brain, spinal cord), animal models (rats, non-human primates), in healthy subjects, and in neuroprosthetic intervention after injury, for both immediate and continual learning over multiple sessions. GP-BO can embed and improve "prior" expert/clinical knowledge to dramatically enhance its performance. These results advocate for broader establishment of learning agents as structural elements of neuroprosthetic design, enabling personalization and maximization of therapeutic effectiveness.
Brain injuries cause hemodynamic changes in several distant, spared areas from the lesion. Our objective was to better understand the neuronal correlates of this reorganization in awake, behaving female monkeys. We used reversible inactivation techniques to “injure” the primary motor cortex, while continuously recording neuronal activity of the ventral premotor cortex in the two hemispheres, before and after the onset of behavioral impairments. Inactivation rapidly induced profound alterations of neuronal discharges that were heterogeneous within each and across the two hemispheres, occurred during movements of either the affected or nonaffected arm, and varied during different phases of grasping. Our results support that extensive, and much more complex than expected, neuronal reorganization takes place in spared areas of the bihemispheric cortical network involved in the control of hand movements. This broad pattern of reorganization offers potential targets that should be considered for the development of neuromodulation protocols applied early after brain injury. SIGNIFICANCE STATEMENT It is well known that brain injuries cause changes in several distant, spared areas of the network, often in the premotor cortex. This reorganization is greater early after the injury and the magnitude of early changes correlates with impairments. However, studies to date have used noninvasive brain imaging approaches or have been conducted in sedated animals. Therefore, we do not know how brain injuries specifically affect the activity of neurons during the generation of movements. Our study clearly shows how a lesion rapidly impacts neurons in the premotor cortex of both hemispheres. A better understanding of these complex changes can help formulate hypotheses for the development of new treatments that specifically target neuronal reorganization induced by lesions in the brain.
Nonhuman primate (NHP) movement kinematics have been decoded from spikes and local field potentials (LFPs) recorded during motor tasks. However, the potential of LFPs to provide network-like characterizations of neural dynamics during planning and execution of sequential movements requires further exploration. Is the aggregate nature of LFPs suitable to construct informative brain state descriptors of movement preparation and execution? To investigate this, we developed a framework to process LFPs based on machine-learning classifiers and analyzed LFP from a primate, implanted with several microelectrode arrays covering the premotor cortex in both hemispheres and the primary motor cortex on one side. The monkey performed a reach-to-grasp task, consisting of five consecutive states, starting from rest until a rewarding target (food) was attained. We use this five-state task to characterize neural activity within eight frequency bands, using spectral amplitude and pairwise correlations across electrodes as features. Our results show that we could best distinguish all five movement-related states using the highest frequency band (200-500 Hz), yielding an 87% accuracy with spectral amplitude, and 60% with pairwise electrode correlation. Further analyses characterized each movement-related state, showing differential neuronal population activity at above-γ frequencies during the various stages of movement. Furthermore, the topological distribution for the high-frequency LFPs allowed for a highly significant set of pairwise correlations, strongly suggesting a concerted distribution of movement planning and execution function is distributed across premotor and primary motor cortices in a specific fashion, and is most significant in the low ripple (100-150 Hz), high ripple (150-200 Hz), and multiunit frequency bands. In summary, our results show that the concerted use of novel machine-learning techniques with coarse grained queue broad signals such as LFPs may be successfully used to track and decode fine movement aspects involving preparation, reach, grasp, and reward retrieval across several brain regions.NEW & NOTEWORTHY Local field potentials (LFPs), despite lower spatial resolution compared to single-neuron recordings, can be used with machine learning classifiers to decode sequential movements involving motor preparation, execution, and reward retrieval. Our results revealed heterogeneity of neural activity on small spatial scales, further evidencing the utility of micro-electrode array recordings for complex movement decoding. With further advancement, high-dimensional LFPs may become the gold standard for brain-computer interfaces such as neural prostheses in the near future.
There is experimental evidence of varying correlation among the elements of the neuromuscular system over the course of the reach-and-grasp task. Several neuromuscular disorders are accompanied by anomalies in muscular coupling during the task. The aim of this study was to investigate if modifications in correlations and clustering can be detected in the Local Field Potential (LFP) recordings of the motor cortex during the task. To this end, we analyzed the LFP recordings from a previously published study on monkeys which performed a reach-and-grasp task for targets with a vertical or horizontal orientation. LFP signals were recorded from the motor and premotor cortex of macaque monkeys as they performed the task. We found very robust changes in the correlations of the multielectrode LFP recordings which corresponded to task epochs. Mean LFP correlation increased significantly during reaching and then decreased during grasp. This pattern was very robust for both left and right arm reaches irrespective of target orientation. A hierarchical cluster analysis supported the same conclusion – a decreased number of clusters during reach followed by an increase for grasp. A sliding window computation of the number of clusters was performed to probe the predictive capacities of these LFP clusters for upcoming task events. For a very high percentage of trials (95.3%), there was a downturn in cluster number following the Pellet Drop (GO signal) which reached a minimum shortly preceding the Start of grasp, hence indicating that cluster analyses of LFP signals could provide online indications of the Start of grasp.
The ventral premotor cortex (PMv) is a key node in the neural network involved in grasping. One way PMv can carry out this function is by modulating the outputs of the primary motor cortex (M1) to intrinsic hand and forearm muscles. As many PMv neurons discharge when grasping with either arm, both PMv within the same hemisphere (ipsilateral; iPMv) and in the opposite hemisphere (contralateral; cPMv) could modulate M1 outputs. Our objective was to compare modulatory effects of iPMv and cPMv on M1 outputs to intrinsic hand and forearm muscles. We used paired-pulse protocols with intracortical microstimulations in capuchin monkeys. A conditioning stimulus was applied in either iPMv or cPMv simultaneously or prior to a test stimulus in M1 and the effects quantified in electromyographic signals. Modulatory effects from iPMv were predominantly facilitatory, and facilitation was much more common and powerful on intrinsic hand than forearm muscles. In contrast, while the conditioning of cPMv could elicit facilitatory effects, in particular to intrinsic hand muscles, it was much more likely to inhibit M1 outputs. These data show that iPMv and cPMv have very different modulatory effects on the outputs of M1 to intrinsic hand and forearm muscles.
