Neural population dynamics reveals disruption of spinal circuits’ responses to proprioceptive input during electrical stimulation of sensory afferents

SUMMARY

While neurostimulation technologies are rapidly approaching clinical applications for sensorimotor disorders, the impact of electrical stimulation on network dynamics is still unknown. Given the high degree of shared processing in neural structures, it is critical to understand if neurostimulation affects functions that are related to, but not targeted by, the intervention. Here, we approach this question by studying the effects of electrical stimulation of cutaneous afferents on unrelated processing of proprioceptive inputs. We recorded intraspinal neural activity in four monkeys while generating proprioceptive inputs from the radial nerve. We then applied continuous stimulation to the radial nerve cutaneous branch and quantified the impact of the stimulation on spinal processing of proprioceptive inputs via neural population dynamics. Proprioceptive pulses consistently produce neural trajectories that are disrupted by concurrent cutaneous stimulation. This disruption propagates to the somatosensory cortex, suggesting that electrical stimulation can perturb natural information processing across the neural axis.

Graphical abstract

In brief

Katic Secerovic et al. examine the impact of electrical stimulation of cutaneous afferents on the processing of proprioceptive information in the monkey spinal cord. They find that proprioceptive processing is significantly disrupted with concurrent cutaneous stimulation. These findings suggest that artificially generated input disrupts neural processes unrelated to the stimulation targets.

INTRODUCTION

Decades of animal and human studies have shown that neurostimulation technologies can restore some level of neurological function in patients with sensorimotor deficits.^1–10 These novel technologies produce immediate assistive effects, achieving a controlled restoration of multifaceted behavioral processes.¹¹ For instance, in humans, peripheral neuroprostheses successfully restore touch sensations,^12–20 and spinal cord stimulation enables the recovery of voluntary motor control.^3–5 While these remarkable results are fueling the translation of these technologies in clinical settings, the short- and long-term effects of injecting electric current into existing neural dynamics are still entirely unknown. In fact, virtually all these interventions suffer from a latent, yet critical, caveat: the input delivered to the neural circuits is artificially generated, being widely different from naturally generated neural activity.

Indeed, electrical stimulation produces synchronized volleys of neural activity in all recruited axons (or cells), rather than the asynchronous bursts of input that govern natural neural activity.^21,22 What is the consequence of this stark difference with respect to neural function? Recently, some studies demonstrated that electrical stimulation actually triggers side effects at the neural level, which were initially unnoticed. For example, new data from epidural spinal cord stimulation for spinal cord injury showed that continuous stimulation of recruited sensory afferents produces a disruption of proprioceptive percepts at stimulation parameters commonly employed in clinical trials.⁶ Similarly, the inability to elicit robust proprioceptive percepts²³ is striking in the application of electrical stimulation of the peripheral nerves for the restoration of somatosensations.

In fact, large-diameter proprioceptive afferents should have the lowest threshold for electrical stimulation. Therefore, these afferents should be the easiest sensory afferents to recruit with neural interfaces.^24–28 However, because of anatomical and geometrical constraints, in practice, electrical neurostimulation leads to the activation of mixed-diameter fiber distributions and, consequently, different sensory modalities.^8,25,29 Therefore, cutaneous afferents are recruited concurrently, along with larger-diameter afferents.³⁰ These fibers converge on interneurons in the spinal cord, where they undergo the first layer of sensory processing, representing a highly shared sensory network node.

It is conceivable that artificially generated patterns of mixed neural activity hinder some of the neural activity within these shared network nodes, thus impairing natural circuit processing, and hence, perception (Figure 1). To demonstrate this conjecture, we need tools that allow us to visualize and identify a direct measure of network activity.³¹ Analysis of population neural dynamics using neural manifolds is commonly employed to study computational objects that process information in the cortex^32–35 and, more recently, in the spinal cord. Indeed, one study showed that intraspinal population responses contain simple structures that enable the examination of complex processes such as walking.³⁶

Figure 1. Electrical stimulation disrupts computations of ongoing network processes.

Neural networks (cyan- and magenta-shaded areas) produce a desired neural function (functions A and B, respectively) within a highly shared neural architecture. These networks may share processing layers (or nodes) to process input information. Top, naturally generated neural activity of unrelated neural functions successfully processes ongoing information input throughout the shared neural architecture. Bottom, stimulation-induced neural activity targeted to restore function B (magenta) artificially processes information input while impairing information processing from an unrelated neural function (function A, cyan). Specifically, artificially induced processing in the shared processing nodes concurrently hinders computations of unrelated ongoing processing of function A, which may also be unselectively recruited by the electrical stimulation.

Therefore, here we employed neural population analysis of intraspinal neural dynamics to (1) visualize spinal network responses to brief proprioceptive percepts elicited by single short pulses of electrical stimulation and (2) study how these responses were altered when concurrent electrical stimulation was delivered to sensory afferents from a different nerve. This experimental design offered a simplified version of the more general problem of the stimulation effects on unrelated neural functions, thus allowing us to execute casual manipulation and quantification of neural variables.

Therefore, we designed a series of electrophysiology experiments in anesthetized monkeys, who share distinguishable projections with the human nervous system distinct from all other animals.^37,38 We recorded and analyzed artificially evoked proprioceptive neural signals in both the cervical spinal cord and the somatosensory cortex. Specifically, we induced proprioceptive input in the hand and forearm by cuff electrode stimulation of the muscle branch of the radial nerve, which does not contain cutaneous afferents.^39,40 Then, we studied how concurrent stimulation of somatosensory afferents in the cutaneous branch of the radial nerve affected the spinal and cortical proprioceptive responses. Using neural population analysis, we examined dorsoventral intraspinal spiking activity in response to muscle-nerve stimulation pulses and performed dimensionality reduction to observe the spinal neural trajectories. Concurrent stimulation of the cutaneous afferents disrupted these neural trajectories, suggesting a significant degradation of proprioceptive information processing in the spinal cord. Changes in proprioceptive information appeared as reduced cortical responses in the somatosensory cortex. Our results show that intraspinal neural population dynamics can capture the processing of sensorimotor information in spinal networks and the disruption of this information processing during artificial electrical stimulation.

RESULTS

Simultaneous brain and spinal neural recordings during electrical nerve stimulation of multiple sensory modalities

We designed an experimental setup in non-human primates as a proxy to understand how artificial inputs can influence neural network function in a controlled fashion. Specifically, we examined how stimulation of cutaneous afferents affects spinal network processing of proprioceptive pulses. The radial nerve, carrying sensory signals from the dorsal part of the forearm and hand, splits in the proximity of the elbow into a pure-muscle and a pure-cutaneous branch (i.e., the deep and superficial branches of the radial nerve,⁴⁰ respectively), offering the opportunity to provide modality-selective sensory stimuli. We implanted cuff electrodes on these two branches to elicit either proprioceptive or cutaneous inputs via electrical stimulation (Figure 2A).

Figure 2. Experimental setup: Schematic illustration of experiments.

(A) Stimulation: we implanted two nerve cuffs for stimulation on the superficial branch (cutaneous nerve) and the deep branch (muscle nerve) of the radial nerve. We stimulated the muscle nerve at ~2 Hz, exclusively or concurrent with ~50 Hz stimulation of the cutaneous nerve branch.

(B) We recorded neural activity with a 32-channel dorsoventral linear probe implanted in the gray matter of the spinal segment C5. Typical intraspinal neural responses were induced by stimulation of the muscle nerve. Zoom insets show examples of detected spike waveforms, e.g., single-unit responses to proprioceptive pulses.

(C) We recorded neural activity with a 32-channel multielectrode array in the somatosensory cortex and provided intracortical neural responses, similar to (B).

We artificially provided brief proprioceptive pulses by stimulating the muscle branch of the radial nerve with single electrical pulses (~2 Hz) below the motor threshold. To assess the influence of artificial cutaneous input on the induced proprioceptive input, we overlapped the muscle-branch stimulation with continuous stimulation of the superficial branch of the radial nerve (i.e., the cutaneous-nerve branch). We provided constant stimulation of the cutaneous-nerve branch at ~50 Hz, which is a typical stimulation frequency used in human studies. Threshold (Thr) was defined as an amplitude that clearly evoked potentials in the spinal cord in response to low-frequency stimulation. We tested two conditions: stimulating the nerve at a low (0.9 × Thr) or a high amplitude (1.1 × Thr). Stimulation amplitude corresponds to the amount of artificially recruited fibers. To study the transmission of artificially induced proprioceptive percepts from the periphery to the cerebral cortex, we recorded the macaque monkeys’ intraspinal neural signals from a dorsoventral 32-channel linear probe implanted in the gray matter of the spinal cord C5 segment (Figure 2B). Furthermore, we extracted intracortical neural signals (Figure 2C) using a 32-channel UTAH array placed in the somatosensory cortex (area S1/S2; Figure S1).

In summary, we recorded neural signals in the spinal cord and the somatosensory cortex of three anesthetized Macaca fascicularis (MK1, MK2, and MK4) and one Macaca mulatta (Mk-Hs) monkey while stimulating only proprioceptive or concurrently proprioceptive and cutaneous afferents.

Proprioceptive inputs elicit robust trajectories in the spinal neural manifold

We explored the effect of brief pulses of artificially generated proprioceptive inputs on the intraspinal neural population dynamics. Because proprioceptive signals enter the spinal cord from the dorsal aspect and project toward medial and ventral laminae,⁴¹ we performed neural population analysis of the multiunit spiking data from all the channels of our linear probe (Figure 3A) in response to 2 Hz muscle-nerve stimulation. Specifically, we applied dimensionality reduction to unveil the latent properties of the spinal neural processing via principal-component analysis (PCA). PCA identified three neural modes that sufficed to explain 54%–65% of the variance of the spike counts of multiunit threshold crossings recorded by the spinal probe for ~350 ms following each proprioceptive stimulus pulse. We then verified whether the neural manifold defined by these neural modes contained simple neural trajectories that captured the changes of time-varying spikes (Figures 3B and 3C).

Figure 3. Intraspinal neural population analysis.

(A) Latent dynamics and neural modes obtained from the multiunit activity recorded per channel. Left, a sketch of the dorsoventral linear probe that recorded the activity of the spinal multiunit neural networks (each circle represents a recorded unit). Each color represents the neural activity recorded by each channel. Right, dimensionality reduction technique identifies the neural modes that define the low-dimensional spaces. In these subspaces, the neural activity followed precise dynamics.

(B and C) We hypothesized that (B) a muscle-nerve stimulation pulse elicits neural trajectories and that (C) these neural trajectories shrink as a function of the stimulation amplitude.

(D) Top, averaged multiunit spike counts across all 32 channels, sorted by the highest spiking activity after the muscle-nerve stimulation, for MK1. Bottom, resultant 10-trial averaged neural trajectories elicited by muscle-nerve stimulation for MK1. This is plotted at both a high- and a low-stimulation amplitude to appreciate the phenomenon of trajectory shrinking.

(E) Statistical quantification of the trajectory length for all monkeys for high- and low-stimulation amplitude of the muscle nerve (***p < 0.001; **p < 0.01; *p < 0.05; Kruskal-Wallis test, with 380 and 231 points for high and low amplitude, respectively, for MK1; 353 and 351 points for high and low amplitude, respectively, for MK2; 353 and 343 points, respectively, for Mk-Hs; and 391 and 394 points, respectively, for MK4). Violin plots: each dot corresponds to the computed trajectory length for a trial, forming a Gaussian distribution of trajectory lengths. The central mark represented as a white dot indicates the median, and the gray line indicates the 25th and 75th percentiles. The whiskers extend to the most extreme data points not considered outliers. Trial corresponds to a stimulation pulse.

In the spinal manifold, the multiunit spike counts elicited very consistent dynamics after each stimulation pulse in the form of closed trajectories that were qualitatively similar in all monkeys (Figure 3D). Because averaged spiking responses initiated and terminated with baseline activity (i.e., no stimulation), the neural dynamics were represented by closed neural trajectories. Given the robustness and reproducibility of these trajectories, we hypothesized that estimated trajectory lengths could be used as a proxy to measure the amount of proprioceptive information processed within the recorded site. The logical consequence of this interpretation is that the length of the trajectories could be proportional to the amount of proprioceptive input processed.

Since the stimulation amplitude controls the number of recruited afferents, we tested this assumption by computing the neural trajectories induced by proprioceptive inputs at both high- and low-stimulation amplitudes (i.e., more or fewer recruited afferents, respectively). As expected, we found that muscle-nerve stimulation at a higher amplitude elicited longer trajectories and vice versa (Figures 3D and S2). This observation was consistent in MK1 (relative mean difference, +14.17%), MK2 (+24.05%), and Mk-Hs (+44.21%; Figure 3E), but not in MK4 (−33.76%), probably due to the higher variability in the overall trajectories for this monkey.

In summary, we showed that population analysis of a dorsoventral linear probe in the spinal cord shows highly robust and reproducible trajectories in the neural manifold in response to artificial proprioceptive pulses. We proposed to quantify the length of this trajectory as a means to assess the amount of proprioceptive information processed in the spinal cord.

Continuous electrical stimulation of the cutaneous nerve disrupts intraspinal proprioceptive neural trajectories

We next evaluated the impact of concurrent artificial cutaneous input on proprioceptive information processing. We projected on the neural manifold neural trajectories elicited by the stimulation of the proprioceptive branch. All four monkeys exhibited robust trajectories in response to proprioceptive inputs and, in all four monkeys, concurrent stimulation of cutaneous afferents significantly reduced the trajectory lengths (Figure 4A) or even completely disrupted their dynamics (Figure S4), albeit with different effect sizes. MK1 (relative mean difference, −66.03%), Mk-Hs (−27.47%), and MK4 (−44.91%) exhibited the largest disruption, while MK2 (−5.89%) was significantly disrupted but to a lower effect size.

Figure 4. Neural trajectory lengths.

(A) Comparison of the neural trajectories induced by muscle-nerve stimulation and concurrent cutaneous stimulation across PC2-PC3 vs. PC1-PC2. Gray dashed lines indicate the average trajectory for muscle- and cutaneous-nerve stimulation at a low amplitude.

(B) Averaged spike counts across all trials and all channels for each stimulation condition for MK1.

(C) Statistical analysis of the trajectory lengths for each stimulation condition. Violin plots: each dot corresponds to the computed trajectory length for a trial, forming a Gaussian distribution of trajectory lengths. The central mark represented as a white dot indicates the median, and the gray line indicates the 25th and 75th percentiles. The whiskers extend to the most extreme data points not considered outliers. Trial corresponds to a stimulation pulse (***p < 0.001; **p < 0.01; *p < 0.05; Kruskal-Wallis test with 381, 470, and 453 points for muscle-nerve stimulation and concurrent cutaneous stimulation at high amplitude and low amplitudes, respectively, for MK1; 369, 410, and 411 points, respectively, for MK2; 353, 376, and 397 points, respectively, for Mk-Hs; and 392, 380, and 371 points, respectively, for MK4).

(D) Top, a graphical representation of the busy-line effect. Left, neural network involved in the neural processing of muscle-nerve stimulation with sufficient capacity in the shared nodes. Right, neural networks involved in the processing of both muscle- and cutaneous-nerve stimulation, which results in a saturation state to process further neural input. Bottom, a representation of the spike counts for each case (data from Mk-Hs across PC1). Although there is an increase in spiking activity during concurrent stimulation, the neural activity is less modulated (i.e., saturated) than during muscle-nerve stimulation alone, which leads to trajectory shrinking.

(E) Quantification of the busy-line effect for each PC for each monkey.

(F) Quantification of the modulation depth computed as the difference between the maximum spike counts after stimulation and the following minimum. Square and error bars indicate the mean distribution of the data and the standard error of the mean. Statistical analysis of the spiking activity for each stimulation condition (***p < 0.001; **p < 0.01; *p < 0.05; Kruskal-Wallis test with 387, 477, and 461 points for muscle-nerve stimulation and concurrent cutaneous stimulation at high amplitude and low amplitude, respectively, for MK1; 374, 410, and 412 points, respectively, for MK2; 353, 400, and 399 points, respectively, for Mk-Hs; and 401, 388, and 376 points, respectively, for MK4).

To validate this result, we repeated the same experiment using lower amplitudes for the stimulation of the cutaneous afferents. Cutaneous stimulation at a low amplitude yielded less disruption (i.e., longer proprioceptive trajectory lengths) than at a high stimulation amplitude (Figure 4C), suggesting that the amount of neural computation disrupted is inversely proportional to the stimulation intensity of cutaneous afferents. This observed trajectory disruption is particularly interesting considering that, during concurrent proprioceptive and cutaneous stimulation, the spinal cord received significantly more artificial input. Indeed, concurrent stimulation of the cutaneous afferents significantly increased overall spike counts in the recorded spinal circuitries (+5,191.68% for MK1, +234.71% for MK2, +68.37% for Mk-Hs, and +754.94% for MK4; Figures 4B and S3). However, this increase was not captured by the neural trajectories, which strengthens the case that those trajectory lengths mainly represent proprioceptive information processing.

In addition, we found that the disruption of information processing was captured in principal component (PC) 2 and PC3, where neural trajectories shrank as a function of stimulation intensity. Moreover, PC1 depicted the displacement of these neural trajectories caused by the amount of concurrent cutaneous input (Figures 4A and S3). In other words, when the spinal cord received inputs induced by concurrent muscle- and cutaneous-nerve stimulation, the neural trajectories were displaced across PC1, away from the proprioceptive neural trajectories. This displacement was proportional to the stimulation intensity and, in turn, to the computed spiking activity in the spinal cord (Figure S3). In particular, MK1 (+94.52%) and MK2 (+45.99%) produced longer proprioceptive trajectory lengths than Mk-Hs (+3.49%) and MK4 (+5.22%) during concurrent cutaneous stimulation at a low amplitude. Indeed, the overall spike counts were very similar in Mk-Hs and MK4 at both low and high stimulation amplitudes (Figure S3, relative mean difference from concurrent high to low amplitude, −84.44% for MK1, −59.82% for MK2, −1.69% for Mk-Hs, and −8.40% for MK4), thereby eliciting similar neural trajectory lengths. These results infer that the main PCs clearly captured the amount of proprioceptive processed information as a function of concurrent stimulation amplitude. Indeed, neural trajectories that were further displaced across PC1 resulted in shorter neural trajectory lengths in PC2-PC3 (during concurrent high stimulation amplitude), whereas those that remained closer to the proprioceptive neural trajectories in PC1 were less disrupted in PC2-PC3 (during concurrent low stimulation amplitude; Figure 4A).

To explain these results, we hypothesized that some of the neurons that respond to both stimuli were now occupied by responses to continuous stimulation of the cutaneous afferents, which would reduce their ability to modulate in response to proprioceptive percepts, hence causing the trajectory collapse (Figure 4D). To validate this hypothesis, we extracted the spike counts of the probe channels that formed the first three PCs, i.e., the most relevant for neural trajectory formation. First, we noticed that their location was scattered across the dorsoventral profile of the probe, suggesting that the PCs were formed by neurons throughout the gray matter and not only from specific laminae (Figure S5). Second, we looked at the total spike counts of the PCs, and we found that the absolute firing of these channels substantially increased in response to the cutaneous stimulation (Figure 4E). Simultaneously, the modulation depth of neural responses to proprioceptive inputs was significantly reduced (black arrows, Figure 4F), suggesting that concurrent artificial recruitment of cutaneous afferents hinders the processing of proprioceptive inputs by reducing the ability of spinal circuits to respond to other concurrent inputs.

Cutaneous electrical stimulation reduced proprioceptive afferent volleys, spinal cord gray matter field potentials, and multiunit responses

To validate our findings, we looked for correlates using classical electrophysiology measures. We first inspected stimulation-triggered average field potentials from the gray matter of the spinal cord, defined as the mean neural response across each single muscle nerve branch stimulation pulse (Figure 5A). Afferent volleys were detected at a latency between 3 and 4 ms after each proprioceptive pulse (Figures S6 and 5A). Continuous electrical stimulation of the cutaneous nerve reduced the peak-to-peak amplitude of these proprioceptive volleys in all four monkeys (Figure S6), and the reduction was proportional to stimulation intensity (muscle-nerve stimulation vs. muscle- and cutaneous-nerve stimulation, high amplitude, mean value difference, MK1, −9%; MK2, −47%; Mk-Hs, −25%; MK4, −14%; muscle-nerve stimulation vs. muscle- and cutaneous-nerve stimulation, low amplitude, MK1, −8%; MK2, −40%; Mk-Hs, −22%; MK4, −1%). Since volleys represent sensory inputs, these results suggest that part of the disruption that we observed in the neural trajectories may be a consequence of reduced proprioceptive inputs in the spinal cord.

Figure 5. Peak-to-peak amplitude suppression of spinal cord gray matter response fields.

(A) MK1 triggered average signal showing afferent volley, and gray matter response fields resulting from muscle-nerve stimulation (cyan), with concurrent cutaneous-nerve stimulation (magenta; high stimulation amplitude, solid color; low stimulation amplitude, semi-transparent).

(B) Peak-to-peak amplitude of gray matter response field in four monkeys, dorsal channel examples. Color coding is the same as in (A). We compared peak-to-peak amplitude values over two conditions with one-way ANOVA with 300 points, where each point represents the peak-to-peak amplitude as a response to a single stimulus pulse. Boxplots: the central mark indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. The whiskers extend to the most extreme data points not considered outliers, and the outliers are plotted individually using open circles. ***p < 0.001.

In addition, local field potential (LFP) responses occurring after the volleys were also substantially suppressed during electrical stimulation. These fields reflect gray matter responses to incoming action potentials and are linked to both synaptic and network activity.⁴² Peak-to-peak amplitude values of these fields were significantly reduced to a much larger extent than the volleys. Again, the suppression correlated to stimulation intensity: high amplitude of cutaneous stimulation resulted in greater suppression of afferent volleys and gray matter response field peak-to-peak values than low stimulation amplitude (muscle-nerve stimulation vs. muscle- and cutaneous-nerve stimulation, high amplitude, mean value difference, MK1, −83%; MK2, −18%; Mk-Hs, −46%; MK4, −56%; muscle-nerve stimulation vs. muscle- and cutaneous-nerve stimulation, low amplitude, MK1, −48%; MK2, −15%; Mk-Hs, −42%; MK4, −43%; Figure 5B). The same trend was found in all four monkeys, suggesting that a significant component of the trajectory disruption may be related to reduced gray matter responses to proprioceptive volleys and not only to a simple reduction of proprioceptive inputs.

Finally, we investigated whether changes in the population neural dynamics and gray matter field potentials could be reflected in changes in single-neuron spiking activity. We utilized multiunit threshold crossing analysis (Figure 6A) and identified channels in which a clear response to proprioceptive pulses was visible. In this multiunit analysis, the peak of neural activity after proprioceptive stimuli occurred at approximately 3–4 ms after each proprioceptive stimulation pulse. We present the neural responses of units that were activated by proprioceptive inputs. When continuous stimulation of the cutaneous nerve was overlapped with muscle-nerve stimulation, we observed a reduction in these responses in all four monkeys, in both the dorsal and the ventral horn of the spinal cord (Figure 6B).

Figure 6. Multiunit activity.

(A) We filtered the signal to extract the spiking component and detected the neural action potentials using the thresholding algorithm (see STAR Methods).

(B) Examples of multiunit activity in two different channels (one in the dorsal, one in the ventral region) for each of the four monkeys. Single muscle-nerve stimulation (cyan, left) and concurrent muscle- and cutaneous-nerve stimulation at a high amplitude (magenta, right). Dashed cyan line represents the muscle- nerve stimulation pulse. Neural activity is presented and quantified with raster plots and peristimulus time histograms (PSTHs). Each row of the raster plots represents the response to a single muscle-nerve stimulation pulse, while each dot corresponds to an action potential. Mean event rate is defined as an average number of spikes within a time frame of one bin (0.2 ms) across all single pulses of muscle-nerve stimulation. Black lines highlight the PSTH bins that are reduced. Black arrows indicate the decreased mean event rate values of PSTH, and their lengths correspond to the amount of reduction. Diagonal lines correspond to the units whose frequency is in line with frequency of stimulation.

In summary, we found that concurrent cutaneous-nerve stimulation reduced peak-to-peak amplitude of afferent volleys, gray matter response fields, and multiunit responses to proprioceptive stimuli. These results suggest that proprioceptive information processing may be disrupted by reducing both sensory input in the spinal cord and gray matter network computations.

Reduction of proprioceptive processing affects somatosensory cortex

We showed that concurrent stimulation of cutaneous afferents suppresses proprioception information processing in the spinal cord and correlates to classic electrophysiology measures. We then hypothesized that this suppression in the spinal cord limits the amount of information transmitted upstream to the brain, which could impact conscious perception of proprioception.

To test this hypothesis, we analyzed intracortical neural signals extracted from area 2 of the somatosensory cortex in all four monkeys. We found cortical evoked potentials with a latency of around 22–25 ms, which is consistent with the longer distance between the cortex and the peripheral nerves and has also been reported in similar experiments.⁴³ Peak-to-peak analysis of the signal amplitude indicated results similar to those in the spinal cord. We observed a reduction of proprioceptive evoked potentials during concurrent high-amplitude stimulation of the cutaneous nerve in all monkeys (Figure 7A).

Figure 7. Somatosensory cortex evoked potentials and cortical spiking activity.

(A) Somatosensory cortex evoked potentials in four monkeys. Examples of signals recorded as a response to muscle-nerve stimulation, with concurrent cutaneous-nerve stimulation (magenta; high stimulation amplitude, solid color; low stimulation amplitude, semi-transparent) or without it (cyan). Evoked potentials appeared with a latency between 22 and 25 ms. Signals are given as an example of a single channel in the somatosensory cortex and are averaged across all muscle-nerve stimulation pulses. We compared peak-to-peak amplitude values of the signal over two conditions with one-way ANOVA with 300 points, where each point represents the peak-to-peak amplitude as a response to a single stimulus pulse.

(B) Spike counts were averaged across all trials and all channels for each stimulation condition for all monkeys.

(C) Statistical analysis of the spiking activity for each stimulation condition (Kruskal-Wallis test with 392, 474, and 460 points for muscle-nerve stimulation and concurrent cutaneous stimulation at high amplitude and low amplitude, respectively, for MK1; 360, 389, and 392 points, respectively, for MK2; 335, 386, and 390 points, respectively, for Mk-Hs; and 391, 383, and 375 points, respectively, for Mk-Hs). Each dot corresponds to the computed trajectory length for a trial, forming a Gaussian distribution of trajectory lengths. Trial corresponds to a stimulation pulse. Boxplots and violin plots: the central mark indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. The whiskers extend to the most extreme data points not considered outliers, and the outliers are plotted individually using open circles. ***p < 0.001; **p < 0.01; *p < 0.05.

Observed suppression was detected in most of the channels in the array. Moreover, when we stimulated the cutaneous nerve at a low amplitude, peak-to-peak values of the signal increased (muscle-nerve stimulation vs. muscle- and cutaneous-nerve stimulation, high amplitude, mean value difference, MK1, −8%; MK2, −19%; Mk-Hs, −30%; MK4, −29%; muscle-nerve stimulation vs. muscle- and cutaneous-nerve stimulation, low amplitude, MK1, +2%; MK2, −18%; Mk-Hs, +3%; MK4, −26%).

Surprisingly, when we inspected the spiking activity extracted from multiunits in the cortex, the spike counts induced by concurrent cutaneous-nerve stimulation at a low amplitude were similar to, or even greater than, those obtained at a high amplitude (relative mean difference from concurrent high to low amplitude, +53.69% for MK1, −7.65% for MK2, +3.14% for Mk-Hs, and −6.40% for MK4; Figures 7B and 7C). This is markedly different from what we observed in the spinal cord (Figures S2 and S3). Indeed, we expected greater spiking activity consistently associated with higher stimulation amplitudes and not the opposite. In fact, this discrepancy seemed to reflect the spinal proprioceptive information processing, where concurrent cutaneous stimulation at a low amplitude yielded longer neural trajectory lengths.

In summary, we observed a reduction of proprioceptive information during concurrent continuous stimulation of the cutaneous nerve also in the somatosensory cortex. This finding suggests that the effects that we observed in the spinal cord propagate through the higher layers of sensorimotor processing.

DISCUSSION

In this experimental study in four monkeys, we combined analysis of neural population dynamics with classical electrophysiology measures to analyze the impact of continuous electrical stimulation of the cutaneous afferents on the processing of proprioceptive information in the spinal cord. We found that the spinal sensorimotor responses to proprioceptive inputs were substantially disrupted when cutaneous afferents were concurrently stimulated and that this interference propagated to the brain. While limited to the stimulation of the cutaneous afferents, our findings suggest that artificially generated neural input may disrupt ongoing neural processes that may be unrelated to the stimulation. More specifically, because of the highly shared neural architecture, even highly selective targeting of neural elements, like, in our case, the cutaneous afferents, can significantly undermine neural processing of seemingly unrelated neural functions, like proprioceptive percepts. Similar phenomena may occur in other regions of the nervous system and should therefore be studied. Hence, these results imply that efforts toward the development of naturalistic or biomimetic stimulation inputs should probably be employed in neurostimulation.

Population analysis as a tool to explain network-level effects of electrical stimulation

Electrical stimulation of the nervous system is widely applied in clinical practice and in clinical research trials to influence neural activity and ameliorate functions in a variety of disorders.^1,2,4,6,9 The most overt applications are sensorimotor neuroprostheses, where a clear relationship can be found between stimulation parameters and strength of elicited movements^3,4,6,44 or evoked sensations.^1,45–48 For example, epidural spinal cord stimulation has been applied for both motor recovery and, more recently, restoration of sensory feedback. We know that spinal cord stimulation recruits sensory afferents. In motor applications, the recruitment of large proprioceptive afferents leads to an increased excitability of spinal motoneurons, thereby promoting movement. Instead, in sensory applications, the recruitment of the same afferents should in principle produce controllable conscious sensory experiences, similar to those elicited by stimulation of the peripheral nerve. However, beyond this simplistic vision, there is a fundamental lack of knowledge of what happens to neural networks that receive inputs from these afferents. In fact, the highly shared neural infrastructures involve neural sub-networks that are meant to produce the desired function.^41,49,50 For instance, the motor network produces movement, but other networks may be involved in other unrelated processes, such as perception, error estimation during movement execution, and autonomic function,⁵¹ among others.⁴⁹ These additional sub-networks may also share inputs from the same afferents and would thus be perturbed by stimulation.

In our work, we constructed a model to study this specific problem in a controlled fashion as a proxy to understand, more generally, how artificial inputs can influence neural network function. We focused on the spinal network effects caused by stimulation of the cutaneous afferents on the neural processing of a proprioceptive pulse. This model exemplifies that an ongoing neural process (proprioceptive input processing) is perturbed when seemingly unrelated electrical stimuli (inducing touch percepts) are applied with a typical stimulation pattern (fixed 50 Hz square pulses). To explore network effects, we used population analysis tools that enable the quantification of information processing in the spinal circuits via analysis of neural trajectories in the neural space.³³ Specifically, we established a measure of proprioceptive information processing by quantifying neural trajectory lengths in spinal neural manifolds using intraspinal population analysis.³⁶ Through the quantification of the trajectory length, we assessed the effect of concurrent cutaneous stimulation on the spinal proprioceptive information. We showed the collapse of proprioceptive neural trajectories during concurrent stimulation of cutaneous afferents, in other words, a suppression of proprioceptive information processed in the spinal cord, according to our interpretation. Importantly, simple analysis of total spike counts showed that the results of our trajectory length quantification were not trivial. Indeed, while multiunit spiking activity was higher during concurrent stimulation of the cutaneous afferents, the modulation depth in response to proprioceptive percepts was significantly reduced. Thus, the actual total neural activity in the spinal cord is higher during cutaneous stimulation. Yet, the population analysis allows us to extract only activity that explains the variance generated by proprioceptive input processes, enabling one to infer the proprioceptive components of the neural dynamics against the background of cutaneous activity. Hence, the use of neural manifolds for population activity enabled the quantification of the stimulation effects on these computations.

We validated results obtained with neural manifold analysis with classical electrophysiology, inspecting peak-to-peak amplitude of afferent volleys, spinal cord gray matter response fields, and multiunit activity. These measures indicated a reduction in proprioceptive information during concurrent cutaneous stimulation.

Potential underlying neural mechanisms

While successful in visualizing network effects, population analysis cannot offer an explanatory value on the specific neural mechanisms responsible for this suppression. Presynaptic inhibition is a likely candidate.⁵² It is a well-known mechanism of sensory input gating that prevents transmission of excitatory postsynaptic potentials to neurons targeted by primary afferents.^39,40 In our experiments, the stimulation amplitude of the muscle nerve was the same across all conditions (i.e., fixed number of recruited afferents). Therefore, the reduction in unit responses to proprioceptive inputs during concurrent cutaneous afferent stimulation could be consistent with a reduction in synaptic inputs to these target units.

Nevertheless, the observed afferent volley reduction was minimal and contrasted with the complete disruption of neural trajectories. This means that the disruption of neural trajectories was not caused only by reduced inputs, but also by reduced neural processing. We refer to this other potential mechanism as the ‘‘busy-line’’ effect. Continuous, non-natural stimulation of the cutaneous afferents produced highly synchronized activity in spinal circuits that significantly increased the firing rates of units involved in producing proprioceptive trajectories (Figures 4D and 4E). However, when artificially synchronized cutaneous inputs reach the spinal cord, they may saturate these circuits and reduce their capacity to respond to additional inputs.⁵³ Indeed, modulation depth in response to proprioceptive inputs was significantly reduced in these network elements (Figure 4F), suggesting that these neurons, occupied in responding to artificial cutaneous inputs, could not be employed to process proprioceptive information as the neural network achieves a saturated state where no further processing can be carried out.

Effects within the brain

We performed a large part of our analysis in the spinal cord, which is the first important layer of sensory processing, particularly in regard to proprioception. However, conscious perception is processed at various layers above the spinal cord. Indeed, peak-to-peak amplitudes of cortex potentials evoked with muscle-nerve stimulation were suppressed when overlapped with cutaneous input also in sensory cortex area 2, which is known to integrate cutaneous and proprioceptive inputs.^49,54 Moreover, if cortical signals were independent of spinal and brain-stem processes, when looking at the global cortical spike counts, we would have expected higher spike counts during high-amplitude stimulation of the cutaneous nerve and lower spiking activity during low-amplitude stimulation of the cutaneous nerve. Instead, we found higher or similar spiking activity when we used lowamplitude stimulation of the cutaneous nerve. This may be indicative of the fact that high-amplitude stimulation may convey more cutaneous input but less proprioceptive input to the cortex because of sub-cortical cancellation.³⁹ In contrast, cutaneous stimulation at a lower amplitude may mean less cutaneous input but more proprioceptive input to the cortex as a consequence of less cancellation occurring in sub-cortical structures.

These overall results support the conclusion that conscious perception of proprioception may be also altered by sub-cortical interference. While this hypothesis cannot be tested in subjects with amputation because of their limb loss, recent data in humans with incomplete sensory spinal cord injury show that spinal cord stimulation, which also recruits sensory afferents,⁵⁵ reduces proprioception acuity during suprathreshold stimulation.⁶ This result in humans further supports our hypothesis, and we believe that it demands further investigation.

Limitations of the study

Our methods directly assess population-level effects and suggest a mechanism by which single neurons become saturated by artificial stimuli, preventing them from responding to other inputs. However, our experiments are not geared toward dissecting neuron-specific effects. Nevertheless, analysis of widespread effects across the probe suggests that these interferences do not occur only at specific laminae, but we still cannot distinguish the specific class of neurons involved in these processes. Finally, our brain data analysis is limited by the specific area we recorded from. Additional brain regions such as Brodmann area 3a that respond solely to proprioceptive input could complement and expand our analysis to validate our findings. While we do not expect that the main research outcomes are in any way dependent on the animal’s sex, our methods were not focused on exploring its influence on the results.

Conclusions and relevance for other brain circuits

Our results showed that electrical stimulation of sensory afferents can alter the processing of proprioceptive information within spinal circuits. Similar phenomena may occur in brain networks during deep brain stimulation, where similar continuous electrical pulses are delivered to thalamocortical projections and other large brain networks. In the brain, these effects, which in the spinal cord indicate the impossibility of appropriately processing proprioception, could potentially alter cognitive processes unrelated to the stimulation goals within the cortex. A potential approach to minimize the interference of stimulation with ongoing neural processes is the use of ‘‘biomimetic’’ and model-based stimulation patterns.^{23,46,56–58} Instead of delivering unstructured and synchronized neural activity, these protocols could produce more naturalistic patterns, thereby potentially avoiding these side effects. In conclusion, future stimulation strategy designs should consider the use of neural population analysis to analyze the effects of particular stimulation patterns on apparently unrelated neural network processes.

STAR★METHODS

RESOURCE AVAILABILITY

Lead contact

Further information and requests should be directed to and will be fulfilled by the lead contact, Marco Capogrosso (mcapo@pitt.edu).

Materials availability

This study did not generate any new materials.

Data and code availability

1. All animal data reported in this paper will be shared by the lead contact upon request.

2. All original code has been deposited at Mendeley Data and is publicly available as of the date of publication. DOI is listed in the key resources table.

3. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, peptides, and recombinant proteins
Propofol	Covetrus	059333; CAS: 2078–54-8
Fentanyl	Covetrus	055012; CAS: 990–73-8
Deposited data
Experimental data and code	This study	https://doi.org/10.17632/f2khpckd8k.1 Link: https://data.mendeley.com/datasets/f2khpckd8k/1
Experimental models: Organisms/strains
Macaca Fascicularis	Kurchatov Complex for Medical Primatology, National Research Center ‘‘Kurchatov Institute’’	http://kcsni.nrcki.ru/en.shtml
Macaca Mulatta	Southwest National Primate Research Center	https://snprc.org/
Software and algorithms
MATLAB R2019b	The Mathworks Inc.	https://www.mathworks.com
Trellis version 1.14	Ripple Neuro	https://rippleneuro.com/
RHX Data Acquisition Software	Intan Technologies	https://intantech.com/index.html
Other
Cuff electrode	Microprobes for Life Science	https://microprobes.com/
Cuff electrode	Micro-Leads	https://www.micro-leads.com
UTAH array electrode	Blackrock Microsystems	https://blackrockneurotech.com/
Pneumatic inserter	Blackrock Microsystems	https://blackrockneurotech.com/
Linear Probe with Omnetics Connector 32 pins - A1×32–15mm-50–177-CM32	NeuroNexus	https://neuronexus.com/
Double linear Probe with Omnetics Connector 64 pins - A2×32–15mm100–200-177	NeuroNexus	https://neuronexus.com/
MM-3 micromanipulators	Narishige	https://usa.narishige-group.com/
MM-3 micromanipulators	David Koff Instruments	https://kopfinstruments.com/
Stereotactic frame Model 1530	David Koff Instruments	https://kopfinstruments.com/
Ripple Neuro Grapevine Neural Interface 1.14.3	Ripple Neuro	https://rippleneuro.com/
RHS Stim/Recording System	Intan Technologies	https://intantech.com/index.html
AM stimulator model 2100	A-M Systems	https://www.a-msystems.com/
Subdermal needle electrode 7mm & 13mm	Rhythmlink	https://rhythmlink.com/

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the local (Research Institute of Medical Primatology) Institutional Ethics Committee (protocol No 38/1, October 31, 2019) and by the University of Pittsburgh Animal Research Protections and IACUC (ISOOO17081).

Three adult Macaca Fascicularis and one Macaca Mulatta monkeys were involved in the study (MK1 - MK 42286, male, 4 years old, 3.5 kg, MK2 - MK 42588, male, 4 years old, 3.35 kg, MK4 - MK 42328, male, 4 years old, 3.48 kg; Mk-Hs – 219–21, male, 7 years old, 11.5 kg). We did not investigate any influence on animals’ sex on the results. Data for all Macaca Fascicularis monkeys were acquired in the Kurchatov Complex for Medical Primatology, National Research Center ‘‘Kurchatov Institute’’ (former name is Research Institute of Medical Primatology), Sochi, Russia, where they were housed. Data for Macaca Mulatta monkey was acquired in the University of Pittsburgh, PA, US and was housed in the primate facility at the Division of Laboratory Animal Resources at the University of Pittsburgh. All animals had unrestricted access to water and food as well as daily enrichments.

METHOD DETAILS

Surgical procedures

All the surgical procedures were performed under full anesthesia induced with ketamine (10 mg/kg, i.m.) and maintained under continuous intravenous infusion of propofol (1% solution in 20 mL Propofol/20 mL Ringer 1.8 to 6 mL/kg/h), in addition to fentanyl (6–42 μg/kg/h) for the Macaca Mulatta, using standard techniques. Throughout the procedures, the veterinary team continuously monitored the animal’s heart rate, respiratory rate, oxygen saturation level and temperature. Surgical implantations were performed during a single operation lasting approximately 8 h. We fixed monkeys’ heads in a stereotaxic frame securing the cervical spine in a prone and flat position. First, we implanted two silicon cuff electrodes (Microprobes for Life Science, Gaithersburg, MD 20879, U.S.A. and Micro-Leads, Somerville, MA 02144, U.S.A.) on the distal ends of the superficial branch and deep branch of radial nerve that we determined via anatomical landmarks. We then inserted EMG electrodes in the Extensor Digit. Communis, the Flexor Carpi Radialis and the Flexor Digit. Superficialis. We stimulated electrically two branches of the radial nerve and looked at the EMG response to verify which branch was the muscle branch and which one was the cutaneous branch. Second, we implanted the brain array using a pneumatic insertion system (Blackrock Microsystem). We performed a craniotomy and we incised the dura in order to get clear access to the central sulcus. We identified motor and sensory brain areas through anatomical landmarks and intra-surgical micro-stimulation. Specifically, we verified that electrical stimulation of the motor cortex induced motor responses in the hand muscles (Figure S1A). We then determined the position of the somatosensory area S1 in relation to this spot and implanted the UTAH array electrode (Blackrock Microsystems, Salt Lake City, UT, U.S.A.) across Areas 1 and 2 (and Areas 3 and 4 for the Mulatta monkey), 1.2 mm lateral to midline and 3.1 mm deep using a pneumatic inserter (Blackrock Microsystems, Salt Lake City, UT, U.S.A.).

Finally, we performed a laminectomy from C3 to T1 and then directly exposed the cervical spinal cord. We implanted a 32-channel linear probe (linear Probe with Omnetics Connector 32 pins - A1×32-15mm-50-177-CM32; NeuroNexus, Ann Arbor, MI, U.S.A.) and a 64-channel linear probe (double linear Probe with Omnetics Connector 64 pins - A2×32-15mm-100-200-177; NeuroNexus, Ann Arbor, MI, U.S.A.) in the gray matter at the C5 spinal segment. To implant the probe, we opened the dura mater and created a small hole in the pia using a surgical needle through which penetration of the probe with micromanipulators was possible. We implanted the arrays using MM-3 micromanipulators (Narishige, Tokyo, Japan; David Koff Instruments for the Mulatta monkey). Experiments in all four monkeys were terminal. At the end the animals were euthanized with a single injection of penthobarbital (60 mg/kg) and perfused with PFA for further tissue processing.

Data acquisition and electrophysiology in sedated monkeys

Monkeys were sedated with a continuous intravenous infusion of propofol that minimizes effects on spinal cord stimulation.⁵⁹ Biphasic, charge balanced, symmetric square pulses (with pulse width of 0.5 ms for the Macaca Fascicularis and 0.1 ms for the Macaca Mulatta) were delivered using an AM stimulator (model 2100 A-M Systems, Sequim, WA, USA). Electromyographic and neural signals were recorded at 30 kHz sampling frequency (RHS recording system with 32-channel headstages, Intan Technologies, Los Angeles, CA, USA for Macaca Fascicularis and Ripple Neuro Grapevine Neural Interface 1.14.3. for Macaca Mulatta monkey).

Data analysis

We applied all data analysis techniques offline, specifically, these analyses were performed offline in MATLAB (version R2019b).

Pre-processing

We filtered raw signals recorded with 32 - electrode array implanted in the spinal cord, as well as signals documented with UTAH array in somatosensory cortex with comb filter to remove artifacts on 50 Hz/60 Hz (depending on the country where the experiments have been done) and its harmonics. We designed a digital infinite impulse response filter as a group of notch filters that are evenly spaced at exactly 50 Hz/60 Hz.

We detected single pulses of the deep branch of the radial nerve and extracted 430 ms of the inta-spinal and intra-cortical signal post stimulation.

Identification of sensory volleys resulting from muscle nerve stimulation

We were able to detect afferent volleys and the resulting gray matter response field evoked with muscle nerve stimulation in the spinal cord. We applied a 3rd order Butterworth digital filter and extracted the signal from 10 to 1000 Hz. Afferent volley is defined as a first volley after the stimulation pulse, occurring 3–4 ms after the stimulation (unique physiology of a single animal causes these variations) and followed with gray matter response field. We quantified the amount of processed proprioceptive information by neural network by measuring peak-to-peak amplitude values of the gray matter response field.

We applied a similar procedure to extract the muscle nerve evoked potentials recorded in the somatosensory cortex.

Characterization and quantification of neural spiking activity

We extracted neural spiking activity by applying a 3rd order Butterworth digital filter to the raw signal, separating the signal in frequency range from 800 Hz to 5000 Hz. We detected the spikes using thresholding algorithm.⁶⁰ We determined the threshold value separately for each recording channel. To detect the accurate threshold value, we concatenated all datasets that we aim to analyze in a single file. All analyzed datasets were concatenated in a single file in order to detect proper threshold values. The same procedure was applied to intra-spinal and intra-cortical recordings.

Multiunit activity is presented in form of rasterplot and quantified with peri-stimulus time histogram (PSTH). Each dot in rasterplot represents a single detected spike. Every rasteplot row corresponds to the intra-spinal or intra-cortical activity perturbed with a single muscle nerve stimulus pulse. PSTH is quantified with mean event rate, defined as the average number of spikes across all single pulses of muscle nerve stimulation, within defined time frame.

Neural manifold and trajectory length

To project the trajectories in the neural manifold, we previously computed multiunit spiking activity for each condition. We calculated the spiking activity for every 100 ms with a sliding window of 10 ms over 430 ms around each muscle stimulation pulse. We zero-padded the first repetition for 90 ms and then overlapped 90ms from the previous repetition for the rest of repetitions. The final step to smooth the spiking activity was the application of a Gaussian kernel (s.d. 20 ms) to the binned square-root-transformed firings (10 ms bin size) of each recorded multiunit. For each condition, this resulted in a matrix of dimensions C x T, where C is the number of channels in the dorsoventral linear probe and T is the number of 10 ms windows in a repetition concatenated for all the repetitions within a condition. Subsequently, we proceed to eliminate noisy repetitions. We discarded those repetitions within each condition whose s.d. was greater than twice the total s.d. across all repetitions plus the total mean of the s.d. across all repetitions for that condition. For cortical data, we previously converted the distribution of s.d. to a lognormal distribution to apply this outlier cleaning rule.

To calculate the latent dynamics for each monkey, we z-scored each condition’s spiking activity before applying dimensionality reduction principal component analysis (PCA) to the concatenated spike counts. We selected the first 3 principal components that explained most of the variance (~65% for all 3 monkeys, 54% for one monkey) as neural modes to define the neural manifold. Convergence points were reached at the first 3 to 5 dimensions according to the eigenspectrum of each monkey. In this low dimensionality space, we proceeded by eliminating repetitions as a function of the distance to the median trajectory. In particular, we computed the median trajectory for each 10 ms window for each condition. For each window, we calculated the distance between the median trajectory and the trajectory elicited by each repetition within a condition. 25th and 75th percentiles of the obtained distances allowed to discard trajectories whose distance was greater than the 75th percentile plus 1.5 times the interquartile range of the averaged trajectory for that repetition across all 10 ms windows. The same criterion was applied for the lower range. Finally, we quantified the trajectory length for the remaining repetitions for each condition and calculated the average trajectory length across all 10 ms windows.

QUANTIFICATION AND STATISTICAL ANALYSIS

The normality of data distributions was verified with a one-sample Kolmogorov-Smirnov test, using MATLAB function ‘‘kstest’’. The function returns a test decision for the null hypothesis that the given data comes from a standard normal distribution, against the alternative that it does not come from such a distribution. The result is 1 if the test rejects the null hypothesis at the 5% significance level, or 0 otherwise. We used quantile-quantile plot (QQ-plot) for visual inspection of normality, using MATLAB function ‘‘qqplot’’. QQ plot quantiles of the data versus the theoretical quantile values from a normal distribution. If the distribution of the data is normal, the plot appears linear. In case of Gaussian distribution, one-way analysis of variance (ANOVA) was applied, using MATLAB function ‘‘anova1’’. Elsewise, we performed Kruskal-Wallis test for data that has two or more groups. Post-hoc correction was executed in case of multiple groups of data.

Multi-group significance comparison of data obtained from the neural manifold for each condition in all four monkeys was tested using Kruskal-Wallis test. The level of significance was set at ***p < 0.005.

Significance of suppressed peak-to-peak amplitude values of afferent volleys was analyzed with one-way analysis of variance revealed (ANOVA). Each point represents the peak-to-peak amplitude as a response to a single stimulus pulse. Boxplots show: the central mark indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. The whiskers extend to the most extreme data points not considered outliers, and the outliers are plotted individually using the ‘o’ symbol. The level of significance was set at ***p < 0.001, **p < 0.01 and *p < 0.05.

All statistical tests used, exact value of n and its definition, together with explained measures used for drawing conclusions are described in figure legends and in the Results section of the manuscript. Levels of significance are labeled on the figures.

Highlights.

• We study the impact of electrical cutaneous stimulation on proprioceptive input processing

• Spinal responses to proprioceptive inputs are disrupted by stimulation of cutaneous axons

• The disruption of proprioceptive responses also occurs in the brain

• Electrical stimulation can perturb neural responses of unrelated neural processes