Multiunit Activity-Based Real-Time Limb-State Estimation from Dorsal Root Ganglion Recordings

A Scoping Review of Machine Learning Applied to Peripheral Nerve Interfaces

Peripheral nerve interfaces (PNIs) can enable communication with the peripheral nervous system and have a broad range of applications including in bioelectronic medicine and neuroprostheses. They can modulate neural activity through stimulation or monitor conditions by recording from the peripheral nerves. The recent growth of Machine Learning (ML) has led to the application of a wide variety of ML techniques to PNIs, especially in circumstances where the goal is classification or regression. However, the extent to which ML has been applied to PNIs or the range of suitable ML techniques has not been documented. Therefore, a scoping review was conducted to determine and understand the state of ML in the PNI field. The review searched five databases and included 63 studies after full-text review. Most studies incorporated a supervised learning approach to classify activity, with the most common algorithms being some form of neural network (artificial neural network, convolutional neural network or recurrent neural network). Unsupervised, semi-supervised and reinforcement learning (RL) approaches are currently underutilized and could be better leveraged to improve performance in this domain.

Decoding Bilateral Hindlimb Kinematics From Cat Spinal Signals Using Three-Dimensional Convolutional Neural Network

To date, decoding limb kinematic information mostly relies on neural signals recorded from the peripheral nerve, dorsal root ganglia (DRG), ventral roots, spinal cord gray matter, and the sensorimotor cortex. In the current study, we demonstrated that the neural signals recorded from the lateral and dorsal columns within the spinal cord have the potential to decode hindlimb kinematics during locomotion. Experiments were conducted using intact cats. The cats were trained to walk on a moving belt in a hindlimb-only condition, while their forelimbs were kept on the front body of the treadmill. The bilateral hindlimb joint angles were decoded using local field potential signals recorded using a microelectrode array implanted in the dorsal and lateral columns of both the left and right sides of the cat spinal cord. The results show that contralateral hindlimb kinematics can be decoded as accurately as ipsilateral kinematics. Interestingly, hindlimb kinematics of both legs can be accurately decoded from the lateral columns within one side of the spinal cord during hindlimb-only locomotion. The results indicated that there was no significant difference between the decoding performances obtained using neural signals recorded from the dorsal and lateral columns. The results of the time-frequency analysis show that event-related synchronization (ERS) and event-related desynchronization (ERD) patterns in all frequency bands could reveal the dynamics of the neural signals during movement. The onset and offset of the movement can be clearly identified by the ERD/ERS patterns. The results of the mutual information (MI) analysis showed that the theta frequency band contained significantly more limb kinematics information than the other frequency bands. Moreover, the theta power increased with a higher locomotion speed.

Control of Mammalian Locomotion by Somatosensory Feedback

When animals walk overground, mechanical stimuli activate various receptors located in muscles, joints, and skin. Afferents from these mechanoreceptors project to neuronal networks controlling locomotion in the spinal cord and brain. The dynamic interactions between the control systems at different levels of the neuraxis ensure that locomotion adjusts to its environment and meets task demands. In this article, we describe and discuss the essential contribution of somatosensory feedback to locomotion. We start with a discussion of how biomechanical properties of the body affect somatosensory feedback. We follow with the different types of mechanoreceptors and somatosensory afferents and their activity during locomotion. We then describe central projections to locomotor networks and the modulation of somatosensory feedback during locomotion and its mechanisms. We then discuss experimental approaches and animal models used to investigate the control of locomotion by somatosensory feedback before providing an overview of the different functional roles of somatosensory feedback for locomotion. Lastly, we briefly describe the role of somatosensory feedback in the recovery of locomotion after neurological injury. We highlight the fact that somatosensory feedback is an essential component of a highly integrated system for locomotor control. © 2021 American Physiological Society. Compr Physiol 11:1-71, 2021.

Decoding hindlimb kinematics from descending and ascending neural signals during cat locomotion

OBJECTIVE

The main objective of this research is to record both sensory and motor information from the ascending and descending tracts within the spinal cord for decoding the hindlimb kinematics during walking on the treadmill.

APPROACH

Two different experimental paradigms (i.e., active and passive) were used in the current study. During active experiments, five cats were trained to walk bipedally while their hands kept on the front frame of the treadmill for balance or to walk quadrupedally. During passive experiments, the limb was passively moved by the experimenter. Local field potential (LFP) activity was recorded using a microwire array implanted in the dorsal column (DC) and lateral column (LC) of the L3-L4 spinal segments. The amplitude and frequency components of the LFP formed the feature set and the elastic net regularization was used to decode the hindlimb joint angles.

MAIN RESULTS

The results show that there is no significant difference between the information content of the signals recorded from the DC and LC regions during walking on the treadmill, but the information content of the DC is significantly higher than that of the LC during passively applied movement of the hindlimb in the anesthetized cats. Moreover, the decoding performance obtained using the recorded signals from the DC is comparable with that from the LC during locomotion. But, the decoding performance obtained using the recording channels in the DC is significantly better than that obtained using the signals recorded from the LC. The long-term analysis shows that robust decoding performance can be achieved over 2-3 months without a significant decrease in performance.

SIGNIFICANCE

This work presents a promising approach to developing a natural and robust motor neuroprosthesis device using descending neural signals to execute the movement and ascending neural signals as the feedback information for control of the movement.

Providing low-dimensional feedback of a high-dimensional movement allows for improved performance of a skilled walking task

Learning a skilled movement often requires changing multiple dimensions of movement in a coordinated manner. Serial training is one common approach to learning a new movement pattern, where each feature is learned in isolation from the others. Once one feature is learned, we move on to the next. However, when learning a complex movement pattern, serial training is not only laborious but can also be ineffective. Often, movement features are linked such that they cannot simply be added together as we progress through training. Thus, the ability to learn multiple features in parallel could make training faster and more effective. When using visual feedback as the tool for changing movement, however, such parallel training may increase the attentional load of training and impair performance. Here, we developed a novel visual feedback system that uses principal component analysis to weight four features of movement to create a simple one-dimensional 'summary' of performance. We used this feedback to teach healthy, young participants a modified walking pattern and compared their performance to those who received four concurrent streams of visual information to learn the same goal walking pattern. We demonstrated that those who used the principal component-based visual feedback improved their performance faster and to a greater extent compared to those who received concurrent feedback of all features. These results suggest that our novel principal component-based visual feedback provides a method for altering multiple features of movement toward a prescribed goal in an intuitive, low-dimensional manner.

Estimation of Bladder Pressure and Volume from the Neural Activity of Lumbosacral Dorsal Horn Using a Long-Short-Term-Memory-based Deep Neural Network

In this paper, we propose a deep recurrent neural network (DRNN) for the estimation of bladder pressure and volume from neural activity recorded directly from spinal cord gray matter neurons. The model was based on the Long Short-Term Memory (LSTM) architecture, which has emerged as a general and effective model for capturing long-term temporal dependencies with good generalization performance. In this way, training the network with the data recorded from one rat could lead to estimating the bladder status of different rats. We combined modeling of spiking and local field potential (LFP) activity into a unified framework to estimate the pressure and volume of the bladder. Moreover, we investigated the effect of two-electrode recording on decoding performance. The results show that the two-electrode recordings significantly improve the decoding performance compared to single-electrode recordings. The proposed framework could estimate bladder pressure and volume with an average normalized root-mean-squared (NRMS) error of 14.9 ± 4.8% and 19.7 ± 4.7% and a correlation coefficient (CC) of 83.2 ± 3.2% and 74.2 ± 6.2%, respectively. This work represents a promising approach to the real-time estimation of bladder pressure/volume in the closed-loop control of bladder function using functional electrical stimulation.

Computationally efficient fully-automatic online neural spike detection and sorting in presence of multi-unit activity for implantable circuits

BACKGROUND

Spike sorting is a basic step for implantable neural interfaces. With the growing number of channels, the process should be computationally efficient, automatic,robust and applicable on implantable circuits.

NEW METHOD

The proposed method is a combination of fully-automatic offline and online processes. It introduces a novel method for automatically determining a data-aware spike detection threshold, computationally efficient spike feature extraction, automatic optimal cluster number evaluation and verification coupled with Self-Organizing Maps to accurately determine cluster centroids. The system has the ability of unsupervised online operation after initial fully-automatic offline training. The prime focus of this paper is to fully-automate the complete spike detection and sorting pipeline, while keeping the accuracy high.

RESULTS

The proposed system is simulated on two well-known datasets. The automatic threshold improves detection accuracies significantly( > 15%) as compared to the most common detector. The system is able to effectively handle background multi-unit activity with improved performance.

COMPARISON

Most of the existing methods are not fully-automatic; they require supervision and expert intervention at various stages of the pipeline. Secondly, existing works focus on foreground neural activity. Recent research has highlighted importance of background multi-unit activity, and this work is amongst the first efforts that proposes and verifies an automatic methodology to effectively handle them as well.

CONCLUSION

This paper proposes a fully-automatic, computationally efficient system for spike sorting for both single-unit and multi-unit spikes. Although the scope of this work is design and verification through computer simulations, the system has been designed to be easily transferable into an integrated hardware form.

Dependence of Neuroprosthetic Stimulation on the Sensory Modality of the Trigeminal Neurons Following Nerve Injury. Implications in the Design of Future Sensory Neuroprostheses for Correct Perception and Modulation of Neuropathic Pain

Amputation of a sensory peripheral nerve induces severe anatomical and functional changes along the afferent pathway as well as perception alterations and neuropathic pain. In previous studies we showed that electrical stimulation applied to a transected infraorbital nerve protects the somatosensory cortex from the above-mentioned sensory deprivation-related changes. In the present study we focus on the initial tract of the somatosensory pathway and we investigate the way weak electrical stimulation modulates the neuroprotective-neuroregenerative and functional processes of trigeminal ganglia primary sensory neurons by studying the expression of neurotrophins (NTFs) and Glia-Derived Neurotrophic Factors (GDNFs) receptors. Neurostimulation was applied to the proximal stump of a transected left infraorbitary nerve using a neuroprosthetic micro-device 12 h/day for 4 weeks in freely behaving rats. Neurons were studied by in situ hybridization and immunohistochemistry against RET (proto-oncogene tyrosine kinase "rearranged during transfection"), tropomyosin-related kinases (TrkA, TrkB, TrkC) receptors and IB4 (Isolectin B4 from Griffonia simplicifolia). Intra-group (left vs. right ganglia) and inter-group comparisons (between Control, Axotomization and Stimulation-after-axotomization groups) were performed using the mean percentage change of the number of positive cells per section [100∗(left-right)/right)]. Intra-group differences were studied by paired t-tests. For inter-group comparisons ANOVA test followed by post hoc LSD test (when P < 0.05) were used. Significance level (α) was set to 0.05 in all cases. Results showed that (i) neurostimulation has heterogeneous effects on primary nociceptive and mechanoceptive/proprioceptive neurons; (ii) neurostimulation affects RET-expressing small and large neurons which include thermo-nociceptors and mechanoceptors, as well as on the IB4- and TrkB-positive populations, which mainly correspond to non-peptidergic thermo-nociceptive cells and mechanoceptors respectively. Our results suggest (i) electrical stimulation differentially affects modality-specific primary sensory neurons (ii) artificial input mainly acts on specific nociceptive and mechanoceptive neurons (iii) neuroprosthetic stimulation could be used to modulate peripheral nerve injuries-induced neuropathic pain. These could have important functional implications in both, the design of effective clinical neurostimulation-based protocols and the development of neuroprosthetic devices, controlling primary sensory neurons through selective neurostimulation.

Neural Networks for Modeling Neural Spiking in S1 Cortex

Somatosensation is composed of two distinct modalities: touch, arising from sensors in the skin, and proprioception, resulting primarily from sensors in the muscles, combined with these same cutaneous sensors. In contrast to the wealth of information about touch, we know quite less about the nature of the signals giving rise to proprioception at the cortical level. Likewise, while there is considerable interest in developing encoding models of touch-related neurons for application to brain machine interfaces, much less emphasis has been placed on an analogous proprioceptive interface. Here we investigate the use of Artificial Neural Networks (ANNs) to model the relationship between the firing rates of single neurons in area 2, a largely proprioceptive region of somatosensory cortex (S1) and several types of kinematic variables related to arm movement. To gain a better understanding of how these kinematic variables interact to create the proprioceptive responses recorded in our datasets, we train ANNs under different conditions, each involving a different set of input and output variables. We explore the kinematic variables that provide the best network performance, and find that the addition of information about joint angles and/or muscle lengths significantly improves the prediction of neural firing rates. Our results thus provide new insight regarding the complex representations of the limb motion in S1: that the firing rates of neurons in area 2 may be more closely related to the activity of peripheral sensors than it is to extrinsic hand position. In addition, we conduct numerical experiments to determine the sensitivity of ANN models to various choices of training design and hyper-parameters. Our results provide a baseline and new tools for future research that utilizes machine learning to better describe and understand the activity of neurons in S1.

A probabilistic recurrent neural network for decoding hind limb kinematics from multi-segment recordings of the dorsal horn neurons

OBJECTIVE

Providing accurate and robust estimates of limb kinematics from recorded neural activities is prominent in closed-loop control of functional electrical stimulation (FES). A major issue in providing accurate decoding the limb kinematics is the decoding model. The primary goal of this study is to develop a decoding approach to model the dynamic interactions of neural systems for accurate decoding. Another critical issue is to find reliable recording sites. Up to now, neural recordings from spinal neural activities were investigated. In this paper, the neural recordings from different vertebrae in decoding limb kinematics are investigated.

APPROACH

In the current study, a new generative probabilistic model with explicit considering the joint density is developed. Then, an adaptive discriminative learning algorithm is proposed for learning the model. It will be shown that the proposed generative process can be implemented by a recurrent neural network (RNN) with specific structure. We record the neural activities from dorsal horn neurons by using three electrodes placed in the L4, L5, and L6 vertebrae in anesthetized cats.

MAIN RESULTS

Information theoretic analysis on single-joint movement and multi-segment recordings implies the rostrocaudal distribution of kinematic information. It is demonstrated that during hip movement, best decoding performance is achieved by L4 recordings. For knee and ankle movements, best decodings are achieved by L5, and L6 recordings respectively. It is also shown that the decoding accuracy using multi-segment recordings outperforms decoding accuracy obtained by single-segment recording in multi-joint movement. The results also confirm the superiority of the proposed probabilistic recurrent neural network (PRNN) over the conventional recurrent neural network and Kalman filter ([Formula: see text]).

SIGNIFICANCE

Multi-segment recordings from dorsal horn neurons as well as the proposed probabilistic recurrent network model provide a promising approach for robust and accurate decoding limb kinematics.

Linear feature projection-based real-time decoding of limb state from dorsal root ganglion recordings

Proprioceptive afferent activities recorded by a multichannel microelectrode have been used to decode limb movements to provide sensory feedback signals for closed-loop control in a functional electrical stimulation (FES) system. However, analyzing the high dimensionality of neural activity is one of the major challenges in real-time applications. This paper proposes a linear feature projection method for the real-time decoding of ankle and knee joint angles. Single-unit activity was extracted as a feature vector from proprioceptive afferent signals that were recorded from the L7 dorsal root ganglion during passive movements of ankle and knee joints. The dimensionality of this feature vector was then reduced using a linear feature projection composed of projection pursuit and negentropy maximization (PP/NEM). Finally, a time-delayed Kalman filter was used to estimate the ankle and knee joint angles. The PP/NEM approach had a better decoding performance than did other feature projection methods, and all processes were completed within the real-time constraints. These results suggested that the proposed method could be a useful decoding method to provide real-time feedback signals in closed-loop FES systems.

Linear Feature Projection-Based Sensory Event Detection from the Multiunit Activity of Dorsal Root Ganglion Recordings

Afferent signals recorded from the dorsal root ganglion can be used to extract sensory information to provide feedback signals in a functional electrical stimulation (FES) system. The goal of this study was to propose an efficient feature projection method for detecting sensory events from multiunit activity-based feature vectors of tactile afferent activity. Tactile afferent signals were recorded from the L4 dorsal root ganglion using a multichannel microelectrode for three types of sensory events generated by mechanical stimulation on the rat hind paw. The multiunit spikes (MUSs) were extracted as multiunit activity-based feature vectors and projected using a linear feature projection method which consisted of projection pursuit and negentropy maximization (PP/NEM). Finally, a multilayer perceptron classifier was used to detect sensory events. The proposed method showed a detection accuracy superior to those of other linear and nonlinear feature projection methods and all processes were completed within real-time constraints. Results suggest that the proposed method could be useful to detect sensory events in real time. We have demonstrated the methodology for an efficient feature projection method to detect real-time sensory events from the multiunit activity of dorsal root ganglion recordings. The proposed method could be applied to provide real-time sensory feedback signals in closed-loop FES systems.

Decoding hind limb kinematics from neuronal activity of the dorsal horn neurons using multiple level learning algorithm

Decoding continuous hind limb joint angles from sensory recordings of neural system provides a feedback for closed-loop control of hind limb movement using functional electrical stimulation. So far, many attempts have been done to extract sensory information from dorsal root ganglia and sensory nerves. In this work, we examine decoding joint angles trajectories from the single-electrode extracellular recording of dorsal horn gray matter of the spinal cord during passive limb movement in anesthetized cats. In this study, a processing framework based on ensemble learning approach is propose to combine firing rate (FR) and interspike interval (ISI) information of the neuronal activity. For this purpose, a stacked generalization approach based on recurrent neural network is proposed to enhance decoding accuracy of the movement kinematics. The results show that the high precision neural decoding of limb movement can be achieved even with a single electrode implanted in the spinal cord gray matter.

Separability of neural responses to standardised mechanical stimulation of limbs

Considerable scientific and technological efforts are currently being made towards the development of neural prostheses. Understanding how the peripheral nervous system responds to electro-mechanical stimulation of the limb, will help to inform the design of prostheses that can restore function or accelerate recovery from injury to the sensory motor system. However, due to differences in experimental protocols, it is difficult, if not impossible, to make meaningful comparisons between different peripheral nerve interfaces. Therefore, we developed a low-cost electronic system to standardise the mechanical stimulation of a rat’s hindpaw. Three types of mechanical stimulations, namely, proprioception, touch and nociception were delivered to the limb and the electroneurogram signals were recorded simultaneously from the sciatic nerve with a 16-contact cuff electrode. For the first time, results indicate separability of neural responses according to stimulus type as well as intensity. Statistical analysis reveal that cuff contacts placed circumferentially, rather than longitudinally, are more likely to lead to higher classification rates. This flexible setup may be readily adapted for systematic comparison of various electrodes and mechanical stimuli in rodents. Hence, we have made its electro-mechanical design and computer programme available online