ORIGINAL RESEARCH published: 15 January 2020 doi: 10.3389/fnsys.2019.00082 Novel Activity Detection Algorithm to Characterize Spontaneous Stepping During Multimodal Spinal Neuromodulation After Mid-Thoracic Spinal Cord Injury in Rats Raymond Chia 1,2 , Hui Zhong 3 , Bryce Vissel 1,2 , V. Reggie Edgerton 1,3,4,5,6,7* and Parag Gad 1,4* 1 Faculty of Science, Centre for Neuroscience and Regenerative Medicine, University of Technology Sydney, Sydney, NSW, Australia, 2 St Vincent’s Centre for Applied Medical Research, Sydney, NSW, Australia, 3 Department of Integrative Biology and Physiology, University of California, Los Angeles, Los Angeles, CA, United States, 4 Department of Neurobiology, University of California, Los Angeles, Los Angeles, CA, United States, 5 Department of Neurosurgery, University of California, Los Angeles, Los Angeles, CA, United States, 6 Brain Research Institute, University of California, Los Angeles, Los Angeles, Edited by: CA, United States, 7 Institut Guttmann, Hospital de Neurorehabilitació, Institut Universitari Adscrit a la Universitat Autònoma Michelle Maria Rank, de Barcelona, Barcelona, Spain The University of Melbourne, Australia Reviewed by: A mid-thoracic spinal cord injury (SCI) severely impairs activation of the lower limb Marin Manuel, Université Paris Descartes, France sensorimotor spinal networks, leading to paralysis. Various neuromodulatory techniques László Demkó, including electrical and pharmacological activation of the spinal networks have been Balgrist University Hospital, Switzerland successful in restoring locomotor function after SCI. We hypothesized that the Andrew C. Smith, combination of self-training in a natural environment with epidural stimulation (ES), Regis University, United States quipazine (Quip), and strychnine (Strych) would result in greater activity in a cage *Correspondence: environment after paralysis compared to either intervention alone. To assess this, we V. Reggie Edgerton vre@ucla.edu developed a method measuring and characterizing the chronic EMG recordings from Parag Gad tibialis anterior (TA) and soleus (Sol) muscles while rats were freely moving in their home paraggad@gmail.com cages. We then assessed the relationship between the change in recorded activity over Received: 13 September 2019 time and motor-evoked potentials (MEPs) in animals receiving treatments. We found Accepted: 16 December 2019 that the combination of ES, Quip, and Strych (sqES) generated the greatest level of Published: 15 January 2020 recovery followed by ES + Quip (qES) while ES + Strych (sES) and ES alone showed Citation: Chia R, Zhong H, Vissel B, least improvement in recorded activity. Further, we observed an exponential relationship Edgerton VR and Gad P (2020) Novel between late response (LR) component of the MEPs and spontaneously generated Activity Detection Algorithm step-like activity. Our data demonstrate the feasibility and potential importance of to Characterize Spontaneous Stepping During Multimodal Spinal quantitatively monitoring mechanistic factors linked to activity-dependence in response Neuromodulation After Mid-Thoracic to combinatorial interventions compared to individual therapies after SCI. Spinal Cord Injury in Rats. Front. Syst. Neurosci. 13:82. Keywords: sub-threshold spinal cord stimulation, spinal cord injury, spontaneous cage activity, EMG, muscle doi: 10.3389/fnsys.2019.00082 coordination, strychnine, quipazine, evoked potentials Frontiers in Systems Neuroscience | www.frontiersin.org 1 January 2020 | Volume 13 | Article 82 Chia et al. Step Detection and Neuromodulation After Paralysis INTRODUCTION and water provided ad libitum. All survival surgical procedures were conducted under aseptic conditions and with the rats Numerous studies of different animal models of spinal deeply anesthetized with isoflurane gas (1.5–2%) administered cord injury (SCI) have demonstrated that modulating the via facemask. All procedures described below are in accordance physiological states of spinal networks, pharmacologically and with the National Institute of Health Guide for the Care and Use electrically and in combination with motor training enables of Laboratory Animals and approved by the Animal Research improved motor performance (Edgerton et al., 2004; Ichiyama Committee at UCLA. et al., 2005; Gerasimenko et al., 2008; Lavrov et al., 2008; Courtine et al., 2009; Harkema et al., 2011; Rossignol and Frigon, 2011; Head Connector and Intramuscular EMG Gad et al., 2013a,b; Angeli et al., 2014; Gerasimenko Y. et al., Electrode Implantation 2015; Capogrosso et al., 2016; Formento et al., 2018; Gill et al., A small incision was made at the midline of the skull. The 2018). Specific pharmacology methods including the use of muscles and fascia were retracted laterally, small grooves were quipazine (Quip) (Fong et al., 2005; Ichiyama et al., 2008b; Gad made in the skull with a scalpel, and the skull was dried et al., 2015; Taccola et al., 2018) and strychnine (Strych) (de thoroughly. Two amphenol head connectors with Teflon-coated Leon et al., 1999; Gad et al., 2013b, 2015) have been effective stainless steel wires (AS 632, Cooner Wire, Chatsworth, CA, in activation in the locomotor spinal neural networks to enable United States) were securely attached to the skull with screws and locomotor activity in rodent with severe paralysis. Similar dental cement as described previously (Ichiyama et al., 2008a). interventional approaches have been observed in humans with a The tibialis anterior (TA, ankle flexor) and soleus (Sol) (ankle severe spinal injury (Gerasimenko Y. et al., 2015; Gerasimenko extensor) muscles were implanted bilaterally with intramuscular Y. P. et al., 2015). Recently, we explored the effects of long-term EMG recording electrodes (Roy et al., 1992). Skin and fascial sub-motor threshold ES in rats with complete mid-thoracic incisions were made to expose the belly of each muscle. Two spinal transections (Gad et al., 2013a). These data demonstrated wires extending from the skull-mounted connector were routed that when rats received sub-motor threshold ES moved around subcutaneously to each muscle. The wires were inserted into their home cages five times more than when they were not the muscle belly with a 23-gauge needle, and a small notch receiving ES, thus demonstrating the possibility to use ES to (0.5–1.0 mm) was removed from the insulation of each wire enable “self-training” in an environment while patients could be to expose the conductor and form the recording electrodes. performing routine activities of daily living. The wires were secured within the belly of each muscle via a The objective of this study was to identify spinal rat’s hind- suture on the wire at its entrance into and exit from the muscle limb stepping like activity based on EMG patterns and to quantify belly. The proper placement of the electrodes was verified during the effects of multi-modal neuromodulation on spontaneous the surgery by stimulation through the head connector and locomotor activity in standard individual housing cages. We postmortem via dissection. hypothesized that multiple neuromodulatory modalities can transform non-functional spinal networks into more excitable physiological states to enable “self-training” in freely moving Spinal Cord Transection and Electrode spinal rats in their home environments. In addition, motor- Implantation Procedures and evoked potentials (MEPs) observed in hindlimb EMG signals Post-surgical Animal Care could present a functional biomarker of these novel physiological A partial laminectomy was performed to expose the T8–T9 spinal states (Gad et al., 2015). Spinal rats that had previously been cord, and then a complete spinal cord transection to include the trained to step on a treadmill in presence of Strych and Quip dura was performed with microscissors. Two surgeons verified were tested in their home cages under five different conditions, the completeness of the transection by lifting the cut ends of the i.e., no ES (Pre), ES, ES + Quip (qES), ES + Strych (sES), spinal cord with fine forceps and passing a glass probe through and ES + Quip + Strych (sqES). To test these hypotheses, the lesion site. Gel foam was inserted into the gap created by we developed a novel Thresholding Offline Kinematic and the transection as a coagulant and to separate the cut ends EMG Data Analysis method (TOKEDA) to characterize motor of the spinal cord. For epidural electrode implantation, partial behaviors to neuromodulatory interventions. TOKEDA has the laminectomies were performed to expose the spinal cord levels capability to be scaled up or down with respect to its sensitivity to L2 and S1. Two Teflon-coated stainless steel wires from the head detect multiple movement and electrophysiological events during connector were passed under the spinous processes and above spontaneous activity under different environments. the dura mater of the remaining vertebrae between the partial laminectomy sites. After a small portion (∼1-mm notch) of the Teflon coating was removed and the conductor was exposed on MATERIALS AND METHODS the surface facing the spinal cord, the electrodes were sutured to the dura mater at the midline of the spinal cord above and Animal Preparation and Care below the electrode sites with 8.0 Ethilon suture (Ethicon, New Data were obtained from four adult female Sprague Dawley Brunswick, NJ, United States). Two common ground (indifferent rats (270–300 g body weight). Pre and post-surgical animal EMG and ES) wires (∼1 cm of the Teflon removed distally) were care procedures have been described previously (Roy et al., inserted subcutaneously in the mid-back region on the right side 1992). The rats were housed individually in cages with food (EMG) and midline (ES) close to the tail. All wires (for both EMG Frontiers in Systems Neuroscience | www.frontiersin.org 2 January 2020 | Volume 13 | Article 82 Chia et al. Step Detection and Neuromodulation After Paralysis and ES) were coiled in the back region to provide stress relief. All 10 kHz (Figure 1F). Data were recorded continuously for 6 h incision areas were irrigated liberally with warm, sterile saline. All between 8 p.m. and 2 a.m. One out of the 20 experiments (n = 4 surgical sites were closed in layers with 5.0 Vicryl (Ethicon, New rats, five experiments/rat) were conducted every night and were Brunswick, NJ, United States) for all muscle and connective tissue randomized with at least 1 day gap between two successive layers and for the skin incisions in the hind-limbs and 5.0 Ethilon recordings for a rat and all 20 experiments were completed within for the back skin incision. Buprenex (0.01–0.05 mg/kg s every 3 weeks. Note that the first 20 min EMG recordings for rat #1 in 8–12 h) was used to provide analgesia. Analgesics were initiated the sqES and all recording channels for rats #2 and #4 for the ES before completion of the surgery and continued for a minimum only case were corrupted and unable to be used. of 2 days. The rats were allowed to recover fully from anesthesia in an incubator. The rats were housed individually in cages that Signal Processing had ample CareFresh bedding, and the bladders of the spinal rats The raw EMG signal was filtered with a third order Butterworth were expressed manually three times daily for the first 2 weeks bandpass filter with cut-off frequencies of 30–1000 Hz after surgery and two times daily thereafter. The hind-limbs of the (Figure 1A). This attenuates DC offset and higher frequency spinal rats were moved passively through a full range of motion noise. To make the signal more sensitive to change, the once per day to maintain joint mobility. All of these procedures bandpass filtered signal was fed into a Teiger Keiser Energy have been described in detail previously (Courtine et al., 2009). Operator (TKEO) algorithm (Solnik et al., 2010). The output was rectified and smoothened across a 50-Hz cut-off low pass Training Procedures filter (Figure 1B). All rats were trained for bipedal stepping and standing on a motor driven rodent body weight supporting treadmill for Thresholding 5 days/week, 20 min/day for 6 weeks starting at 12 days post- To determine the threshold of activity across each channel, the injury (dpi), including the days of testing (de Leon et al., 2002). 6 h recordings were broken up in to 10 s bins (Figure 1F). Bipolar ES between L2 and S1 (current flowing from L2 to A threshold sampling method was used to determine periods of S1) at frequency of 40 Hz, pulse width 0.2 ms was used in rest throughout each bin. To find periods of rest, the data were combination with Quip (0.3 mg/kg; de Leon et al., 1999) and averaged across 10 ms and the minimum was selected as the first Strych (0.5 mg/kg; Gad et al., 2013a) injected intraperitoneally point of interest. A sliding window procedure was implemented 10 min before each training/testing session. ES was only delivered to approximate the entire period of rest within the 10 s bin. during the 20 min/day training periods as described previously This was limited by a first differential magnitude of 0.1 mV/s or (Ichiyama et al., 2005). Chronic step training was used to logarithmic threshold of 300 ln (mV). It was found that during train and reinforce locomotor neural networks that generate periods of slower, more gradual ramping EMG activity, the first spontaneous cage activity (Gad et al., 2013a). At the early and second differential values were not sufficiently sensitive. To stages of training, ES intensity was set at threshold or at super- account for slow ramping activity, the logarithmic threshold was threshold to invoke locomotive activity. As training continued, implemented. This enables the software to detect both quick and the stimulation intensity was gradually reduced, dependent upon slower adapting changes across the EMG signal. the stepping performance of each rat until the stimulation A double thresholding mechanism was used to detect bursts intensity of below threshold. of EMG. The mean and standard deviation of the detected rest period was calculated and used to determine a threshold for EMG Stimulation and Testing Procedures magnitude. Given the circumstance of recording and filtering, The threshold for eliciting muscle twitch and corresponding a scaling constant (J) of 7 was used. Finally, to complete the time linked EMG response in the Sol was identified. The sub- double thresholding, a time criterion was implemented for each threshold level was set to 20% below the motor threshold during channel, where the TA channel time threshold was 0.1 s and Sol the recording of spontaneous cage activity. ES was delivered only channel time threshold was 0.3 s. If the signal magnitude was during the training and testing periods. Each rat prior to testing below the time threshold, it was not considered bursting activity. was injected intraperitoneally 10 min prior to beginning the The minimum criteria for an EMG magnitude to remain above testing, with the prescribed neuromodulating pharmacology. The the motor threshold as defined by: spontaneous activity of the spinal rats was determined in their home cage. Spinal rats that had previously been trained to step threshold = µrest + J ∗ σrest on a treadmill in presence of Strych and Quip were tested under five different conditions, i.e., no ES (Pre), ES, ES + Quip (qES), For an example of a period of an EMG signal which met ES + Strych (sES), and ES + Quip + Strych (sqES). The head the magnitude threshold but not the time criteria, see the RSol connector was coupled to a set of amplifiers and a stimulator. channel of Figure 1A. The swivel arrangement was attached to allow the rats to roam freely in the cage. There was food distributed throughout the Activity Detection cage floor to encourage movement and exploration. IR video To detect instances of step-like behavior, a dataset was generated data were recorded using a camcorder for select conditions. using IR video data. A total of 26.7 h of video data was EMG data were amplified and recorded using custom LabView- collected. Step-like, standing, and rest events were compiled into based data acquisition software with a sampling frequency of a dataset of activity. These data were then synchronized with Frontiers in Systems Neuroscience | www.frontiersin.org 3 January 2020 | Volume 13 | Article 82 Chia et al. Step Detection and Neuromodulation After Paralysis FIGURE 1 | (A) Representative band-pass filtering of TA and Sol. The boxed area shows a close-up example of the thresholding and the normalized difference plot used to characterize step-like activity. Regions of the signal below threshold (blue), above threshold (red), and rest (black) were recorded. (B) TKEO signal conditioned, rectified, and smoothed EMG signal from the TA and Sol muscle of a spinal rat during spontaneous cage activity. A representative example of a singular hind-limb step within the first hour of recording spontaneous activity. (C) The relative difference between the normalized values of the TA and Sol from graphs (B). The gray line y-axis = 0. The blue line is a trace of the relative difference plot and in red is the periods of active extension, identified through thresholding. Green triangles represent points of local maxima during active flexion. These features were used to detect spontaneous hind-limb step-like activity. (D) A singular identified step using the TOKEDA method as summarized in graphs A–C. This plot illustrates motor evoked potentials (MEPs) over time for the tibialis anterior (TA) and Soleus (Sol), respectively. (E) A representative MEP from the TA recording shown in panel D. Color scaling reflects amplitude in mV. Note the windows marking the early, middle, and late responses and peaks (∗ ) and troughs (o). (F) A breakdown of the volume of data involved within the experiments and the process of data structure and analysis. Out of four possible channels, displayed in the top-most subplot is the full 6 h EMG recording of RSol from a singular treatment. Highlighted within the dashed blue box is a 10-min section in which the total hour was sectioned to give a total of 36 files per 6 h recording. The dashed red line bins the 10-min files for thresholding, burst detection, and step detection. The dashed magenta box represents a singular instance of step-like activity from the beginning of swing to the end of the stance phase. Within the dashed green box is a singular MEP measured within a singular instance of the step-like event. Frontiers in Systems Neuroscience | www.frontiersin.org 4 January 2020 | Volume 13 | Article 82 Chia et al. Step Detection and Neuromodulation After Paralysis EMG recordings of the TA and Sol of both left and right hind- 40 Hz stimulation pulse was used to separate the EMG signal limbs to produce 300 EMG samples from a combined total of into 25 ms MEPs. The early response (ER), middle response 217 step-like events and 83 non-functional activity events. While (MR), and LR time windows were defined as 1–4, 4–7, and there were several standing events detected throughout the 6 h 7–25 ms, respectively (Figure 1E). To successfully detect the period, analyses were limited to step-like events only. peaks and troughs of the MEP signal, a threshold using the second A relative difference signal between a normalized TA and Sol differential was used. Each MEP was interpolated to double the recording was calculated to determine instances of spontaneous number of points. Using the interpolated signal, the local maxima step-like activity (Figure 1C). The TA and Sol signals were and local minima with a minimum separation of 0.6 ms and normalized to account for missing force-EMG measurements. minimum prominence of 0.2 were found. These local maxima This signal was used as a representation of coordination between and minima were filtered through a second differential threshold extension and flexion of the ankle. It was assumed that if the of 1.5 × 106 such that only peaks and troughs that met the signal was positive, the ankle was applying a greater flexion threshold criteria were accepted (Figure 1C). This was used to force on the joint compared to extension, vice versa for negative calculate the area under the curve (IEMG), peak-to-peak, and magnitudes. A peak during the positive region of a signal total number of peaks. followed by a sharp dip into the negative portion of the signal was associated with the reduction in flexion activity during swing Statistical Analyses phase and an activation of extensors during the stance phase. All data reported as means ± standard error (SE). Statistically Peaks from the relative difference signal were only considered if significant differences were determined using a one-way they were positive in magnitude and during an active TA EMG ANOVA with the Tukey–Kramer post hoc test, correlation burst. By using the active burst regions as a “filter,” many false coefficients were calculated using Pearson’s linear coefficient. positives of locomotion initiation were removed. All statistical difference was set at P < 0.05. All results For conservative activity detection, a step was only registered reporting step-like activity were normalized to no intervention when a positive peak was followed by a negative signal during (Pre) measurements. an active extension EMG burst achieved within a threshold of 0.5 s. Only peaks with a magnitude greater than the difference threshold of 0.01, in the relative difference signal and only RESULTS magnitude differences greater than the difference threshold were considered. The time and magnitude thresholds were Task Recognition Validity characterized from numerous test examples in the given dataset. To test the validity of the task recognition algorithm derived In order to filter coordination and reflect false positives, a time from the relative difference signal, the point of change from threshold for active Sol burst was characterized at 0.25 s. Finally, flexion phase to extension phase during a “step-like” event was a minimum period between each local maxima of 0.2 s was set. logged (difference signal equals zero). A dataset of IR video This was to ensure that only the last local maxima within the TA and EMG-derived step-like examples were collected to form burst was selected. the human-observation-based dataset. In this dataset, the onset Accuracy was defined as the ratio between correctly detected and completion of a step-like or rest event for each limb events over the total number of sampled events. Precision was was logged to a millisecond precision. The human-observation defined as the number of true positives out of both true and dataset comprised of 300 EMG samples from a combined total false positives; thus, an estimate of the number of actual step- of 217 step-like events and 83 non-functional activity events. like events detected. Recall was defined as the number of true A validation script checked the algorithm event log against the positives out of false negatives and true positives; thus, an human-observation dataset. To gather a sense of relevance in estimate for detecting all relevant step-like events within the detected step-like patterns, accuracy, precision, and recall were existing dataset. calculated from the validation script. The relative difference For a complete algorithm process illustration, refer signal detected step-like activity performed with an accuracy of to Figure 2. 83%, precision of 88%, and recall of 89%. MEP Analysis For granular electrophysiological analysis, the entire step cycle Electropharmacological Treatments was defined from the beginning of the swing phase to the end Modulate the Functional State of Spinal of the stance phase. This was set at the point where the relative Circuitry difference signal first reaches zero before the detected peak and All interventional conditions increased the number of where the negative signal reaches zero or the extension EMG spontaneous step-like activity compared to baseline (Pre- burst has ended (Figure 1C). With these defined windows of step- intervention) (Figure 3). Quip or Strych significantly increased like activity, MEPs were analyzed to extract electrophysiological the number of recorded step-like activity in all animals tested. features such as the number of positive peaks and integrated A significant level of variability was observed among rats, EMG for both the middle and late response (LR) (Figures 1D,E). especially in the qES and sqES experiments. However, sqES The MEPs were extracted from the bandpass-filtered EMG data. treatment generated the highest number of steps recorded when Using a recording of the stimulation pulses, the end of each compared to Pre and ES (P < 0.05). Frontiers in Systems Neuroscience | www.frontiersin.org 5 January 2020 | Volume 13 | Article 82 Chia et al. Step Detection and Neuromodulation After Paralysis FIGURE 2 | Flowchart of the signal processing and algorithm used to detect beginning and end of swing and stance phases. For access to the code, see the following GitLab repository, https://code.research.uts.edu.au/13035516/tokeda.git. Since a single bolus dose of a pharmacological treatment was variance was present in the sqES stance period compared to qES administered at the start of the experiment, we observed a time- within the first 2 h. In addition, during the entire 6-h period, dependent phenomenon in the number of steps detected with the overall left–right symmetry was maintained suggesting a bipedal greatest level of activity observed in the first 120 min (Figure 4). response (Figure 4). These observations were specific to the rapid registration of step- Inclusion of multiple pharmacological agents in the presence like activity for sqES and qES treatments (Figure 5B). This time- of ES resulted in a more consistent and greater IEMG response dependent phenomenon was not observed in Pre or ES cases. (Figure 5A). While sES reduced the variability across both The stepping activity occurred stochastically throughout the 6 h extensor and flexor muscles in the hindlimbs, the inclusion (Figures 4, 5B). Spontaneous step-like activity during the first of Quip significantly increased IEMG magnitude compared 120 min occurred with longer, more consistent step lengths and to Pre (P < 0.05) (Figures 5A, 6A). For the ES and Pre increased muscle activity during sqES and qES (Figures 4, 5A,B). treatments, the IEMG of TA differed in shape compared to Sol During continuous stepping activity, an average stance period of with wider error bands while the sES, qES, and sqES relatively ∼2 s was observed (Figure 4). Both the step activity and stance coordinated activity in the TA and Sol muscles with narrower period across the population had lower variability during sqES error bands were observed (Figure 5A). Differences between compared to other conditions tested (Figures 4, 5B). A smaller TA and Sol may be explained by the higher occurrence of Frontiers in Systems Neuroscience | www.frontiersin.org 6 January 2020 | Volume 13 | Article 82 Chia et al. Step Detection and Neuromodulation After Paralysis IEMG and step-like activity, while sES, ES, and Pre reflect a linear correlation. sqES intervention had the largest SE across both the x- and y-axes of Figure 6C whereas sES and Pre have minimally discernible spread among the sampled population. Increased baseline tone in Sol during rest correlates (R = 0.6943, P < 0.05) with the increased step-like activity across all conditions and rats and consistent with the overall summed IEMG across the 6 h (Figure 6B). There appears to be a breakaway condition given the logarithmic response in the basal tone in Sol represented (Figure 6B). Evoked Responses Correlate With Spontaneous Step-Like Activity Motor-evoked potentials were analyzed across the Sol during spontaneous step-like activity as these signals more accurately reflect the state of the neural networks associated with the trained task. The detected peaks in the middle and LRs (MRs and LRs) of the Sol during spontaneous step-like activity in their home cages were exponentially correlated (Figure 7). Due to the sub- threshold nature of spinal stimulation, the evoked responses appeared to be non-time linked to the stimulation pulses when compared to suprathreshold stimulation on a treadmill (Gad FIGURE 3 | Mean% change in step-like events for each rat over the 6-h et al., 2013b). Each of the treatments had linear regression applied period for each treatment (∗ ). The light gray line is set at 100%, normalized to pre-treatment for each rat. Results for sqES were significantly different to both the LRs and MRs. Both ES and sES display a negative compared to Pre and ES (P < 0.05). gradient associative with decreasing change in step-like activity when compared with Pre while qES and sqES have a steep positive gradient. The combination of all data points was fitted to an exponential equation where both the LRs and MRs have a similar relationship with step-like activity, the MR trend has a lower first-order coefficient when compared to the LR. As anticipated, the sqES and qES evoked the highest number of LRs and MRs which corresponded to a greater number of detected spontaneous step-like activity. DISCUSSION We developed a method for identifying spinal rat’s hind- limb stepping like activity based on EMG patterns recorded from the Sol and TA to determine the effects of multi-modal neuromodulation on spontaneous locomotor activity in standard individual housing cages. These results describe the functional and electrophysiological changes in response to activity-specific FIGURE 4 | An example of the step-like activity across interventions plotted training in combination with enabling ES and pharmacological as number of steps per minute and mean stance periods for the left (black) interventions. For access to longitudinal functional analyses, refer and right (red) hind limb for rat #3. For other rat performances, refer to Gad et al. (2015). This experiment was conducted to determine Supplementary Figures 1–3. the feasibility of developing a quantitative electrophysiological assessment tool of spontaneous cage activity of rats as an estimate of neuromuscular activity. co-contractions and organized coordinated activity during the Several strategies have been developed for detecting rat activity pharmacological interventions. The steep gradient was only within a caged setting using a variety of sensing methods. These observed during the first 2 h of sqES and qES and becomes more include the use of vibration/tilt sensing (Parreno et al., 1985; linear during hours 2–6, possibly indicating the effective half- Megens et al., 1987; Ganea et al., 2007), IR beams (Clarke et al., life of the drug. Significant correlation was observed between 1985), IR and non-IR video tracking (Tamborini et al., 1989; the overall muscle activity (IEMG) and the steps registered Morrel-Samuels and Krauss, 1990; Aragão et al., 2011; Gad et al., across both right hind-limb muscles (Figure 6C, P < 0.05). 2013a; York et al., 2013), capacitive flooring (Pernold et al., 2019), Both sqES and qES have an exponential relationship between optical touch sensors (Mendes et al., 2015), RFID (Redfern et al., Frontiers in Systems Neuroscience | www.frontiersin.org 7 January 2020 | Volume 13 | Article 82 Chia et al. Step Detection and Neuromodulation After Paralysis FIGURE 5 | All cumulative sum plots are normalized per animal from 0 to 1. (A) Cumulative sum of IEMG in of the left and right TA and Sol. (B) On the left is a cumulative sum plot of the step-like events occurring over time for each treatment (blue, red, gold, and magenta represents rat #1, #2, #3, and #4, respectively). On the right is the averaged cumulative sum plot. Shaded is the standard error. 2017), and radar technology (Martin and Unwin, 1980; Rose et al., notion has been reflected in past research papers (Lavrov et al., 1985; Young et al., 1996; Genewsky et al., 2017). Most pattern 2006, 2008; Gad et al., 2015; Alam et al., 2017). Unfortunately, recognition approaches based on EMG recordings have been without the use of kinematic recordings, information such as used to control prosthetics (Huang et al., 2009, 2010). A more locomotion speed, step quality, step length, and other kinematic- complex approach involves the use of Hidden Markov Models related metrics remain remarkedly rare given the availability of (HMMs) to represent stochastic processes using time series data the appropriate technologies. to predict the probability of model states (Lee, 2008). However, Long-term recordings of electrophysiological data in the detection of spontaneous self-recovering movement after laboratory animals and humans before and after treatment paralysis outside of a controlled clinical environment is rarely for a dysfunction have been performed previously (Alaimo explored quantitatively due to the limitation associated with et al., 1984). The existing hypothesis regarding the possibility securely attaching sensors or markers at specific locations, etc. of electrophysiological biomarkers is the emergence of MRs Use of antennas attached to a rat have been used to track and LRs in MEPs during activity-driven training (Lavrov et al., gross body movements through a maze of tubes. It was shown 2006, 2008; Gad et al., 2015). There appear to be underlying that self-motivated training within a “enriched environment” neurological mechanisms involved in time-related modulation lead to superior performance in skilled movement compared to in MEPs during treadmill activities. In Gad et al. (2015), it restricted task-specific training. This “RatTrack” system allows was shown that re-emergence in LRs and reduced MRs with for testing self-initiated and task-specific training and dose– changes in modulation of flexor extensor motor pools reflect responses (Starkey et al., 2014). While the mentioned efforts plasticity of the neural networks. Another study reports a can be extended in the direction of automation, each of these correlation between the MR and LR with changes in the methods lack the ability to have a direct measurement of the EMG activity level of the corresponding muscle, perhaps, neuromotor parameters. TOKEDA relies solely on chronically reflecting in part monosynaptic and polysynaptic pathways, implanted EMG electrodes for sensing to assess in vivo responses respectively, to motor pools (Gerasimenko et al., 2006). In to different combinations of electrical and pharmacological the present paper, we explore the characteristics of the same neuromodulatory interventions. Having access to neuromuscular signals during spontaneous in vivo step-like activity within an activity synchronized with stimulation pulses directly linked enriched caged environment. Paralleled between these findings, to behaviors provides a direct and realistic measurements of Figure 7 presents an exponentially increasing trend between reorganizing neural networks throughout a chronic period as the the emergence of LRs and the functional response to hind-limb nervous system becomes more functional (Edgerton et al., 2008). step training. Notably, while Gad et al. (2015) concludes a As the MEPs and associated behavior are directly linked from an reduced reliance on MRs may indicate a change in targeted “input–output” relationship, we hypothesize that a biomarker for neural networks and neuro-plastic mechanisms, Figure 7 reorganization may be discernible from the presented data. This illustrates no distinct difference between LRs and MRs beside a Frontiers in Systems Neuroscience | www.frontiersin.org 8 January 2020 | Volume 13 | Article 82 Chia et al. Step Detection and Neuromodulation After Paralysis Additionally, note that the highlighted outliers in Figure 7 were not included in the regression calculations. These data points suggest that at below average low MEP peak counts, high levels of activity can occur. These outliers were measured from the same rat within the same experiment. During analysis of the video data, an unusual frequency of induced, air-stepping was observed (Supplementary Videos 1, 2). The fact that some number of these counted step-like events were not enabled but rather induced by ES may result in minimal polysynaptic pathways being activated during air-stepping, i.e., when the proprioceptive and tactile afferents typically associated in over- ground, weight-bearing hind-limb locomotion were not present. A number of possible reasons could provide explanation as to why repetitive, non-weight bearing cycles could readily occur. The exponential trend visible in Figure 7 alongside the data in Figures 6B,C present the possibility of a “breakaway” point where the minimum level of neuronal activity and polysynaptic mechanisms is required for significant functional improvement. It was only when Quip was introduced where a significant improvement in spontaneous step-like activity was observed. Given the extensive research provided with training under the influence of Strych, one would expect that Strych experiments would provide equal benefit when compared with Quip (de Leon et al., 1999; Gad et al., 2013b, 2015). Albeit a small sample size, these data extend the hypothesis FIGURE 6 | (A) The average sum of IEMG across all the animals in each treatment in the TA and Sol channels for both sides and the 95% confidence of the existence of electrophysiological biomarkers and interval. The Sol channel for sqES and qES was significantly different from Pre warrant the extension of further investigation into longer (∗ P < 0.05). There is a trend of increasing mean and standard error in IEMG in term longitudinal studies and the involvement of more robust both TA and Sol as more interventions are introduced on top of ES. (B) The algorithms and in vivo sensing technologies. These observations graph depicts the relationship of the calculated EMG amplitude threshold required to determine burst activity of the neural networks in the Sol and the alongside the previously discussed representation of plastic change in step-like activity normalized to pre intervention for all rat subjects in mechanisms in MEPs agree with our initial hypothesis of both hindlimbs. The treatments, sqES, qES, sES, and ES are plotted in blue, multiple neuromodulatory modalities transforming non- red, black, and magenta, respectively. A linear regression of the natural functional spinal networks to a more excitable state to logarithm of Sol threshold values against the normalized change in step-like enable “self-training.” activity results in R = 0.6943. Inclusion of pharmacology and ES increases the ln(Sol) threshold, only qES and sqES have drastic change in detected Presented here is a novel method of simultaneously identifying step-like events. (C) Ln(Total IEMG/10 min) plotted against step-like step-like activity while recording electrophysiological biomarkers events/10 min across 6 h for sqES, qES, sES, ES, and Pre in the right of stepping. The value of these types of data can be further hind-limb. In the corresponding shaded area is the standard error for the x- improved with the introduction of additional sensors in order and y-axis represented in an ellipse. A Pearson’s linear correlation coefficient to compose a more comprehensive assessment of movements reveals significant positive correlation in all interventions for both hind-limbs (P < 0.05 in all cases). While the sES, ES, and Pre interventions are linearly as occurs in vivo in a variety of environments. As gait patterns correlated, the qES and sqES have an exponential relationship between the have a state-specific nature, repeating the task recognition natural log of the total IEMG/10 min and the step-like events/10 min. Note exercise using a stochastic probability based method such that sqES, qES, and ES have a distinct y-axis error bound while the x-axis as HMMs (Byung-Woo et al., 1997; Sy Bor et al., 2006; error bound is less notable. This is especially notable in the RSol graph. Naeem and Bigham, 2007; Lee, 2008) or long short-term memory (LSTM) (How et al., 2014; Tsironi et al., 2017) with the existing datasets may improve the overall accuracy of the task recognition. The present data provide an example slightly lower exponential coefficient. These contrasting reports of a “proof of concept” approach to examine the level of may be due to the small sample size or due to difference in activity dependence that is present when testing the efficacy experimental environment. While Gad et al. (2015) performed of an intervention. It is technically becoming more feasible experiments within a weight-supported treadmill setting, the to chronically record EMG from many muscles to detect presented experiment was conducted in a weight bearing, free how the reorganization of neuronal networks that control roaming environment. This difference in locomotive setting may locomotion can be focused on the patterns of coordination introduce disruptions to the reorganization of neural networks of flexion/extension, abduction/adduction, and non-repetitive due to the loss of highly organized and predictable patterns of tasks such as grip and pinch maneuvers as well as repetitive sensory input that is normally generated with step training on a tasks such as cycling, etc. Moreover, the ability to measure treadmill (Edgerton and Roy, 2009; Taccola et al., 2018). the state of the locomotive neural circuitry to determine the Frontiers in Systems Neuroscience | www.frontiersin.org 9 January 2020 | Volume 13 | Article 82 Chia et al. Step Detection and Neuromodulation After Paralysis FIGURE 7 | (A) Change in the number of late and (B) middle peaks responses (normalized to pre values) for each rat, treatment, and left and right hind-limb of the Sol MEPs. Circled in red are two outliers identified as rat #3 during the sqES treatment. Additionally, linear regressions for each of the respective treatments in their corresponding color are plotted. An exponential curve (dashed line) encompasses all treatments. direct relationship between the treatment provided and the While the presented data are based on data derived underlying neuronal mechanics serves as detailed insight as the from spinally injured rats that have had multiple training subject undergoes training, providing the opportunity to adjust sessions under the influence of ES and pharmacological agents, treatments in order to maintain an enabling effect, maximizing longitudinal functional data and analysis can be found in the activity-dependent recovery. paper presented previously (Gad et al., 2015). Longitudinal In addition, our algorithm was used to examine the effect of functional analyses were performed at 12, 22, 35, and 49 days pharmacology in combination with ES by directly comparing past injury with hindlimb joint kinematics tracked with EMG the occurrence of step-like activity within a chronic self- recordings of the TA and Sol under ES, qES, sES, and sqES training environment. The ability to record chronically for conditions during standing and step training in a weight 6 h in a natural environment highlights the functional and supported treadmill. Future work should include continual electrophysiological changes over time and could represent a measurements of functional neurological changes as reported functional biomarker of recovery post paralysis. This leads here, but throughout a training period. The present data are to a potentially critical question in the rehabilitation strategy largely focused on the acute effects of spontaneous activity after post paralysis. Can daily periods of sub-threshold electrical completion of electrical and pharmacological neuromodulation neuromodulation of the spinal circuitry enhance the performance which facilitated self-training (Figure 4). A repetition of this of motor skills as well as or even more than when a task- experiment with a larger sample size and with a regimented specific rehabilitation paradigm is imposed? If that is the level of a specific motor task would be of significant interest. case, then the results will have an immediate impact on how Adaptation of the existing algorithm with machine-learning neuromodulation can be used to enhance sensorimotor recovery algorithms such as HMMs and LSTM could further enhance in individuals with other neural disorders than SCI. Spinal our understanding of the more critical variables intrinsic to the circuits controlling stepping and standing (locomotion and mechanisms of activity dependent modulation of neuromotor posture) can be improved after a SCI by practicing those tasks, facilitated reorganization. i.e., by increasing the activation of those circuits (Ahissar and Hochstein, 1997; Bayona et al., 2005). The data presented here demonstrate that there is a minimal amount of spontaneous CONCLUSION activity in the sensorimotor circuits that control standing and stepping after a mid-thoracic spinal cord transection in adult rats We successfully demonstrated the proof of concept algorithm (Figures 4, 5) that can be further enhanced using multimodal to detect specific functional hind-limb spontaneous locomotion neuromodulatory mechanisms. In effect this enhanced level using minimal biosignals while monitoring biomarkers that of spontaneous activity can be viewed as a “self-training” represent a polysynaptic neurophysiological system input– phenomenon. Clinically, the presence of a “self-training” effect output response of the spinal circuitry that controls locomotive would have an immediate and significant impact on designing activity. In addition, we observed a correlation between these rehabilitation strategies. biomarkers and the functional responses to spinally evoked Frontiers in Systems Neuroscience | www.frontiersin.org 10 January 2020 | Volume 13 | Article 82 Chia et al. Step Detection and Neuromodulation After Paralysis potentials. By incorporating complex machine-learning methods performed the surgeries. RC, PG, and VE interpreted the data. and measuring kinematic output, the algorithm developed can be RC, HZ, PG, BV, and VE wrote the manuscript. All authors read readily scaled to detect more complex tasks. and approved the final manuscript. DATA AVAILABILITY STATEMENT FUNDING The datasets generated for this study are available on request to This study was funded by the Dana & Albert R. Broccoli the corresponding author. Charitable Foundation, Nanette and Burt Forester, including matching by PwC LLP and the Australian Department of Education Research Training Program. ETHICS STATEMENT The animal study was reviewed and approved by the Animal ACKNOWLEDGMENTS Research Committee at UCLA. All procedures described are in accordance with the National Institute of Health Guide for the The authors thank Valentine Zinchenko for his technical support. Care and Use of Laboratory Animals. SUPPLEMENTARY MATERIAL AUTHOR CONTRIBUTIONS The Supplementary Material for this article can be found RC developed the algorithm and performed the data analysis. PG online at: https://www.frontiersin.org/articles/10.3389/fnsys. performed the experiments and consulted in data analysis. HZ 2019.00082/full#supplementary-material REFERENCES eds McKerracher, L. Doucet, G. and Rossignol, S. (Amsterdem: Elsevier) 137, 141–149. doi: 10.1016/s0079-6123(02)37013-4 Ahissar, M., and Hochstein, S. (1997). Task difficulty and the specificity of de Leon, R. D., Tamaki, H., Hodgson, J. A., Roy, R. R., and Edgerton, V. R. (1999). perceptual learning. Nature 387, 401–406. doi: 10.1038/387401a0 Hindlimb locomotor and postural training modulates glycinergic inhibition Alaimo, M. A., Smith, J. L., Roy, R. R., and Edgerton, V. R. (1984). EMG activity of in the spinal cord of the adult spinal cat. J. Neurophysiol. 82, 359–369. doi: slow and fast ankle extensors following spinal cord transection. J. Appl. Physiol. 10.1152/jn.1999.82.1.359 Respir. Environ. Exerc. Physiol. 56, 1608–1613. doi: 10.1152/jappl.1984.56.6. Edgerton, V. R., Courtine, G., Gerasimenko, Y. P., Lavrov, I., Ichiyama, R. M., Fong, 1608 A. J., et al. (2008). Training locomotor networks. Brain Res. Rev. 57, 241–254. Alam, M., Garcia-Alias, G., Jin, B., Keyes, J., Zhong, H., Roy, R. R., et al. (2017). Edgerton, V. R., and Roy, R. R. (2009). Activity-dependent plasticity of spinal Electrical neuromodulation of the cervical spinal cord facilitates forelimb locomotion: implications for sensory processing. Exerc. Sport Sci. Rev. 37, skilled function recovery in spinal cord injured rats. Exp. Neurol. 291, 141–150. 171–178. doi: 10.1097/JES.0b013e3181b7b932 doi: 10.1016/j.expneurol.2017.02.006 Edgerton, V. R., Tillakaratne, N. J. K., Bigbee, A. J., de Leon, R. D., and Roy, R. R. Angeli, C. A., Edgerton, V. R., Gerasimenko, Y. P., and Harkema, S. J. (2014). (2004). Plasticity of the spinal neural circuitry after injury. Annu. Rev. Neurosci. Altering spinal cord excitability enables voluntary movements after chronic 27, 145–167. complete paralysis in humans. Brain 137(Pt 5), 1394–1409. doi: 10.1093/brain/ Fong, A. J., Cai, L. L., Otoshi, C. K., Reinkensmeyer, D. J., Burdick, J. W., Roy, awu038 R. R., et al. (2005). Spinal cord-transected mice learn to step in response to Aragão, R. D. S., Rodrigues, M. A. B., de Barros, K. M. F. T., Silva, S. R. F., Toscano, quipazine treatment and robotic training. J. Neurosci. 25, 11738–11747. doi: A. E., de Souza, R. E., et al. (2011). Automatic system for analysis of locomotor 10.1523/jneurosci.1523-05.2005 activity in rodents—A reproducibility study. J. Neurosci. Methods 195, 216–221. Formento, E., Minassian, K., Wagner, F., Mignardot, J. B., Le Goff-Mignardot, doi: 10.1016/j.jneumeth.2010.12.016 C. G., Rowald, A., et al. (2018). Electrical spinal cord stimulation must preserve Bayona, N. A., Bitensky, J., Salter, K., and Teasell, R. (2005). The role of task- proprioception to enable locomotion in humans with spinal cord injury. Nat. specific training in rehabilitation therapies. Top. Stroke Rehabil. 12, 58–65. Neurosci. 21, 1728–1741. doi: 10.1038/s41593-018-0262-6 doi: 10.1310/bqm5-6ygb-mvj5-wvcr Gad, P., Choe, J., Shah, P., Garcia-Alias, G., Rath, M., Gerasimenko, Y., et al. Byung-Woo, M., Ho-Sub, Y., Jung, S., Yun-Mo, Y., and Toshiaki, E. (1997). (2013a). Sub-threshold spinal cord stimulation facilitates spontaneous motor “Hand gesture recognition using hidden Markov models,” in Proceedings of activity in spinal rats. J. NeuroEng. Rehabil. 10, 108–108. doi: 10.1186/1743- the 1997 IEEE International Conference on Systems, Man, and Cybernetics. 0003-10-108 Computational Cybernetics and Simulation. Orlando, IL. Gad, P., Lavrov, I., Shah, P., Zhong, H., Roy, R. R., Edgerton, V. R., et al. (2013b). Capogrosso, M., Milekovic, T., Borton, D., Wagner, F., Moraud, E. M., Mignardot, Neuromodulation of motor-evoked potentials during stepping in spinal rats. J.-B., et al. (2016). A brain-spine interface alleviating gait deficits after spinal J. Neurophysiol. 110, 1311–1322. doi: 10.1152/jn.00169.2013 cord injury in primates. Nature 539, 284–288. doi: 10.1038/nature20118 Gad, P., Roy, R. R., Choe, J., Creagmile, J., Zhong, H., Gerasimenko, Y., et al. (2015). Clarke, R. L., Smith, R. F., and Justesen, D. R. (1985). An infrared device for Electrophysiological biomarkers of neuromodulatory strategies to recover detecting locomotor activity. Behav. Res. Methods Instr. Comput. 17, 519–525. motor function after spinal cord injury. J. Neurophysiol. 113, 3386–3396. doi: doi: 10.3758/bf03207645 10.1152/jn.00918.2014 Courtine, G., Gerasimenko, Y., van den Brand, R., Yew, A., Musienko, P., Zhong, Ganea, K., Liebl, C., Sterlemann, V., Müller, M., and Schmidt, M. (2007). H., et al. (2009). Transformation of nonfunctional spinal circuits into functional Pharmacological validation of a novel home cage activity counter in mice. states after the loss of brain input. Nat. Neurosci. 12:1333. doi: 10.1038/nn.2401 J. Neurosci. Methods 162, 180–186. doi: 10.1016/j.jneumeth.2007.01.008 de Leon, R. D., Reinkensmeyer, D. J., Timoszyk, W. K., London, N. J., Roy, R. R., Genewsky, A., Heinz, D. E., Kaplick, P. M., Kilonzo, K., and Wotjak, C. T. (2017). A and Reggie Edgerton, V. (2002). “Chapter 11 Use of robotics in assessing the simplified microwave-based motion detector for home cage activity monitoring adaptive capacity of the rat lumbar spinal cord,” in Progress in Brain Research, in mice. J. Biol. Eng. 11:36. doi: 10.1186/s13036-017-0079-y Frontiers in Systems Neuroscience | www.frontiersin.org 11 January 2020 | Volume 13 | Article 82 Chia et al. Step Detection and Neuromodulation After Paralysis Gerasimenko, Y., Gorodnichev, R., Moshonkina, T., Sayenko, D., Gad, P., and of the 2007 2nd International Conference on Pervasive Computing and Reggie Edgerton, V. (2015). Transcutaneous electrical spinal-cord stimulation Applications. Birmingham. in humans. Ann. Phys. Rehabil. Med. 58, 225–231. doi: 10.1016/j.rehab.2015. Parreno, A., Saraza, M., and Subero, C. (1985). A new stabilimeter for small 05.003 laboratory animals. Physiol. Behav. 34, 475–478. doi: 10.1016/0031-9384(85) Gerasimenko, Y., Roy, R. R., and Edgerton, V. R. (2008). Epidural stimulation: 90215-x comparison of the spinal circuits that generate and control locomotion in rats, Pernold, K., Iannello, F., Low, B. E., Rigamonti, M., Rosati, G., Scavizzi, F., et al. cats and humans. Exp. Neurol. 209, 417–425. doi: 10.1016/j.expneurol.2007. (2019). Towards large scale automated cage monitoring – Diurnal rhythm 07.015 and impact of interventions on in-cage activity of C57BL/6J mice recorded Gerasimenko, Y. P. I, Lavrov, A., Courtine, G., Ichiyama, R. M., Dy, C. J., Zhong, H., 24/7 with a non-disrupting capacitive-based technique. PLoS One 14:e0211063. et al. (2006). Spinal cord reflexes induced by epidural spinal cord stimulation in doi: 10.1371/journal.pone.0211063 normal awake rats. J. Neurosci. Methods 157, 253–263. doi: 10.1016/j.jneumeth. Redfern, W. S., Tse, K., Grant, C., Keerie, A., Simpson, D. J., Pedersen, 2006.05.004 J. C., et al. (2017). Automated recording of home cage activity and Gerasimenko, Y. P., Lu, D. C., Modaber, M., Zdunowski, S., Gad, P., Sayenko, temperature of individual rats housed in social groups: the rodent D. G., et al. (2015). Noninvasive reactivation of motor descending control after big brother project. PLoS One 12:e0181068. doi: 10.1371/journal.pone. paralysis. J. Neurotrauma 32, 1968–1980. doi: 10.1089/neu.2015.4008 0181068 Gill, M. L., Grahn, P. J., Calvert, J. S., Linde, M. B. I, Lavrov, A., Strommen, Rose, F., Dell, P., and Love, S. (1985). Doppler shift radar monitoring of activity J. A., et al. (2018). Neuromodulation of lumbosacral spinal networks enables of rats in a behavioural test situation. Physiol. Behav. 35, 85–87. doi: 10.1016/ independent stepping after complete paraplegia. Nat. Med. 24, 1677–1682. doi: 0031-9384(85)90175-1 10.1038/s41591-018-0175-7 Rossignol, S., and Frigon, A. (2011). Recovery of locomotion after spinal cord Harkema, S., Gerasimenko, Y., Hodes, J., Burdick, J., Angeli, C., Chen, Y., et al. injury: some facts and mechanisms. Ann. Rev. Neurosci. 34, 413–440. doi: (2011). Effect of epidural stimulation of the lumbosacral spinal cord on 10.1146/annurev-neuro-061010-113746 voluntary movement, standing, and assisted stepping after motor complete Roy, R. R., Hodgson, J. A., Lauretz, S. D., Pierotti, D. J., Gayek, R. J., and paraplegia: a case study. Lancet 377, 1938–1947. doi: 10.1016/S0140-6736(11) Edgerton, V. R. (1992). Chronic spinal cord-injured cats: surgical procedures 60547-3 and management. Lab. Anim. Sci. 42, 335–343. How, D. N. T., Sahari, K. S. M., Hu, Y., and Loo Chu, K. (2014). “Multiple Solnik, S., Rider, P., Steinweg, K., DeVita, P., and Hortobágyi, T. (2010). Teager- sequence behavior recognition on humanoid robot using long short-term Kaiser energy operator signal conditioning improves EMG onset detection. Eur. memory (LSTM),” in Proceedings of the 2014 IEEE International Symposium on J Appl. Physiol. 110, 489–498. doi: 10.1007/s00421-010-1521-8 Robotics and Manufacturing Automation (ROMA). Ipoh. Starkey, M. L., Bleul, C., Kasper, H., Mosberger, A. C., Zörner, B., Giger, S., et al. Huang, H., Kuiken, T. A., and Lipschutz, R. D. (2009). A strategy for identifying (2014). High-impact, self-motivated training within an enriched environment locomotion modes using surface electromyography. IEEE Trans. Bio-Med. Eng. with single animal tracking dose-dependently promotes motor skill acquisition 56, 65–73. doi: 10.1109/TBME.2008.2003293 and functional recovery. Neurorehabil. Neural Repair 28, 594–605. doi: 10.1177/ Huang, H., Zhang, F., Sun, Y. L., and He, H. (2010). Design of a robust EMG 1545968314520721 sensing interface for pattern classification. J. Neural Eng. 7, 056005–056005. Sy Bor, W., Quattoni, A., Morency, L., Demirdjian, D., and Darrell, T. (2006). doi: 10.1088/1741-2560/7/5/056005 “Hidden conditional random fields for gesture recognition,” in Proceedings of Ichiyama, R. M., Courtine, G., Gerasimenko, Y. P., Yang, G. J., van den the2006 IEEE Computer Society Conference on Computer Vision and Pattern Brand, R. I, Lavrov, A., et al. (2008a). Step training reinforces specific spinal Recognition (CVPR0 06). New York, NY. locomotor circuitry in adult spinal rats. J. Neurosci. 28, 7370–7375. doi: 10.1523/ Taccola, G., Sayenko, D., Gad, P., Gerasimenko, Y., and Edgerton, V. R. (2018). JNEUROSCI.1881-08.2008 And yet it moves: recovery of volitional control after spinal cord injury. Prog. Ichiyama, R. M., Gerasimenko, Y., Jindrich, D. L., Zhong, H., Roy, R. R., and Neurobiol. 160, 64–81. doi: 10.1016/j.pneurobio.2017.10.004 Edgerton, V. R. (2008b). Dose dependence of the 5-HT agonist quipazine in Tamborini, P., Sigg, H., and Zbinden, G. (1989). Quantitative analysis of rat activity facilitating spinal stepping in the rat with epidural stimulation. Neurosci. Lett. in the home cage by infrared monitoring. Application to the acute toxicity 438, 281–285. doi: 10.1016/j.neulet.2008.04.080 testing of acetanilide and phenylmercuric acetate. Arch. Toxicol. 63, 85–96. Ichiyama, R. M., Gerasimenko, Y. P., Zhong, H., Roy, R. R., and Edgerton, V. R. doi: 10.1007/bf00316429 (2005). Hindlimb stepping movements in complete spinal rats induced by Tsironi, E., Barros, P., Weber, C., and Wermter, S. (2017). An analysis epidural spinal cord stimulation. Neurosci. Lett. 383, 339–344. doi: 10.1016/j. of convolutional long short-term memory recurrent neural networks for neulet.2005.04.049 gesture recognition. Neurocomputing 268, 76–86. doi: 10.1016/j.neucom.2016. Lavrov, I., Dy, C. J., Fong, A. J., Gerasimenko, Y., Courtine, G., Zhong, H., et al. 12.088 (2008). Epidural stimulation induced modulation of spinal locomotor networks York, J. M., Blevins, N. A., McNeil, L. K., and Freund, G. G. (2013). Mouse short- in adult spinal rats. J. Neurosci. 28, 6022–6029. doi: 10.1523/jneurosci.0080-08. and long-term locomotor activity analyzed by video tracking software. J. Vis. 2008 Exp. 20:50252. doi: 10.3791/50252 Lavrov, I., Gerasimenko, Y. P., Ichiyama, R. M., Courtine, G., Zhong, H., Roy, Young, C., Young, M., Li, Y., and Lin, M. (1996). A new ultrasonic method for R. R., et al. (2006). Plasticity of spinal cord reflexes after a complete transection measuring minute motion activities on rats. J. Neurosci. Methods 70, 45–49. in adult rats: relationship to stepping ability. J. Neurophysiol. 96, 1699–1710. doi: 10.1016/s0165-0270(96)00102-1 doi: 10.1152/jn.00325.2006 Lee, K. (2008). EMG-based speech recognition using hidden markov models with Conflict of Interest: VE holds shareholder interest in NeuroRecovery global control variables. IEEE Trans. Biomed. Eng. 55, 930–940. doi: 10.1109/ Technologies and holds certain inventorship rights on intellectual property TBME.2008.915658 licensed by The Regents of the University of California to NeuroRecovery Martin, P., and Unwin, D. (1980). A microwave doppler radar activity monitor. Technologies and its subsidiaries. PG and VE hold shareholder interest in Behav. Res. Methods Instrument. 12, 517–520. doi: 10.3758/bf03201826 SpineX Inc. Megens, A., Voeten, J., Rombouts, J., Meert, T. F., and Niemegeers, C. (1987). Behavioural activity of rats measured by a new method based on the piezo- The remaining authors declare that the research was conducted in the absence of electric principle. Psychopharmacology 93, 382–388. any commercial or financial relationships that could be construed as a potential Mendes, C. S., Bartos, I., Márka, Z., Akay, T., Márka, S., and Mann, R. S. (2015). conflict of interest. Quantification of gait parameters in freely walking rodents. BMC Biol. 13:50–50. doi: 10.1186/s12915-015-0154-0 Copyright © 2020 Chia, Zhong, Vissel, Edgerton and Gad. This is an open-access Morrel-Samuels, P., and Krauss, R. M. (1990). Cartesian analysis: a computer-video article distributed under the terms of the Creative Commons Attribution License interface for measuring motion without physical contact. Behav. Res. Methods (CC BY). The use, distribution or reproduction in other forums is permitted, provided Instr. Computers 22, 466–470. doi: 10.3758/bf03203196 the original author(s) and the copyright owner(s) are credited and that the original Naeem, U., and Bigham, J. (2007). “A comparison of two hidden markov publication in this journal is cited, in accordance with accepted academic practice. No approaches to task identification in the home environment,” in Proceedings use, distribution or reproduction is permitted which does not comply with these terms. Frontiers in Systems Neuroscience | www.frontiersin.org 12 January 2020 | Volume 13 | Article 82