Long-Term Training with a Brain-Machine Interface-Based Gait Protocol Induces Partial Neurological Recovery in Paraplegic Patients

Brain-machine interfaces (BMIs) provide a new assistive strategy aimed at restoring mobility in severely paralyzed patients. Yet, no study in animals or in human subjects has indicated that long-term BMI training could induce any type of clinical recovery. Eight chronic (3–13 years) spinal cord injury (SCI) paraplegics were subjected to long-term training (12 months) with a multi-stage BMI-based gait neurorehabilitation paradigm aimed at restoring locomotion. This paradigm combined intense immersive virtual reality training, enriched visual-tactile feedback, and walking with two EEG-controlled robotic actuators, including a custom-designed lower limb exoskeleton capable of delivering tactile feedback to subjects. Following 12 months of training with this paradigm, all eight patients experienced neurological improvements in somatic sensation (pain localization, fine/crude touch, and proprioceptive sensing) in multiple dermatomes. Patients also regained voluntary motor control in key muscles below the SCI level, as measured by EMGs, resulting in marked improvement in their walking index. As a result, 50% of these patients were upgraded to an incomplete paraplegia classification. Neurological recovery was paralleled by the reemergence of lower limb motor imagery at cortical level. We hypothesize that this unprecedented neurological recovery results from both cortical and spinal cord plasticity triggered by long-term BMI usage.

of SCI), aged between 26-38 years, seven of them with complete injury (ASIA A) and one with incomplete injury (ASIA B), were included in an intensive gait training routine involving a BMI paradigm applied to both immersive virtual reality environments and robotic gait systems enriched with the ability to provide tactile feedback. The complexity of activities were increased over time, to ensure cardiovascular system stability and better postural control of the patients, starting with orthostatic training at stand-in-table device and progressing to robotic gait training, using a greater body-weight support (BWS) system on a treadmill (Lokomat) 2 . We then gradually decreased the BWS over time on a fixed overground track gait BWS system (ZeroG) 3 . A robotic exoskeleton was also used for the gait training. Clinical and BMI activities were integrated by proposing a multi-step BMI protocol, starting with training patients to use their EEG signals to interact with a brain-controlled 3D avatar, rendered in an immersive virtual reality setup.
Next, patients learned to use the same EEG signals to control key locomotion movements generated by brain-controlled robotic gait devices (Lokomat and exoskeleton).

B1. Clinical Protocol
Clinical activities included traditional physical therapy with stretching and strengthening exercises and the use of a body-weight support (BWS) system for gait training: BWS ambulation on a treadmill and BWS ambulation on a fixed overground track 3 .

Gait training using BWS on a treadmill system
The BWS used in combination with a treadmill system (Lokomat, Hocoma) included an electrically controlled gait orthosis composed of a hip support and two running orthotics, which allowed command of the hip and knee movements.
Physiotherapists were able to control parameters such as the treadmill speed (1 to 2km/h), the body weight support (0 to 100%), the guidance force (0 to 100%) and the range of motion of the robotic joints (42-45° for hip flexion, 63-66° for knee flexion).
Patients were also allowed to actively move the orthosis, in synergism with the computer-generated orthosis movements. The same equipment allowed us to assess spasticity, muscular strength during static and dynamic states, estimate active range of motion of the joints and distance walked. We fixed the range for BWS from 75-100%.
Guidance force was fixed at 100% and the speed of the treadmill was set at 1km/h with the goal of promoting the safest possible training environment. Because subjects were not able to report on pain sensitivity, BWS was limited by the maximum knee extension without joint collapse during the stance phase of gait (knee collapse was observed about 75% of BWS, establishing a limit for further BWS reductions in this device). By using 100% of guidance force, the range of joint motion was adjusted by the physiotherapist who fixed the minimum and maximum values. That meant that joint control was maintained by the equipment. As guidance force was reduced and some of their weight was not supported directly by the BWS, subjects were in charge of adjusting their own posture.

Gait training using overground BWS system
The overground BWS system (ZeroG, Aretech LLC, Ashburn, VA) employed in our study uses a static or dynamic BWS while it rides along an overhead U-shaped fixed track. This device allows a greater freedom of movement and a better interaction between therapist and subject because there are no mechanical barriers among them.
In this context, this equipment challenges the subjects more, by requiring postural control, trunk control, upper limb strength and dynamic balance during gait training. The device allows the control of BWS, speed, and provides a measurement of the distance walked. It also offers gait and balance training (static and dynamic) on different surfaces if necessary. The BWS contains an anti-fall system which provides greater security during all activities.
During gait training, subjects wore lower limb orthosis on both legs and a walking assistive device (hip-knee-ankle-foot orthosis [HKAFO] or ankle-foot orthosis [AFO] with knee extension splint and wheeled triangular walker). The following parameters were used during the session: range for BWS from 30 to 75%, fixed speed at 3km/h and fall distance between 5 and 10cm (which means that the displacement of the string at 5cm or 10cm triggers the anti-fall feature). Since patients started training with the Lokomat, achieving 75% of BWS, the ZeroG training started with 75%, progressing to 30% in some cases. The walking distance varied depending on the subject and on the subsequent sessions (individual upper limb strength and resistance, and cardiopulmonary performance).

Brain Machine Interface (BMI) algorithm
A 16-channel EEG cap was used for all the experiments that involved the recording of cortical signals and their use in controlling virtual or robotic devices. In the initial phases of training, patients were instructed to imagine movements of their hands or arms. EEG electrodes were placed in order to maximize the recording area over the arm representation in the sensorimotor cortex. EEG patterns produced by this motor imagery were then decoded using a linear discriminant analysis (LDA), using features extracted by a 6-dimensional common spatial pattern (CSP) to construct an EEG classifier. Patients used the same motor imagery strategy to control both the simulated virtual avatar, the Lokomat, and the exoskeleton. They first selected a high level state of the actuator (for example 'walk' state) and then confirmed and triggered the execution of this motor command by performing an isometric contraction of the arm triceps. Using this simple strategy, patients could perform four primary commands 'stand' (except with the Lokomat), 'walk', 'stop', and 'kick' A second control paradigm was introduced 7 months after onset of training. EEG signals were recorded over the leg sensorimotor cortex area. In this second strategy, subjects imagined moving their left and right legs to control the stepping of the ipsilateral legs. By alternating between left and right, stepping patients controlled the walking pattern of the avatar or the exoskeleton.

Generation and delivery of tactile feedback related to locomotion
The key objective of our strategy for tactile feedback was to provide SCI patients with key sensory information, lacking from their lower body, in a non-invasive way to help generate the most realistic walking experience possible. During training in the virtual environment, virtual tactile signals were generated every time the avatar feet touched the ground. During training with robotic devices, artificial tactile information was generated by distance sensors placed on the patients immobilized legs and feet, in the case of the Lokomat, or in key locations of the exoskeleton, such as the plantar surface of the robotic feet. In both cases, contact with the ground generated a wave of pressure signals that could be delivered to the patients' forearm skin via a haptic display. By using this haptic display, all patients were able to sense the position of their legs in space and the contact of their (or the exoskeleton's) feet with the ground. In some experiments, the same arrangement was also employed to allow patients to experience the contact of the exoskeleton feet with a soccer ball during a "kicking" movement.
The haptic display employed vibrators (ERM vibrator consisting of a DC motor rotating an eccentric mass at different angular velocities) similar to the ones found in cellphones. Three vibrators spaced 6cm apart were integrated in the long sleeves of a shirt. These vibrators are coin-shaped and because of their small size (10mm x 2 mm) they can be easily integrated into a wearable tactile interface. This allowed the generation of various amplitudes and frequencies of vibration. ERM frequency and amplitude were coupled, and the maximum amplitude was reached at a frequency of 220 Hz which corresponds to the peak response frequency of Pacinian Corpuscles.
Exploiting a haptic illusion called the 'Apparent Movement Illusion' 4,5 , it was possible to produce the sensation of one continuous tactile feedback moving along the patients' forearm by sequentially triggering the three equidistant vibrators. While brain controlling the virtual avatar, the Lokomat or the exoskeleton, patients received tactile feedback moving from their wrist to their elbow in synchrony with the rolling of the foot on the virtual or physical floor 6 .

Virtual reality environment
Three virtual avatars (one female and two males) were designed and rendered

Brain-controlled exoskeleton sensorized to deliver tactile feedback
A custom brain-controlled robotic exoskeleton (EXO) was designed for the execution of this project. To maximize power-to-weight ratio, this EXO employed electric motors and oil transmission hoses for a 12 degree of freedom actuation. The EXO was designed to be anatomically coherent with the body of our subjects. The hip-to-knee segments of the legs could be adjusted to accommodate a variety of different leg lengths. The EXO was stable in single support stance without the need of crouch, liberating the arms of the patients to execute any type of upper limb behavior. Patients could control the high level states of the EXO using EEG signals, while low level stabilization was done automatically by the robot.
Pressure sensors, wire sensors and gyroscopes were used for the PID controller of the exoskeleton to insure that the exoskeleton followed the correct trajectory.
Strain gauges and multimodal sensors 7 covered the exoskeleton's feet to detect forces exerted on the ground and confirm single stance and double stance positions.
The information was conveyed to the patient through the tactile shirt.

C1. Medical Evaluation
The medical team provided integral support for the subjects during the research,   contraction, or sacral sensory sparing with sparing of motor function more than three levels below the motor level, and the majority of key muscles have muscle grade less than 3. The ASIA allows even non-key muscle function more than three levels below the motor level to be used in determining motor incomplete status (B versus C); ASIA D is defined by presence of motor functions below the neurological level of the SCI and by at least half of key muscles below the neurological level having a muscle grade greater than or equal to 3; ASIA E is defined by normal sensory and motor functions (an individual without an initial SCI does not receive an ASIA grade).

Semmes-Weinstein Monofilament Test
This evaluation was done using nylon filaments of different thickness, distinguished by color and weight (blue 0.20g, purple 2.0g, red 4.0g, orange 10.0g, pink 300g). Originally used to test sensitivity in extremities (hand and foot), this test was applied to the trunk area in the present study, in order to expand and specify our tactile sensitivity evaluation. Each filament refers to a particular level of tactile sensitivity. As such, this test was used in order to better investigate multiple aspects of somatosensory sensation/discrimination. The monofilaments were used in a similar manner to the pin, by carrying out a comparison between the first stimulus done on the face (reference of normal sensitivity) and the second done in a thoracic dermatome or lower limb. The participant reported their perception of the second stimulus: normal, altered sensitivity, or absence of sensitivity.

Temperature, Pressure, Vibration, Proprioception
Temperature was evaluated using a dry cotton ball for warm sensation and an alcohol-soaked swab for cold sensation in every dermatome, on both sides of the body (right and left). Pressure pain (deep pain sensitivity) was assessed by using a dynamometer (10g/mm2, maximum 8kg) applied to every dermatome (right and left sides). Vibration was evaluated using a diapason in a bone surface. The initial stimulation was delivered in an area proximal to the level of the SCI: upper limb (elbow) and upper trunk (third or fourth rib). Then, the patient's perception was compared to the stimulation delivered to a distal area of the level of the injury. The sites that were stimulated were standardized in advance, but the sequence of the stimulation was

C2. Physical Therapy Evaluation
Our physical therapy and medical teams managed the physical training and

Lokomat L-force Evaluation
In order to promote a more accurate quantitative analysis on muscle strength, we Concomitant with the assessment of L-force, we also recorded electromyographic (EMG) activity generated by the lower limb musculature. A total of eight surface electrodes were used to capture EMG signals of four muscle groups in each leg. These included: rectus femoris proximal portion, gluteus maximus, medial hamstring and rectus femoris distal portion. The choice of these muscles was based on their motor functions studied by L-force: hip flexion (in this case, the most superficial muscles which assists in hip flexion corresponds to the proximal rectus femoris), hip extension (gluteus maximus), knee flexion (medial hamstring is the muscle group of choice in these patients, because it is easier for its identification) and knee extension (distal rectus femoris).

Thoracic-Lumbar Control Scale
The

Walking Index Spinal Cord Injury II
The WISCI II is a revised version of the original measurement WISCI. It evaluates gait performance in SCI patients on a 10 meters route, based on the need to use assistive walking devices, braces, or physical assistance from a therapist. The rank ranges from 0, which means inability to keep a standing position, to 20, which refers to ambulation with no devices, no braces and no physical assistance. The measurement was applied 5 times (beginning and end of first semester of 2014, and three times in the second semester of 2014).

Spinal Cord Independence Measurement III (SCIM third version)
This test was used to evaluate the level of independence of the participants in daily activities and monitor their progress throughout the training. The scale was applied five times in 2014, three times in the first half and twice in the second half of the year. It classifies the ability to perform daily tasks, such as personal care, respiratory condition and sphincter control (bladder and bowel) and mobility in the home and community environment. The score ranges from 0 (totally dependent) to 100 (fully independent).

The McGill Pain Questionnaire / Visual Analogue Scale (VAS)
This is a self-report pain questionnaire that contains three sections. The evaluations were performed during five periods over the 12 months of training, in order to verify the existence, intensity, location and behavior of pain and how the patient was connected with it. The study of pain was important to guide its treatment when necessary and to assist in the investigation of neurological recovery. The Visual Analogue Scale is a simple assessment used to quantify pain. The patient is asked to draw a mark on a 10cm size horizontal line. The perception of pain increases on a scale that ranges from 0 to 10.

Medical Research Council Scale
This measurement has a component that quantitatively evaluates joint mobility (range of motion) passively, using a goniometer (CARCI®). The joints studied included the hip, knee and ankle, bilaterally, in the following movements: flexion/extension, abduction/adduction, internal/external rotation of the hip, flexion/extension of the knee, dorsi/plantar flexion of the ankle. This evaluation was performed during six periods over the 12 months of the study, in order to monitor the mobility during physical interventions.

Modified Ashworth Scale (MAS) and Lokomat L-stiff Evaluation
The The purpose of applying both instruments was to contribute to the treatment of spasticity and to monitor the physical safety of patients 7 .

C3. Psychological Evaluation
Our psychology team provided continuous support to the patients during the to 3, which imply increasing degrees of depression. The total score allows classifying the level of depression as minimal (score: 0 to 9), mild (score: 10 to 16), moderate (score: 17 to 29) and severe (score: 30 to 63) 21 .
The Rosenberg Self-Esteem Scale (RSE) is a one-dimensional instrument that consists of 10 statements with four options that assess global self-esteem. The four possible answers are evaluated also using a Likert scale, ranging from "strongly agree" to "strongly disagree" and a score that ranges between 0 and 3. The final result allows the classification of self-esteem in low (score: 0 to 14), normal (score: 15 to 25) and high (score: 26 to 30) 22 .

EEG data analysis was performed with custom Matlab R2012a (The MathWorks
Inc., MA) routines and EEGLAB 13.3.2b functions 23 . First, the raw EEG data was bandpass-filtered with a causal zerophase delay correction using a 1 Hz FIR high-pass filter (order 4000) and a 50 Hz FIR low-pass filter (order 36).
After that, we removed bad channels based on the kurtosis of data points from each channel using 5 standard deviations from the mean threshold limit (good channels present voltage values that are closer to a Gaussian distribution than noisy channels).
Data was re-referenced to a common average reference calculated from the remaining EEG channels, and one channel was removed to avoid rank-deficiency in the Independent Component Analysis (ICA) algorithm. We then extracted -1 to 3s epochs from the continuous EEG data with respect to the event times. Bad epochs were identified using an amplitude threshold of -500 to 500 uV and a probability test with a 5 standard deviation from the mean threshold. In sequence, we decomposed the data into independent components (ICs) by performing an ICA using the JADE algorithm 24 , followed by an equivalent dipole current source fitting procedure in which we only kept dipoles located within the brain and that had a maximum residual variance from the IC scalp projection of 15%; for this latter calculation, the DIPFIT EEGLAB toolbox 25 was used with a spherical head model in which each subject's electrode coordinates were aligned with the surface of a standard brain template. In order to cluster ICs from different subjects, we built a feature vector with their 5-25 Hz power spectrum, equivalent dipole position, and ICs scalp projection. The resulting vector dimension was reduced with a principal component analysis, and the first 10 principal components were used as inputs to a k-means clustering algorithm. ICs that were located more than 3 standard deviations from the nearest cluster center were considered outliers and discarded. Clusters with components from less than half of subjects were discarded.
For each cluster and condition, we performed a series of measurements to highlight task-related brain dynamics modulations. Time-frequency maps revealing event related spectral perturbations (ERSPs) were calculated with respect to a baseline of 1s prior to the event and normalized by the average power across trials at each frequency; a 3-cycle Morlet wavelet was employed to obtain the frequency spectrum in every time window. Significant perturbations (p<.05) were determined using a nonparametric permutation method with 2000 surrogate data sets and this was subsequently used for masking the ERSP plots for significance. In addition, average event related potentials (ERPs) were determined for each subject in all conditions and submitted to a permutation statistical test for assessing significant effects of conditions in each channel ERP. Figure S1. Details per patient of the sensitivity to vibration in the lower limbs. Patient's sensitivity to vibration on eight leg bones presented in a craniocaudal sequence. Score convention was the following: 0 for absence of sensation, 1 for altered sensation and 2 for normal sensation. The measurement was introduced at month 2 of the study. The patients were asked to compare a first stimulation performed in an area proximal to the SCI level with a