Participant screening
Ethical approval was obtained from the University of California, Irvine Institutional Review Board (Irvine, CA, USA). Candidates were recruited from a population of individuals with chronic T6 – T12 SCI. They underwent several screening procedures to rule out severe spasticity, contractures, restricted range of motion, lower extremity fractures, pressure ulcers, severe osteoporosis, orthostatic hypotension, as well as affirm neuromuscular responsiveness to FES (see Additional file 1 for details). A physically active 26-year-old male with a T6 AIS B SCI, with no motor function in the lower extremities and no sensation below the injury level except for minimally preserved bladder fullness sensation, passed all the screening requirements. He provided informed consent to participate in the study. He also consented to the publication of the biomedical data and media, including photographs and videos (consent to publish was also obtained from every person featured in these photographs and videos).
Training procedure
The participant underwent BCI training to learn how to ambulate within a VRE using attempted walking and idling (i.e. relaxing) as a control strategy. This procedure also generated an EEG decoding model that was subsequently used in BCI-FES experiments. In addition, since the supraspinal areas underlying human gait can become suppressed after chronic SCI, it has been suggested that motor imagery practice may facilitate their reactivation [17]. Therefore, the purpose of the BCI-VRE training was to also facilitate the reactivation of the brain areas responsible for gait control. Finally, the participant simultaneously underwent FES training to recondition his lower extremity muscles in order to be able to stand and walk overground using a FDA-approved commercial FES system (Parastep I System, Sigmedics, Fairborn, OH).
BCI training
Similar to our prior studies [10, 15, 16], the participant first underwent a BCI screening procedure to determine if he could control the BCI in a VRE. Subsequently, he underwent BCI training in order to further master BCI-VRE control. Each BCI screening and training visit entailed the same procedure that began with a 10-min electroencephalogram (EEG) recording. During this period, the participant engaged in 30-s-long alternating epochs of attempted walking and idling while seated in his wheelchair [10, 16]. A detailed description of this procedure is given in Additional file 1.
Based on these data, an EEG decoding model was generated offline using the methods described in [10, 15, 16]. Briefly, the EEG epochs were segmented into 4-s-long trials of “Idle” and “Walk” class, transformed into the frequency domain, and their power spectral densities (PSDs) were integrated from 6 to 40 Hz in 2-Hz bins. These spatio-spectral data were then subjected to dimensionality reduction using classwise principal component analysis (CPCA) [18, 19], and discriminating features were extracted using approximate information discriminant analysis (AIDA) [20]. Note that this feature extraction method is rooted in information theory [21] and has been extensively tested in our prior BCI studies [10, 15, 16, 22, 23]. More formally, one-dimensional (1D) features \(f\in \mathbb {R}\) were extracted by:
$$ f = \mathbf{T} \Phi(\mathbf{d}), $$
((1))
where \(\mathbf {d}\in \mathbb {R}^{B\times C}\) is a single trial of spatio-spectral data (B–number of frequency bins, C–number of electrodes), \(\Phi :\mathbb {R}^{B\times C}\rightarrow \mathbb {R}^{m}\) is a mapping from the data space to an m-dimensional CPCA-subspace, and \(\mathbf {T}:\mathbb {R}^{m}\rightarrow \mathbb {R}\) is an AIDA transformation matrix.
A Bayesian classifier was then designed as follows:
$$ f^{\star}\in\left\{{\vphantom{\begin{aligned} &\mathcal{S}_{1},\quad \text{if}\quad P(\mathcal{S}_{1}|\,\,f^{\star})>P(\mathcal{S}_{2}|f^{\star})\\ &\mathcal{S}_{2}, \quad \text{otherwise} \end{aligned}\quad,}}\right. \begin{aligned} &\mathcal{S}_{1},\quad \text{if}\quad P(\mathcal{S}_{1}|\,f^{\star})>P(\mathcal{S}_{2}|\,f^{\star})\\ &\mathcal{S}_{2}, \quad \text{otherwise} \end{aligned}\quad, $$
((2))
where \(P(\mathcal {S}_{1}|\,f^{\star })\) and \(P(\mathcal {S}_{2}|\,f^{\star })\) are the posterior probabilities of idling and walking classes, respectively, given the observed feature, f
⋆. They were found using the Bayes rule \(P(\mathcal {S}_{i}|\,f^{\star })=p(\,f^{\star }|\mathcal {S}_{i})P(\mathcal {S}_{i})/p(\,f^{\star })\), i=1,2, where \(p(\,f^{\star }|\mathcal {S}_{i})\) is a conditional probability density function (PDF) evaluated at f
⋆, \(P(\mathcal {S}_{i})\) is the prior probability of the class, \(\mathcal {S}_{i}\), and p(f
⋆) is the (unconditional) PDF. To simplify calculations, the conditional PDFs were modeled as Gaussians with equal variances. Note that this rendered the Bayesian classifier (2) linear [24]. The performance of the classifier was evaluated offline through stratified ten-fold cross-validation [25].
Each visit continued with online BCI operation, where 0.75-s-long segments of EEG data were wirelessly acquired in real time every 0.25 s using a sliding window approach. The PSDs of the EEG channels were then calculated and integrated in 2 Hz-bins for each of these segments, and used as the input for the EEG decoding model. The posterior probabilities, \(P(\mathcal {S}_{1}|\,f^{\star })\) and \(P(\mathcal {S}_{2}|\,f^{\star })\), were calculated using the Bayes rule (see above), and were averaged over a 1.5–2.0 s window to minimize false alarms and omissions [10, 15, 16]. Before online BCI operation, the BCI-VRE system was calibrated using a short procedure (see Additional file 1 for details). During each online experiment, the participant performed between one and five goal-oriented, real-time BCI walking tasks. Specifically, he was instructed to utilize attempted walking and idling to control the linear ambulation of an avatar and make sequential stops at ten designated points within the VRE [14–16]. The goal of the task (see Fig. 1) was to walk the avatar at a constant speed and complete the course as quickly as possible, while dwelling at each stop for at least 2 s. The online performances, expressed as the number of successful stops and course completion time, were compared to the results of Monte Carlo simulations to ascertain whether control of the BCI system was purposeful (details in Additional file 1). Note that despite demonstrating purposeful control during the BCI screening process, the participant continued the BCI-VRE training throughout the study. This provided the EEG decoding model for subsequent BCI-FES experiments. It also allowed the participant’s BCI-VRE performance to be tracked over time and the presumed reactivation of the cortical gait areas to occur.
FES training
To better understand the FES training procedures, a brief description of the Parastep system’s operation is first provided. Namely, the Parastep achieves ambulation by activating the quadriceps and tibialis anterior muscles. This is accomplished by placing electrode pairs bilaterally over the femoral (immediately proximal to the knee) and deep peroneal (immediately distal to the knee) nerves. Simultaneous bilateral activation of the quadriceps is used to maintain the knee extension necessary for standing, while a front-wheel walker is used for upper body stabilization. A step is achieved with the following sequence: 1. the user performs an anterior-lateral weight shifting maneuver; 2. a brief electrical stimulation is delivered unilaterally to the deep peroneal nerve while the corresponding quadriceps are deactivated, thereby eliciting a triple-flexion reflex of the leg (i.e. combination of foot dorsiflexion, knee flexion, and hip flexion); 3. the user’s leg swings forward due to the anteriorly shifted center of gravity; 4. the quadriceps are reactivated to maintain a standing position. The Parastep system’s adjustable parameters are the step duration (controlled manually by the subject via buttons) and stimulation current for bilateral femoral and deep peroneal nerves. Based on these five parameters, the system generates pre-programmed stimulation sequences for walking movements.
The FDA-approved guidelines for the Parastep system require users to recondition their muscles prior to engaging in FES-mediated walking. This reconditioning also facilitates improved cardiopulmonary endurance. To this end, the participant performed strength and endurance training of the quadriceps using the FES device. Once the participant regained sufficient strength and endurance, and demonstrated the ability to stand using the FES system, the training sessions progressed to FES-assisted overground walking. This included learning the coordination of movements such as weight shifting, front-wheel walker advancement and leg swing, which facilitate FES-mediated walking. A more detailed description of these procedures is provided in Additional file 1. It should be noted that the FES training was also used to empirically determine the stimulation parameters. More specifically, the time necessary to perform the weight shifting, walker advancement, and leg swing determined the step rate. The stimulation amplitude for each femoral nerve was determined as the minimal amount of current necessary to achieve a standing posture. Similarly, the stimulation amplitude for each peroneal nerve was determined by finding the minimal current necessary to elicit an adequate triple-flexion response and step. Note that these parameters were later used in the BCI-FES experiments as described below.
The FES training continued until the participant could walk the length of the overground walking course (3.66 m) without any intervention from the physical therapist. To prevent falls and provide partial body-weight support, FES walking was performed while the participant was mounted in a body-weight support system (ZeroG, Aretech, Ashburn, VA).
BCI-FES Experiments
The BCI-FES walking experiments were initiated once the participant completed the FES training. This was accomplished by first integrating the BCI and FES systems using a dedicated microcontroller. In addition, the step rate and stimulation amplitudes (as determined above) were pre-programmed into the microcontroller such that the left and right steps cycle automatically. A motion sensor system was then developed and synchronized with the BCI-FES system for the purpose of facilitating the performance assessment. A more detailed description of these steps is provided in Additional file 1. Finally, the EEG decoding model from the most recent BCI training session was loaded into the BCI system. The participant then undertook suspended BCI-FES walking tests followed by overground BCI-FES walking tests.
Suspended walking tests
Prior to overground walking, suspended walking tests were performed to establish whether the participant could purposefully operate the BCI-FES system. First, the participant was positioned ∼1 m from a computer screen and suspended using the ZeroG support system so that his feet were ∼5 cm off the ground (see Fig. 2). This allowed the execution of BCI-FES-mediated walking and standing without having to maintain postural stability, perform weight shifting, or advance the front-wheel walker. The participant then followed 30-s-long alternating “Idle” and “Walk” visual computer cues for a total of 180 s with the goal of controlling the standing and walking functions of the BCI-FES system in real time. Finally, the participant’s performance (details below) was assessed using video, BCI state, and motion sensor data.
Overground walking tests
For overground walking tests, the participant utilized the system to walk along a 3.66-m-long linear course with three cones positioned 1.83 m apart (Fig. 1). He was instructed to walk and stand at each cone for 10–20 s via verbal cues given by the experimenter. Subsequently, he used an attempted walking strategy to initiate BCI-FES-mediated walking to progress to the next cone. Note that the duration of standing at each cone was randomized to minimize anticipation by the participant. Also note that the ZeroG system was used during these tests to provide partial body-weight support and prevent falls. Overground walking tests were repeated as tolerated by the participant. Video, BCI state, and motion sensor data were recorded to assess the performance during this task.
Performance assessment
The subject’s performances in the suspended and overground walking tests were derived based on the video, BCI state, and gyroscope data. Specifically, they were quantified by calculating the cross-correlation and information transfer rate (ITR) between the externally supplied cues and BCI-FES-mediated responses. In the suspended walking tests, the timings of the visual cues were obtained from the BCI computer. In the overground walking tests, the timings of the auditory cues were extracted from the video recordings. In both types of tests, the epochs of BCI-FES mediated responses were extracted from the gyroscope data. Similar to above, purposeful BCI-FES control was ascertained by comparing these cross-correlations to those achieved by Monte Carlo simulations (details in Additional file 1). In addition, the instances of false alarms and omissions were recorded, where a false alarm was defined as the presence of BCI-FES-mediated walking within any intended idling epoch, while an omission was defined as the absence of BCI-FES-mediated walking within any intended walking epoch. Finally, in the overground walking tests, the laser distance meter was used to confirm that the subject ambulated along the course and stopped at the cones.