A Systematic Review Establishing the Current State-of-the-Art, the Limitations, and the DESIRED Checklist in Studies of Direct Neural Interfacing With Robotic Gait Devices in Stroke Rehabilitation

Introduction

Stroke, a disease with substantial personal and societal consequences, remains the leading cause of acquired disability worldwide. With 13.7 million new cases each year, the associated economic costs of treatment and post-stroke care are significant (Wilkins et al., 2017; Johnson et al., 2019). At 3 months after a stroke, 20% of people remain wheelchair dependent and ~70% walk with a reduced capacity (Mehrholz et al., 2017). Task-specific training is critical for recovery, and the intensity of practice is strongly associated with improved functional gait outcomes (Kwakkel et al., 2004; Veerbeek et al., 2014).

Providing high intensity restorative exercises for a larger share of the stroke population is part of the Action Plan for Stroke in Europe 2018–2030 (Norrving et al., 2018), yet the delivery of an adequate dosage of gait training for physically dependent patients is challenging in the rehabilitation sector, from manual handling and human resource perspectives. Robotic gait devices, which enable people to walk with electromechanical assistance to achieve a healthy gait trajectory, can potentially overcome some of these practical difficulties (Mehrholz et al., 2017; Cervera et al., 2018) and allow an intensive, high repetition of the gait cycle with reduced therapist involvement (as they no longer need to set the paretic limbs or assist trunk movements) (Sarasola-Sanz et al., 2017). The addition of robotic-assisted gait training (RAGT) to usual rehabilitation has been shown by a systematic review to improve the likelihood of regaining independent walking after a stroke [odds ratio 1.94, 95% confidence interval (CI), 1.39 to 2.71], with a subgroup analysis suggesting that people in the acute phase and non-ambulatory individuals benefit most from the intervention (Mehrholz et al., 2017). Of note is that the improvements in walking velocity and walking capacity did not match the observed improvements in independence in gait.

At present, RAGT alone has not been shown to be superior to equally dosed routine rehabilitation despite the increased intensity of stepping in RAGT (Taveggia et al., 2016; Bruni et al., 2018). When motor function is considered specifically as an outcome, the upper limb robotic devices have proven efficacy in contrast to the lower limb robotic training, where no treatment effect for motor function has been demonstrated (Lin et al., 2019). Current RAGT therapies have focused on providing high-intensity training and repetition but not on patient engagement, motivation, and reward, which are important factors for inducing cortical plasticity (Hogan et al., 2006; Goodman et al., 2014). Limitations in randomized controlled trials (RCTs) in this area to date have been identified (Molteni et al., 2018) and many RAGT protocols were criticized for allowing the trainee to be too passive, with lower metabolic costs, muscle activations, and subject effort reported in comparison to therapist-assisted treadmill training (Cai et al., 2006; Israel et al., 2006; Krishnan et al., 2013). However, rehabilitative robotics, when deployed correctly, have the ability to generate bottom up and top down complex and controlled multisensory stimulation to modify the plasticity of neural connections through the experience of movement (Molteni et al., 2018).

Direct human machine interfaces (HMIs) can translate electrical, magnetic, or metabolic activity at the brain or the muscle level into control signals for external devices (e.g., computers or neuroprosthetic and robotic devices) to replace, restore, or enhance the natural neural output (Wolpaw, 2012; Soekadar et al., 2015). Brain interfacing technology, primarily electroencephalography (EEG)-based brain computer interfaces (BCI,) can offer new strategies for robotic gait re-education after a stroke that can promote more active engagement in movement intent and/or by neurophysiological feedback. In stroke, BCI exploitation has mainly used motor imagery to drive brain activity levels (with no overt motor output) in combination with visual, auditory, or haptic feedback or to control an external device which executes the movement and provides proprioceptive feedback (Prasad et al., 2010; Van Dokkum et al., 2015). Of the nine upper limb studies identified in a recent systematic review of BCI for motor rehabilitation after a stroke, only three used BCI to control a robotic or orthotic device with large to moderate effect sizes noted for improved motor impairment (Cervera et al., 2018) and emerging evidence in upper limb rehabilitation now points to the superiority of BCI robotic training after a stroke to robotic training alone in motor recovery (Varkuti et al., 2013; Ang et al., 2014). No lower limb robotic RCT studies using BCI were reported in this review (Cervera et al., 2018).

Motor intent can also be determined non-invasively by electromyography (EMG) activity and responded to in triggered motion (Hussein et al., 2009) and thus has potential to enhance RAGT. EMG-based robotic movement has emerged as a well-developed field in upper limb rehabilitation in stroke (Ho et al., 2011; Vaca Benitez et al., 2013; Hu et al., 2015), and when used in robot-assisted rehabilitation has achieved a significantly higher completion rate compared to torque control for the participants with severe to moderate impairment in the upper limb (Paredes et al., 2015). EMG has also been combined with EEG in human–machine interactive force to improve the recognition of movement intent (Mrachacz-Kersting et al., 2012; Jiang et al., 2014; Bhagat et al., 2016).

As reported in a 2018 review of human intent-based control in motor rehabilitation after a stroke, most studies are in the laboratory stage (Li et al., 2018), and a systematic review of RCTs of BCI interfaces after a stroke identified no RAGT studies (Cervera et al., 2018). Therefore, the aim of this systematic review was to establish the current state-of-the-art in EMG and/or EEG neural biosignal deployment during robotic gait training post-stroke as described in the literature (with no limitation by study design applied). Contributing to this review is a panel of relevant stakeholders from the fields of rehabilitation, neurology, biomedical engineering, and BCI engineering who, in providing a summary of available data, comment and make important recommendations to standardize reporting and advance this important and emerging area in robotic-assisted gait rehabilitation in stroke.

The primary question that this review asks is:

what is the current state-of-the-art in neural–exoskeleton interface (non-invasive EEG and/or EMG) during robotic gait training after a stroke?

The secondary review questions asked are the following:

1. What is the profile of the stroke patients in the included studies?

2. What robotic gait devices are deployed?

3. What biosignals are measured in conjunction with the robotic gait devices and what devices (hardware and/or software) are used to capture these biosignals?

4. What protocols are used for recording and processing these biosignals?

5. For what purpose is the acquired biosignal data collected?

As a nascent area, the inclusive approach to study type was taken in this review to allow a true reflection of bioengineering translational research in gait rehabilitation robotics in a clinical population. A compendium of current data collection and signal processing procedures will be developed to allow recommendations for the standardization of future research in this field.

The systematic review was pre-registered with PROSPERO (PROSPERO 2018 CRD42018112252) (Heilinger et al., 2018).

Methods

Results

Discussion

This systematic review compiled the current state of the art in the use of neural biosignals during robotic gait training after a stroke. No studies that used EEG signals to close the loop between human and robotic gait device were identified. Two BCI studies that show promise (with adequate training) were identified for the classification of gait in an exoskeleton after a stroke with a view toward a future BCI application (He et al., 2014; Contreras-Vidal et al., 2018). The work presented by Contreras-Vidal (Contreras-Vidal et al., 2018) builds on a previously published framework proposed by this study group (Contreras-Vidal and Grossman, 2013). Otherwise, as with the majority of the EMG studies identified, the EEG signals were used as an outcome measure to evaluate RGT devices in stroke rehabilitation, for example, as an index of fronto-parietal connectivity to quantify neuroplastic changes (Calabrò et al., 2018). A recent systematic review of BCI rehabilitation in stroke supports this finding, where EEG was used to trigger neuromuscular electrical stimulation in the lower limb but not robotic gait devices to date (Cervera et al., 2018).

Specifying a search strategy that must include individuals with stroke in this systematic review yielded very limited EEG data. While this is informative with respect to the current state of the art in this area in stroke rehabilitation, it does not reflect the broader field of EEG-based control for robotic gait devices well. A recent systematic review by Al-Quraishi et al. (2018) comprehensively reported on EEG-based control for upper and lower limb exoskeletons and prostheses. In this review, 14 studies that used EEG-based control for lower limb movement, primarily in healthy subjects and individuals with spinal cord injury, were identified. Among those, nine studies targeted robotic gait-assistive devices (alone or in conjunction with an avatar), three used motor-imagery-only protocols with event-related desynchronization/resynchronization (ERD) (Do et al., 2013; Gordleeva et al., 2017; Lee et al., 2017), four used a movement-based protocol—with the EEG signal analysis undocumented in one (He et al., 2018b), and in the remaining three as ERD (Garcia-Cossio et al., 2015), combined ERD and movement-related cortical potential (MRCP) (López-Larraz et al., 2016), and exogenous steady-state visually evoked potentials with the visual stimuli representing robotic commands (Kwak et al., 2015). Two studies identified used a combination of motor imagery and movement using sensorimotor rhythms and MRCP (Liu et al., 2017) and event-related spectral pertubations (Donati et al., 2016). Notably, in one patient with a spinal cord injury, EEG signals were used to detect gait initiation to trigger the exoskeleton movement (López-Larraz et al., 2016). In another study with healthy individuals, online control of an overground exoskeleton using ERD in sensorimotor networks to train a classifier to identify two different mental states of walking forward intention or turning were demonstrated. In one body-weight-supported exoskeleton system, the user's intention was classified into active and passive walking phases using 62-channel EEG and power spectrum analysis in 8–30 Hz, normalized to the baseline condition to calculate ERD (Garcia-Cossio et al., 2015). The classification accuracies for active and passive walking with baseline were 94 and 93%, respectively, demonstrating the capability of BCI-assisted robotic training. The majority of EEG-based control in lower limb studies (N = 11; 79%) included in this cited review (Al-Quraishi et al., 2018) were markedly published from 2015 onwards, indicating a relatively new research area and, in part, explaining the poor penetration in the stroke population identified in this current systematic review. Another review of brain–machine interfaces for controlling lower limb powered robotic systems (He et al., 2018a) identifies that the most common studies in this area are classification-based studies of walk vs. stand tasks in healthy subjects and system performance is not clearly presented in these studies. Several challenges were summarized in this review, including EEG denoising, safety, and responsiveness. Furthermore, it concluded that suitable performance metrics and more clinical trials were required to advance research and development in the field.

One study that investigated closed-loop control between human and robotic gait device involving three stroke survivors was identified in this systematic review (Fan and Yin, 2013). EMG activity levels from knee flexor and extensor muscle groups were measured and a multisystem decoding paradigm, which included EMG in addition to joint angle and force production data, allowed the robotic command to be altered. High error rates in the commands generated during active movement were observed when flexion and extension activity alternated rapidly, limiting application in the clinical setting (Fan and Yin, 2013). EMG methods for motor intent identification have previously been noted to have significant limitations in stroke in that they are only appropriate for patients who can produce some voluntary movement or high-enough levels of muscle activity and are not suitable for individuals with severe motor impairment, profound muscle fatigue, or abnormally coactivated muscles (Li et al., 2018). Concerns have also been raised that continuous EMG control may indeed reinforce pathological movement rather than encouraging the recovery of normal motion patterns (Krebs et al., 2003).

EMG has been combined with EEG to improve the recognition of movement intent in the upper limb (Bhagat et al., 2016) in the BCI literature. The current review identified two studies that recorded EEG and EMG. However, the two neural biosignals were not used in conjunction in either study to decode movement. One study reported these measures separately (Calabrò et al., 2018), while the other decoded EMG activity in the lower limbs using EEG during robotic walking in one stroke subject (He et al., 2014).

One example was identified in the literature where the best power transfer between subject and robot was achieved through a fine-tuning procedure for robotic parameters based on optimized EMG activity during the gait cycle (Gandolla et al., 2018); in this context, sEMG could prove to be a useful tool to optimize the patient–robot interaction in the clinical setting. However, the current lack of personalization of robotic gait command derived from neural biosignals and the limited ability to tailor robotic training to participant effort and to rehabilitative goals aligned with motor (re)learning principles limit their capacity as truly restorative devices in stroke rehabilitation. Robotic gait devices and the technological advancements enabling their continued development have been the preserve of the field of engineering (Pons, 2010). Translational research that examines deployment of devices in a clinical population must now also draw from expertise in rehabilitation and clinical research. This paper includes input from experts in the field of neurology, rehabilitation, bioengineering, and BCI engineering, discusses shortcomings in the papers identified, and makes recommendations to advance this field of research. A quick reference guide DESIRED (Table 6) has been developed by the group to identify a minimum reporting data set as a standard for future studies and the rationale and evidence base guiding these recommendations are described in detail next.

The majority of papers identified reported methodologies related to the devices, biosignals, and/or model development as appropriate to the domain of engineering. As a consequence, when considered as clinical studies in a stroke population and assessed using a broadly applicable quality rating tool (Thomas et al., 2004), the majority of studies were deemed to have a weak methodology. Consistent problems identified across studies related to the selection of stroke subjects and to the reporting of the validity and the reliability of the outcome measures employed. Guidelines with quality control checklists are available across a range of clinical study methodologies, for example, RCTs (Campbell et al., 2012), observational studies (Von Elm et al., 2014), and qualitative methodologies (Booth et al., 2014). When introducing participants with stroke or other neurological pathologies to robotic and/or neural signal-based studies, it is recommended that the authors familiarize themselves with the criteria expected based on the study type to be reported in the paper.

It is interesting to note that none of the papers reviewed provided a rationale for their selection of the stroke participants, and limited details on stroke pathology, stroke laterality, and stroke severity levels were documented. The time from stroke, for example, is something that further warrants attention. After focal damage, there is a brief, approximately 3 months, window of heightened plasticity, the so-called opportunity window which, in combination with training protocols, leads to large gains in motor function (Zeiler and Krakauer, 2013). Emerging evidence now supports smaller, plastic, and non-compensatory recovery in the chronic stages after a stroke also (Mrachacz-Kersting et al., 2015; Carvalho et al., 2018). To compare the neural biosignals and their utility in robotic gait training after a stroke across studies and to allow the results to be interpreted correctly, it is imperative to report this information. No consensus was observed across studies with respect to the gait impairment level of those included in the studies and ranged from those fully and independently mobile to those who are wheelchair dependent, again limiting the conclusions that can be drawn across studies. To stratify the findings from future studies, a minimum data set for participants with stroke is recommended and summarized as: stroke type, laterality, time from stroke to inclusion in the study, and functional ambulatory category (Mehrholz et al., 2007). The impairment of sensation also needs to be taken into account, given that accurate motor control can only be exerted with correct sensory and proprioceptive input. An index of stroke severity would also be a useful addition, for example, the National Institutes of Health Stroke Scale score (Ortiz and Sacco, 2014), as well as the level of cognitive function of the participants, if this is not a stated inclusion/exclusion criterion. To best replicate clinical application, it is advised that only the participants with gait impairment are included in the research.

A review of the brain–machine interface for lower limb systems after a stroke, published in 2015, concluded that additional research and development were required to advance this field (Soekadar et al., 2015). This systematic review now identifies that EEG data use during robotic gait after a stroke remains sub-optimal to closing the loop between person and robot. It is acknowledged that EEG activity during walking is not well-understood in general and discordance exists in the literature on the temporal and the spectral patterns of cortical dynamics during walking (Wagner et al., 2012, 2014; Seeber et al., 2014, 2015; Bradford et al., 2015; Bruijn et al., 2015; Bulea et al., 2015; Storzer et al., 2016; Winslow et al., 2016; Artoni et al., 2017; Luu et al., 2017; Oliveira et al., 2017). A number of research groups (Bradford et al., 2015; Bruijn et al., 2015; Bulea et al., 2015; Winslow et al., 2016; Artoni et al., 2017; Luu et al., 2017; Oliveira et al., 2017) report event-related (de)synchronization (ERD/S) (i.e., an event-related power change) in 8–12 Hz (alpha/mu) and 12–28 Hz (beta) oscillations after the onset of stepping/walking task (Bradford et al., 2015; Bruijn et al., 2015; Bulea et al., 2015; Winslow et al., 2016; Artoni et al., 2017; Luu et al., 2017; Oliveira et al., 2017), while other research groups report ERS at 28–40 Hz (low gamma) during early and mid-swing and ERD toward the end of the swing phase and during double support (Wagner et al., 2012, 2014; Seeber et al., 2014, 2015; Storzer et al., 2016). The literature does call attention to the importance of the central pre-motor/motor/sensorimotor and posterior parietal cortical areas in neural signal generation during the walking tasks. Thus, for the separation of walking from rest periods, we recommend the above-described cortical areas and the frequency bands are represented in the data collected and processed.

Decoding the 3D motion trajectory of the lower limbs is a more challenging objective (Georgopoulos et al., 2005). In BCI, this method poses an ideal solution for controlling a robotic device as the applied signal processing algorithm reconstructs the track of the intended movement. To date, most joint trajectory decoding studies have focused on reconstructing the movement of the upper limbs (Bradberry et al., 2010) and fingers (Paek et al., 2014) using 0.5–2 Hz slow cortical potentials (SCP) or band-power time-series of mu and beta bands (Korik et al., 2015, 2016, 2018). Motion trajectory prediction has successfully been applied to lower limb kinematics during treadmill walking in healthy individuals by Presacco et al. (2011, 2012) using SCP time-series. Here topographical analysis did not identify a significant pattern of lower limb movement-related cortical areas. Two studies included in this current review identified the utility of the 0.1–3 Hz frequency band for decoding kinematic data (He et al., 2014; Contreras-Vidal et al., 2018) and EMG kinetics during robotic walking after a stroke (He et al., 2014). One of these studies primarily focused on the frontal, temporal, and parietal brain regions (He et al., 2014), while others removed the peripheral channels susceptible to facial/cranial movement-related noise (Contreras-Vidal et al., 2018). Thus, as decoding the motion of lower limbs during walking is a nascent area, we still recommend a global analysis to identify the most suitable features (potential or band-power time-series, low-delta, mu, or beta frequency band and all cortical areas). However, the SCP/mu and beta band-power time-series extracted from the central motor and posterior parietal areas most likely contain maximal information for decoding lower limb movement trajectories.

From the present review, the lack of a standardized EMG recording protocol when applied to people with stroke-related disability in interaction with exoskeletons is evident. This limitation hampers the constitution of shared database repositories and pooling of data. A protocol and reported methodology should include a minimum dataset of muscles, dimensions, and positioning criteria of the surface EMG electrodes, interelectrode distance, techniques to verify the system selectivity, and technical sampling requirements (e.g., amplifier characteristics, electrode diameter, and impedance limits). The guidelines also suggest the inclusion of additional details relating to signal analysis pipelines, such as filtering and signal quality checks (Blanc and Dimanico, 2010; Merletti and Farina, 2016; Benedetti et al., 2017).

Currently, no consensus exists on targeting specific muscle groups during gait analysis in stroke survivors. The surface EMG of agonist and antagonist lower limb muscle activity during gait is emerging as an effective way of defining motor control during spontaneous movement in stroke (Srivastava et al., 2019). A minimum set of agonist and antagonist muscles in the distal and the proximal compartment of the leg needs to be defined to provide comprehensive muscle recruitment patterns and muscle synergies during the gait cycle after a stroke, which could be helpful for future HMI. Additionally, for individuals after a stroke, sEMG data from both legs (stroke affected and non-affected sides) should be collected as motor deficits are not only associated with the stroke-affected side but also of the non-affected side during spontaneous walking (Parvataneni et al., 2007; Bagnato et al., 2009; Tseng and Morton, 2010; Raja et al., 2012). Recent reviews of muscle synergies in post-stroke gait and robotic gait devices support the need for better standardization of muscles chosen for EMG data capture (Molteni et al., 2018; Van Criekinge et al., 2019). Tibialis anterior, soleus, gastrocnemius, and rectus femoris were noted to be most commonly assessed in all muscle synergy studies after a stroke (Van Criekinge et al., 2019). Considering best clinical practice and the need to record agonist and antagonist muscles during gait, a minimum representative muscle set to be targeted in future studies is recommended as bilateral: tibialis anterior, soleus, rectus femoris, and vastus lateralis, where possible. Alternate muscle/s selection may need to be defined by the participant's stroke-related muscle impairment/s or the robotic gait device and its positioning at specific anatomical landmarks for sensor placement, leading to muscle group exclusion.

Only three studies included in this review referred to a guideline document used for the correct positioning of electrodes on muscles. Failure to do this limits the reliable recording of muscle signals and does not address the challenge of avoiding “crosstalk” (diffused signal components coming from co-active or inactive adjacent muscles) (Basmajian, 1983). Correct sensor positioning aims to minimize this phenomenon and allows the researchers to identify a real co-contraction of agonist and antagonist muscle groups, which is common after a stroke. Basmajian and Blummenstein provide instructions to identify the minimal crosstalk area (MCA) for electrode placement on superficial muscles during gait (Basmajian and Blumenstein, 1980; Basmajian, 1983; Campanini et al., 2007; Blanc and Dimanico, 2010). Although the Surface EMG for Non-invasive Assessment of Muscles (SENIAM) guidelines referenced in the included studies are readily available and easy to use (www.seniam.org; Merletti, 2000), the MCA locations defined by Basmajian and Blummenstein, subsequently updated by Blanc and Dimanico (Blanc and Dimanico, 2010), have been proven to be superior to the SENIAM guidelines (Campanini et al., 2007). MCA identification would now be a minimum standard recommendation to follow in this field when studying the EMG timing during gait (Campanini et al., 2007), with the axis of the electrodes directed parallel to the muscle fibers to increase selectivity (Blanc and Dimanico, 2010). Additional quality assurance measures during robotic gait after a stroke, where feasible with the constraints of the device itself, could include sensor placement checks—performed by eliciting contractions of a single muscle and checking for crosstalk on the other collected traces (Benedetti et al., 2017) and strongly recommended where spasticity or muscle shortening may alter placement accuracy—and data check to ascertain the shape of the power spectral density (PSD) of the signal to ensure meaningful content (Merlo and Campanini, 2010), free of movement artifacts. When recording surface EMG in sessions that include the use of exoskeletons, electrical interference on EMG signals coming from electrical actuators, battery packs, and cables is not unexpected, and the PSD computation could prove to be a powerful tool to unravel such unwanted events.

Limitations

The authors acknowledge that while no language limits were applied when searching across databases, no papers were returned in languages other than English. As such, it is possible that additional manuscripts exist that were not identified through this search strategy. The search also returned studies with heterogeneous use of neural bio-signals, including as an outcome measurement only. While these were included in the review, their purpose was not in line with the primary focus of this review. However, in unifying all studies in this area irrespective of their set purpose, biosignal collection and interpretation in this field could be generalized and commented on constructively.

Conclusion

Overall, while there are ever-growing technological advances in robotics, actuators, and sensors, advances in applications to entrain robotic commands with biosignals for gait training in clinical populations such as stroke have been considerably slower. EEG recording in stroke, where the pathology is at the brain level, has been problematic when compared to other neurological pathologies such as spinal cord injury (Castermans et al., 2014), and similarly EMG recording on the stroke-affected side can be problematic (Sarasola-Sanz et al., 2017; Li et al., 2018). Uncertainty still exists in the literature on the best choice of EEG metric (Goh et al., 2018) and in the ability of EMG to respond accurately in real time (Fan and Yin, 2013). This review, summarizing the current state of the art in neural interface during robotic-assisted gait training after a stroke, identifies a lack of standardization in data collection in this field and provides guidance for study design and reporting future studies. Promising findings for decoding movement during robotic gait after a stroke and potential for EMG, in conjunction with other measurement modes to close the loop, have been elucidated.

Data Availability Statement

All datasets generated for this study are included in the article/supplementary material.

Author Contributions

OL, AD, EG, FM, CG, and DC developed the concept for this systematic review. OL, MT, RO, and CF contributed to the development of the search strategy, review of papers, data extraction, and synthesis sections. All the authors had equal involvement in the drafting and the revising of the review article by unique contribution to the intellectual content and consensus among all the authors on the content.

Conflict of Interest

CG and RO are employed at g.tec Medical Engineering GmbH, Austria. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.