Determining optimal mobile neurofeedback methods for motor neurorehabilitation in children and adults with non-progressive neurological disorders: a scoping review

A medical librarian at the National Institutes of Health was consulted to develop the optimal search strategy to address our research question. A title and abstract keyword search was conducted utilizing the following search terms and general strategy adapted as needed for the PubMed, Web of Science, and Scopus databases: “motor” AND “Brain Computer Interface” OR “BCI” OR Neurofeedback” OR “BMI” OR “EEG biofeedback”. Only articles published in the English language were considered. There was no restriction on the date of publication with April 15, 2021 as the final search date.

All clinical studies on the application of non-invasive, mobile (i.e. EEG or fNIRS) BCI-NFT for motor neurorehabilitation of individuals (children or adults) with non-progressive neurological injuries (e.g., stroke or CP) were included. We excluded studies that only enrolled healthy participants or those with progressive neurological conditions such as Parkinson’s Disease or Amyotrophic Lateral Sclerosis. Studies using Magnetic Resonance Imaging (MRI) or magnetoencephalography (MEG) to deliver neurofeedback were excluded. Systematic or scoping reviews were not included; however, reference lists of relevant reviews were scanned for studies that may not have been captured in the initial search. Further criteria for inclusion were that the interventions had to utilize a feature of the participants’ cortical activity within the training session, and that participants had to be attempting to perform a voluntary motor task. Studies in which participants were using motor imagery or action observation to generate the brain activation signals used for neurofeedback were excluded. Finally, since the goal was to examine how methodological differences might affect motor outcomes, only studies that reported these measures were included.

Duplicates were initially eliminated within ENDNOTE. Titles and abstracts were screened independently in EndNote by two authors (AB, DD) to remove additional duplicates and to identify studies that potentially met the inclusion criteria. Disagreements on which articles to retain were resolved through discussion. Full texts of all potentially eligible papers were independently assessed by the same two review authors, with disagreements again resolved through discussion. Reference lists of the final set of papers, as well as of relevant systematic reviews, were also scanned to ensure that no studies were missed. Then, three authors (AB, VH, WL) extracted data independently from the final group of studies satisfying all inclusion and exclusion criteria, with each assigned a group of articles to extract data from using a comprehensive pre-piloted data extraction form, in Google Sheets format, and a group to verify data extracted by another author.

Extracted information included: study population and participant demographics; number of participants in intervention and control conditions, motor task performed during BCI-NFT and whether it was an upper limb or lower limb task, control condition, dosage of BCI-NFT, cortical activity feature(s) extracted to generate the feedback, cortical region(s) the feature was selected from, signal processing technique used for extracting the feature(s) and generating feedback, feedback timing, type of feedback (e.g., visual, robotic, functional electrical stimulation), whether any other additional training (e. g., conventional physical therapy) was provided before and/or after BCI-NFT , and finally, all reported motor outcome measures.

Where possible, mean differences in motor outcome measures across specific feature categories (e.g. feedback types) were analyzed using a General Linear Model (GLM) or independent t-tests. Pearson correlation procedures were also used to relate specific training aspects or features to motor outcomes (p<0.05 for all analyses).

Table 1 provides a summary of details extracted from all included studies. A total of 223 individuals participated across studies; 153 of whom were in the BCI-NFT condition. All participants were adults, with 153 males and 70 females. Except for one study, on 12 participants with incomplete spinal cord injury, seven of whom were in the experimental group [64], the rest were on the stroke population. Nineteen studies only included participants more than 6 months post-stroke, and three studies, all RCTs, only focused on participants within the first 6 months of stroke; with 24 [26], 14 [23], and 20 participants [63]. Remsik et al., included four participants less than 6 months and 17 more than 6 months post-stroke [41]. Thus, in all there were 149 participants more than 6 months post stroke and 62 participants less than 6 months post stroke with 113 and 33 of those, respectively, participating in the BCI-NFT.

Thirteen studies reported the mean FMA difference in the experimental group, pre to post intervention, ranging from 0.77 to 17.0 across studies, six of which showed significant improvements (see Table 3). The mean change across the 13 studies was 6.53 ± 4.46. Four studies additionally reported the FMA pre-post change score difference between the control and experimental groups [4, 22, 23, 26], two of which showed significantly greater improvements in the neurofeedback group [22, 26]. Mean change across the three studies was 4.06 ± 0.57. The mean Action Research Arm Test (ARAT) change score from pre to post intervention within the experimental group, reported in six studies [41, 65‐67, 73, 74], ranged from 1.3 to 23.8 with the ARAT improvement significant in three [65‐67]. Mean change across the six studies was 8.34 ± 9.00.

Several other motor outcomes were reported less frequently across studies. Grip strength (GS) was reported in five studies. Mean GS improvement, pre to post intervention, ranged from 3.87 kg [41] to 9.83 kg [67], both of which were statistically significant, p = 0.046 and p < 0.005, respectively. Mean GS post intervention increases were not significant in the other three studies [65, 66, 72]. Muscle spasticity was assessed in five studies using the Modified Ashworth Scale (MAS); however, no statistically significant changes were reported. Active dorsiflexion range of motion (ROM) improved in two studies [70, 77]: in McCrimmon et al., 5 of 9 participants showed an increase of \(2.5^\circ\) or greater with a significant positive linear trend from pre to post intervention for the group as a whole (p < 0.01), and in Takahashi et al., the mean change of \(8.7^\circ\) was also significant (p < 0.001) [77]. Mean active wrist extension ROM was improved by \(16.8^\circ\) for the BCI-FES group versus \(3.4^\circ\) for the FES group in the study by Osuagwu et al.; however, no statistical comparison was reported [64].

Mean Stroke Impact Scale (SIS) improvements of 5.4 [41], 10.5 [68] and 21.3 [71] were reported in three studies, none of which were significant. Despite no significant changes immediately post intervention, Remsik et al. [41] showed a significant SIS improvement of 6.2 (p = 0.05) at follow-up. A small, statistically significant improvement in gait speed of 0.08 m/s (p = 0.007) was reported only in the neurofeedback group by Mrachacz-Kersting et al. (2016), as measured by the 10 Meter Walk Test (10mWT) after only 20 min (30 pairs of MA and associative feedback) of neurofeedback training [61]. McCrimmon et al. in contrast showed no significant improvement in walking speed in the neurofeedback group [70] after 12 training sessions; the linear trend from pre to post intervention, however, was significantly positive. Mrachacz-Kersting et al. (2019) found significant 10mWT improvements in the control and neurofeedback groups (both p < 0.008) with no statistical between group difference [26]. Further, they reported that five participants in the experimental group and three in the control group who could not walk pre-intervention were able to walk after 12 training sessions [26]. Significant 10mWT improvements were also found in the control group with no between group difference. All other motor outcomes reported in each study are listed in Table 1. The trends for these outcomes were mostly positive or unchanged except for reaction time, which worsened significantly after training [59, 60].

Independent t-tests were conducted to assess the effect of feedback type (robotic: n = 3, and FES: n = 9) and co-interventions (yes or no) on FMA differences within experimental groups in the 13 studies which reported these data. Despite higher mean FMA values for FES compared to robotic feedback (7.6 ± 4.8 vs 5.3 ± 2.8; p = 0.47), outcomes were not significantly better. Conversely, the ARAT mean difference was higher for robotic (n = 3) vs FES feedback (n = 3) (11.59 ± 10.53 vs 5.10 ± 7.34), but was also not significant. The inclusion of co-interventions showed no consistent or statistically significant effect on FMA and ARAT scores (for FMA: none=6.4 ± 5.6 [n = 9], yes = 6.7 ± 2.04 [n=4]; p = 0.92, for ARAT: none=5.16 ± 6.32 [n = 4], yes = 14.7 ± 12.8 [n=2]; p = 0.47). Mean FMA change also did not differ significantly by the level of evidence as assessed with a general linear model (II = 6.8 ±2.4 (n = 4), III = 0.77 (n = 1), IV = 4.8 ± 4.6 (n = 4), V = 9.5 ± 5.1 (n = 4); p = 0.27).

The effect of differences between classic versus associative learning paradigms on FMA within the experimental group was also assessed using an independent t-test. The two studies that used associative learning [26, 61] had a slightly lower mean FMA score when compared with studies that used operant conditioning (n=8) [4, 22, 23, 65, 68, 70, 71, 75, 76] of 4.6 vs 6.9, respectively. The difference, however, was not significant (p = 0.54).

Pearson correlation between FMA improvement after BCI-NFT and the number of training sessions showed a moderate positive relationship (r = 0.67, p = 0.01). Despite a similar correlation value between ARAT score and the number of sessions; this was not statistically significant (r = 0.70, p = 0.20).

The small number of studies with the same functional outcomes included in this review limited the ability to identify specific protocols or features with superior efficacy or effectiveness. Many paradigms aimed to produce reliable (minimal false positives or negatives) but varied in the timing of activation of an external assistive device with movement onset, ideally recommended to occur within a 300 ms window [98]. Therefore, it seems reasonable that paradigms that use EEG activity prior to movement to predict movement intention rather than real-time detection of movement onset would be preferable, if not essential, considering the time delays related to EEG processing and communication between system software and hardware. MRCP contain signals that precede movement; however, the time between these and movement onset or device activation can fluctuate within and across individuals, and perhaps are even more variable for those with brain injuries. To account for this, machine leaning algorithms, such as the Gaussian classifier utilized by Biasiucci et al. [22], and Naïve Bayes deployed by Ibanez et al. [68] that can predict or classify the motor attempt online, might be a more precise alternative than relying on the consistency of the MRCP signals such as peak negativity timing with respect to motor onset. These could reduce the calibration time and thereby optimize therapy time, and account more effectively for individual variability, This could perhaps be improved even further using transfer learning algorithms to train a predictive model once and then transfer this across participants, thus eliminating the calibration phase [99]. Proprioceptive feedback was used almost exclusively for these motor rehabilitation applications, rather than visual feedback alone, which is logical since these also serve assistive as well as sensory-enhancing roles. More data on the short- and longer-term efficacy of FES compared to robotic feedback are still needed. Modulation and progression or weaning of feedback over time are important future considerations to maximize motor learning and neuroplasticity. While we did not compare effectiveness of motor imagery to motor attempt, the latter is more intuitive and clearly more feasible across a broader range of patients.

Some limitations are that this scoping review did not, by design, include all studies on the use of BCI-NFT, but it does provide a comprehensive review on the current state of the science on motor attempt paradigms for neurorehabilitation. Another possible limitation is that we did not restrict the studies by the level of evidence; however, significant mean differences in outcomes across levels were not found. The limited use of consistent outcomes across studies also restricts the number of studies included in any meta-analysis, reinforcing the need for greater efforts in rehabilitation research to enable the accumulation of larger datasets with common data definitions and outcomes. Still, given the limitations, several studies demonstrated clinically significant functional changes after short durations of training, far shorter than typically needed for producing the same magnitude of effects with motor training alone.

In conclusion, the specific focus on enhancing neuroplasticity within a task-specific paradigm with BCI-NFT provides a solid neurophysiological mechanism for potential behavioral changes that we believe are only beginning to be realized. Future efforts should be directed towards deploying these in younger patient populations with greater neuroplastic potential and designing these systems for broad clinical implementation and for larger efficacy trials that compare these to other forms of neuromodulation or other evidence-based motor training approaches at equivalent doses.