Edited by: Timothée Levi, University of Bordeaux 1, France
Reviewed by: Ren Xu, Guger Technologies, Austria; Ning Jiang, University of Waterloo, Canada
*Correspondence: Jose L. Pons
This article was submitted to Neuroprosthetics, a section of the journal Frontiers in Neuroscience
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Upper-limb function recovery after a stroke is in many cases insufficient despite intensive physical therapy. In order to actually get meaningful functional changes in these patients, it has been suggested that traditional physical therapies need to be paralleled with brain modulation interventions aimed to guide plastic changes in the brain (Belda-Lois et al.,
Experimental neuromodulation paradigms using electrophysiological acquisition and stimulation techniques to produce long-term plastic changes at supraspinal and spinal levels have been proposed to treat motor dysfunction in stroke (Lefaucheur,
The electroencephalographic (EEG) activity over the premotor and motor cortical areas presents characteristic variations in the periods before self-initiated movements. Two main motor related EEG patterns are known to reflect mental states related to motor planning and execution processes: the Bereitschaftspotential (BP; Shibasaki and Hallett,
Mostly, studies of low-latency detectors of motor intentions with the upper limb have only been carried out in offline conditions. Moreover, no interventions so far have tested BMI platforms decoding pre-movement BP and ERD patterns online in patients with brain damage due to a stroke. While BP detections online in healthy subjects doing ankle dorsiflexions have demonstrated to be reliable for BMI approaches, BPs in upper-limb movements (Hadsund et al.,
In this study, the ability of a BMI system to assist upper-limb functional movements of stroke patients based on their pre-movement cortical changes decoded online and on a single-trial basis was tested for the first time. Moreover, the study reports results of a multisession intervention (eight sessions during 1 month) using FES to assist upper-limb reaching movements of four chronic stroke patients when EEG-based low-latency estimations of motor intentions are detected. The BMI system performance, patients' functional changes as well as their subjective reports regarding the received intervention are used to discuss on potential benefits of the proposed intervention.
Four chronic stroke patients (age 54 ± 12 years, mean ±
P1 | 54 | Male | Ischem. | L | 3 | 61 | 64 | 2 |
P2 | 54 | Male | Hemorr. | R | 4 | 83 | 66 | 2 |
P3 | 69 | Male | Hemorr. | L | 4 | 65 | 44 | 0 |
P4 | 40 | Male | Hemorr. | L | 5 | 81 | 73 | 2 |
The experimental protocol for this study was approved by the Ethical Committee of the “Universidad Rey Juan Carlos” (Alcorcón, Spain) and warranted to be in accordance with the Declaration of Helsinki. All patients signed a written informed consent.
All experiments were carried out in a sound- and light-attenuated ward of a clinical university. During BMI interventions, patients were seated on comfortable seats and with their arms resting on a desk and movements were performed with the affected upper limbs of the patients.
Patients participated in 10 sessions carried out in different days during 1 month (see Figure
Clinical experts performed functional tests in the first and last sessions to analyse possible patients' improvements.
Upper-extremity sensorimotor function was assessed using Fügl-Meyer Assessment for Upper Extremities (FMA-UE). The four domains evaluated included: upper-extremity motor function (maximum score = 66), sensory function (maximum score = 12), passive joint motion (maximum score = 24), and joint pain (maximum score = 24). Items were scored on a 3-point ordinal scale from 0 (cannot perform) to 2 (performs fully). Summative scores were generated for each domain, scores ranged between 0 and 126 (Duncan et al.,
The Stroke Impact Scale 16 (SIS-16) was used to assess patients' health status following stroke. Duncan et al. (
To analyse kinematics, patients performed five repetitions of the reaching task while being measured with a motion capture platform based on the optoelectronic system Vicon Motion (Oxford Metrics, Oxford, UK). Patients sat on a comfortable chair close to a desk. The patient-to-desk distance was 8-10 cm and the angle of the chair was 90°–100°. In the starting position the patient's trunk rested firmly against the back of the chair. Patients were asked to put their hands on the desk (palms down) with shoulder at around 20° of abduction and elbow at around 95° of flexion. A hard plastic glass (diameter = 5.5 cm, height = 15 cm) was used as target. The glass was placed on the desk in line with the patient's sternum and at a distance equal to 75% of the maximum reachable distance with the paretic arm.
Patients were instructed to reach the glass from the starting position using their paretic hand. All patients practiced the reaching task before motion capture trials. Once this phase was completed, a static calibration recording was performed. Using this recording, it was checked that each marker was visible from the scanning cameras and analyzed movements were registered. In these, after the verbal instruction “Get ready…go,” patients had to lift the arm and reach and grasp the glass at a comfortable speed (similar to the one used in the BMI-FES interventions). Three seconds after reaching the target patients had to move back to the initial position. The time needed to perform the movement was defined as the time interval between the hand movement onset until the hand reached the glass. We analyzed the shoulder, elbow and thorax positions when the hand reached the glass.
We evaluated patients' perceived comfort and acceptability of the BMI-FES platform. Five items are rated on a Likert-type scale from 1 to 5 (strongly disagree—strongly agree): (1) “you are satisfied with the intervention”; (2) “this intervention has been useful in order to carry out activities of daily living”; (3) “you would recommend this intervention to other subject in the same situation”; (4) “The instrumentation is uncomfortable.” The arithmetic mean across all items provides the total satisfaction score.
During the intervention sessions movements of the paretic arm were measured with solid-state gyroscopes, which allowed easy and robust recordings of transitions between resting and movement phases (Ibáñez et al.,
EEG signals were recorded from 31 positions (AFz, F3, F1, Fz, F2, F4, FC3, FC1, FCz, FC2, FC4, C5, C3, C1, Cz, C2, C4, C6, CP3, CP1, CPz, CP2, CP4, P3, P1, Pz, P2, P4, PO3, PO4, and Oz, all according to the International 10–20 system) with active Ag/AgCl electrodes (Acticap, Brain Products GmbH, Germany). The reference was set to the voltage of the earlobe contralateral to the arm moved. AFz was used as ground. The signal was amplified (gUSBamp, g.Tec GmbH, Austria) and sampled at 256 Hz. A standard PC was used to acquire and process the EEG data using a custom-made Simulink model (The Mathworks Inc., Natick MA, USA). This PC sent digital signals to the real-time PC using a USB DAQ (USB-6008, National Instruments, Austin TX, USA).
FES was delivered at the anterior deltoids, triceps and wrist extensors with a multichannel monopolar neurostimulator with charge compensated pulses (UNA Systems, Belgrade, Serbia). Traditional surface electrodes (Pals Platinum—rectangle 5 × 5 cm) were used. The common electrode was located on the oleocranon. Pulse width and frequency were set to 350 μs and 30 Hz, respectively. FES current was adapted in each session with each patient to achieve comfortable stimulation levels that elicited muscle contractions. Current values ranged between 20 and 50 mA (depending on the motor threshold of each muscle and to the patients' acceptance of the received stimulation). Due to the FES configuration and to the weight of the patients' arms, FES alone was not able to lift the arm unless it was successfully triggered by the BMI when patients attempted to perform the reaching task (in which case it provided assistance to the attempted movement). The stimulator was controlled by the PC storing gyroscopic data, which in turn received activation commands from the computer recording the EEG activity via a digital signal. Each time FES was activated it was done in a sequential manner (first deltoids and 250 ms later triceps and wrist extensors) so that the arm could first be lifted from the table and then extended toward the target.
The classifier used to detect movement intentions from EEG was based on the one presented in Ibáñez et al. (
Locations of onsets of voluntary movements were estimated based on the gyroscopic signals. For this, gyroscopic recordings were low-pass filtered (Butterworth, order 3, <10 Hz). For each patient, the sensor that first showed changes during the execution of reaching movements (between the one placed on the forearm and the one on the arm) was used. The peak amplitude performing the movements was estimated in each session. A threshold amplitude for the detection of the onsets of the movements was set to 5% of this peak amplitude. Finally, visual correction of the detected onsets was carried out to ensure that involuntary and residual movements were not taken into account for training and BMI validation purposes.
EEG signals recorded in the pre-intervention calibration trials in each session were used to extract the best features for the posterior online decoder.
For the ERD detection, band-pass filtering (Butterworth, 3th order, 6 Hz < f1, 35 > f2) and small Laplacian filtering were applied to the EEG signal. Power values in segments of 1.5 s and in the frequency range 7–30 Hz (with steps of 1 Hz) were obtained from the frontal, fronto-central, central, centro-parietal, and parietal channels. The Welch's method was used to get power estimations (Hamming windows of 1 s; 50% overlap). The estimations in the training trials from –3 to –0.5 s (with respect to the movement onsets) were labeled as examples of the resting state. Estimations at the movement onsets were labeled as examples of the movement state. The Bhattacharyya distance was used to select the 10 best channel-frequency pairs to build the classifier, i.e., the 10 with the largest distance between the resting state and the movement onset estimations.
For BP, a Butterworth low-pass filter (1 Hz > fc, 1st order) was applied to extract the low-frequency components of EEG signals. A modified version of the large laplacian filter using as reference the average information from eight peripheral channels in the EEG electrode layout was used in order to minimize the weight of individual reference channels (Ibáñez et al.,
A naïve Bayes classifier of independent features was used to detect the ERD pattern preceding the onset of the reaching movements by using 10 channel-frequency pairs previously selected.
A matched filter of length 1.5 s was designed using the selected channel for BP detection. The filter was obtained by removing the baseline level (first 500 ms) of all 1.5 s trials in the training dataset and then averaging the BPs.
To train the logistic regression classifier that combined ERD- and BP-based estimations of motor intentions, training examples of the resting condition were taken from outputs of both classifiers (ERD and BP) between −3 s and −0.5 s, and estimations between ± 125 ms with respect to the onsets of movements were used to model the movement state.
In the online decoding phase, the logistic regression classifier generated estimations of movement intentions every 100 ms. The decoder yielded a binary output depending on whether the probabilistic output from the logistic regression classifier was over or under a certain threshold, which in turn activated the FES. The threshold was obtained from the training dataset and, if needed, it was further adjusted based on the reports of the patient in a few number of pre-intervention calibration trials. The threshold was initialized following the criterion of maximizing the percentage of good trials (GT), i.e., trials with a true positive (TP), and with no false positives (FP). TP were movements detected by the BMI with a detection latency within the range of ±750 ms with respect to the reference onsets estimated with the gyroscopes. EEG-based movement intention detections during resting phases were considered FPs. The precision of the detector was characterized by computing the number of FP per minute (FP/min). The percentage of GT was obtained by counting the amount of trials with no FP and a TP. Finally, latencies of the TP with respect to the onsets of movements were computed to analyse the temporal accuracy of the system. The definition of all these metrics is further elaborated in Ibáñez et al. (
To achieve a stable BMI system, outputs of the BMI were processed by a block ensuring that consecutive FES stimuli were separated by at least 5 s of time.
Once the BMI system had been calibrated, the intervention phase of the session begun. Patients performed 60–80 movements assisted with FES triggered by the BMI. The specific amount of trials performed in each session depended on the patients' arousal and their willingness to continue. Patients were allowed to rest and talk in the middle of the sessions if they needed to. Throughout the intervention trials, patients were asked to concentrate and have the FES activated with their movements.
Each time a trigger from the BMI was received by the PC controlling the FES system, the stimulation pattern described in Section BMI-FES Platform was triggered.
In order to validate the performance of the EEG-based decoder of motor intentions during the interventions, the GT and the detection latencies during the intervention trials were computed. To facilitate the evaluation of the BMI system, patients were instructed not to perform movements when FES stimulation arrived before they have planned to start the movement. The times at which FES stimuli were triggered were compared to the onsets of movements according to the data from gyroscopes.
All statistical analyses were performed using SPSS 17.0 (IBM Corp., New York USA) and Matlab2011 (The Mathworks Inc., Natick MA, USA). Due to the small sample size, Shapiro-Wilk test was applied to check normality of BMI performance and clinical scores. Given that all analyzed samples violated the statistical normality, Wilcoxon signed-rank test was used to compare the clinical scales scores before and after the intervention. The Friedman test for repeated measures was used to compare the BMI performance scores between sessions. Only the sessions with data from all patients were included in the analysis. A linear least square fitting was applied to estimate the tendency of the BMI performance measures along sessions, obtaining the squared error
All results are reported as the mean ±
ERD and BP patterns were used by the BMI platform to control FES assistance. Figure
Average percentages (across sessions) of GT for P1, P2, P3, and P4 were 67.4 ± 15.5, 52.8 ± 6.7, 81.1 ± 12.1, and 66.1 ± 14.8%, respectively (left panel in Figure
Results improved along sessions in three cases (see Table
P1 | 5.54 | 0.460 | 1.21 | 0.480 | 4.50 | 0.388 | −4.99 | 0.016 |
P2 | 1.88 | 0.397 | 0.61 | 0.166 | 1.06 | 0.157 | −4.13 | 0.043 |
P3 | 0.25 | 0.001 | −0.37 | 0.029 | 0.89 | 0.029 | −10.09 | 0.125 |
P4 | 5.64 | 0.506 | −1.66 | 0.434 | 4.80 | 0.564 | −38.79 | 0.636 |
The average detection latencies (considering all sessions) were 202 ± 266 ms (P1), 130 ± 316 ms (P2), 3 ± 190 ms (P3), and 103 ± 254 ms (P4). No pair of sessions differed statistically in the average detection latency [
A tendency toward smaller detection latencies could be observed in all patients (P3, P4, and less in P1, P2) when analysing the evolution along the different sessions (Table
Unsuccessful results of the BMI-FES intervention were only observed in the sixth session with patient P2 (unreliable estimations of motor intentions were generated in that case). As a result of this unreliable BMI function, this intervention session was interrupted since the patient reported an uncomfortable interaction with the FES.
Figure
Table
P1 | 12 | 26 | 9 | 10 | 16 | 16 | 24 | 24 | 61 | 76 | 64 | 74 |
P2 | 31 | 32 | 7 | 8 | 21 | 24 | 24 | 24 | 83 | 88 | 66 | 79 |
P3 | 24 | 22 | 10 | 12 | 7 | 24 | 24 | 24 | 65 | 82 | 44 | 63 |
P4 | 28 | 32 | 9 | 10 | 20 | 24 | 24 | 24 | 81 | 90 | 73 | 73 |
Avg. ± |
72 ± 11 | 84 ± 6 | 62 ± 12 | 72 ± 6 | ||||||||
FMA-UE | 72 ± 11 | 84 ± 6 | ||||||||||
SIS | 62 ± 12 | 72 ± 6 |
Table
Shoulder flexion | 43.8 ± 17.7 | 44.5 ± 18.4 |
Shoulder abduction | 62.5 ± 42.1 | 57.1 ± 36.6 |
Elbow flexion | 85.1 ± 16.9 | 83.0 ± 20.2 |
Thorax flexion | 12.1 ± 4.3 | 7.1 ± 4.1 |
The perceived comfort and acceptability of the intervention platform proposed here varied across patients. Three patients (P1, P2, P3) “agreed or strongly agreed” with the intervention, while P4 “neither agreed nor disagreed” with it. One patient (P2) reported that the intervention was useful in order to carry out activities of daily living (“strong agreement”), but the other three patients reported “strong disagreement” (P3, P4) or “neither agreement nor disagreement” (P1) in this regard. Regarding the degree of recommendation of the received intervention; all participants declared to “agree or strongly agree.” Finally, two participants (P1 and P4) reported to be in “agreement or strong agreement” with the instrumentation process carried out during the intervention; however, the other two participants indicated to be in “disagreement or strong disagreement.”
This study has tested for the first time the ability of a BMI system to assist upper-limb functional movements of stroke patients based on their pre-movement cortical changes decoded online and on a single-trial basis. Such use of anticipatory EEG activity allows the timely assistance of patients during the motor tasks performed. The study aimed to evaluate the usability of the proposed technology and its potential effects when applied in a prolonged in time intervention. Overall, patients could reliably control the interface by spontaneously performing movements and low average detection latencies (<200 ms) were obtained. Moreover, measured FMA-UE changes were higher than the minimal clinically important difference.
Most BMI interventions involving patients with a stroke have used synchronous paradigms, that is, patients performed movements paced by an external signal, and ERD patterns were used to characterize the movement phases (Ang et al.,
Importantly, although a small sample is considered here, results are comparable (and better in some cases) to those obtained in previous similar studies with healthy subjects (Ibáñez et al.,
Importantly, the potential impact of motor-related artifacts in the obtained BMI performances, although possible, is estimated to be small. Muscle artifacts, on the one hand, lead to increases of the cortical activity in frequencies within the beta band, which are in the opposed direction to ERD changes learned from the pre-movement EEG signals in the training stage. On the other hand, post-movement low frequency components in the EEG are typically spread along the scalp (which allows spatial filtering techniques to cancel them) and also they typically start with positive changes of the EEG amplitudes (Shibasaki et al.,
Patients' ERD and BP were marked and started around 0.5–1 s before movements could be observed from gyroscopes. Both patterns showed alterations compared to ERD and BP patterns in healthy subjects in terms of spatial and temporal distributions and in line with previous findings (Serrien et al.,
Changes in patients as a result of the intervention were observed in terms of changes in the BMI performance across sessions. GT in patients P1, P2, and P4 improved in the first 3–4 sessions, and remained high in P3. In addition, average detection latencies decreased with sessions in P3, P4, and less consistently in P2. These results indicate that patients were able to modulate preparatory cortical activity that released the movements, and therefore reinforces the idea that asynchronous BMI approaches as the one here are suited to reinforce and maximize motor planning in stroke.
Apart from changes in BMI performance, no other neurophysiological changes are described here. Using ERD and BP phenomena simultaneously gives raise to the hypothesis that the intervention might induce changes in both motor planning and execution concurrently, contrasting the existing approaches that singled only one of the former phenomena out. However, despite changes in EEG due to the intervention period were explored, no consistent results were found in terms of ERD, BP, or power in motor cortical rhythms during rest (results not included here). Such changes, if they exist, should be derived from a larger sample of patients, given the high intra-patient variability in EEG information across days (Shenoy et al.,
Regarding functional changes, the analysis of reaching kinematics led to small but consistent results. Reaching movements need an adequate range of motion toward thorax extension, shoulder flexion, and elbow extension. The altered kinematic of reaching movements in stroke are typically characterized by compensatory trunk and shoulder movements (Roby-Brami et al.,
With respect to the satisfaction assessments, according to the patients two aspects should be improved in order to carry out this intervention in further studies. First, it is necessary to carry out more sessions and more arm movements to achieve a better learning transfer. Second, the instrumentation has to be simplified because this aspect may generate fatigue and discomfort in the participants.
This study was carried out with a small sample of patients. This implies that, despite consistent changes could be observed across patients, no statistically significant changes were observed in the clinical or functional metrics, likely due to the small statistical power. To demonstrate the benefits of the intervention proposed here as compared to (or in conjunction with) more traditional therapies, future studies in line with the present manuscript should involve larger groups of patients, with sample-matched control groups and blind assessments to avoid observer's bias. Moreover, the analysis of the proposed intervention in subacute stroke would also be relevant to test if the functional impact is larger in patients more susceptible to neuromodulation interventions.
This manuscript represents an approach to BMI-FES interventions for the upper limb in stroke patients, exploiting the predictive properties of EEG signals related to motor processes. Results show a potentially beneficial effect of the BMI-FES intervention in terms of clinical scales and kinematic analysis. In addition, the study demonstrates the suitability of the proposed EEG-based decoding algorithms for their use with patients.
JI, FM, EM, and JS participated in the design of the experiment, in the data collection process, in the data analysis, and in writing the manuscript. MD and JP designed the experiments, carried out the experimentation, and wrote the manuscript. RC, AC, MC, IA, and JM collected the patients, carried out the experiments, and edited the manuscript.
The 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.
This research has been supported by Spanish Ministry of Science project HYPER-CSD2009-00067. We acknowledge support of the publication fee by the CSIC Open Access Publication Support Initiative through its Unit of Information Resources for Research (URICI).