Rehabilitation of hand in subacute tetraplegic patients based on brain computer interface and functional electrical stimulation: a randomised pilot study

A standard cue-based paradigm was implemented with rtsBCI, a part of the open source Biosig toolbox (Vidaurre et al 2011), implemented under Simulink, MATLAB. Patients sited in their wheelchairs approximately 1.5 m from a computer screen. A trial started at t = −3 s and ended at t = 3 s. At t = −1 s a warning cue (a cross) was presented at the screen, followed by an execution cue (an arrow) at t = 0 s. The warning cue stayed on the screen till t = 3 s, and thereafter the screen stayed blank for a random period which was between 1 and 3 s before the next trial began. The total time between two trials was random, between 7 and 9 s. There were two types of arrows i.e. execution cues, an arrow pointing to the right for MA of the right hand and to the left for the MA of the left hand. Patients were instructed to attempt waving their hand continuously from t = 0 s till t = 3 s, i.e. while they saw a cross on the screen. Note that unlike able-bodied persons, paralysed people can differentiate between the imagination of movement, a movement attempt, and an overt (executed) movement. We considered MA being more appropriate task than the imagination of movement, because the aim of the study was the restoration of voluntary hand movement. There were 120 trials (60 for the right and 60 for the left hand) divided in 4 runs each consisting of 30 trials (15 for each hand).

During this task patients' multichannel EEG was measured with 48 electrodes placed according to 10/10 system (Jurcak et al 2007) using usbamp device (Guger technologies, Austria). Electrodes covered the central region of the sensory-motor cortex, the parietal cortex and sparsely covered the frontal and occipito-temporal cortices. Forty seven electrodes were used to record EEG while one electrode was placed at the lateral cantus of the orbicularis oculi of the right eye to record electrooculogram (EOG). EEG was recorded with the sampling frequency of 256 samples s^–1 with respect to the linked-ear reference. Impedance was initially set and later kept under 5 kΩ, occasionally checking its values on a graphical user interface of a usbamp proprietary Simulink block in between experimental sub-sessions. A ground electrode was placed at the electrode location AFz. EEG signal was filtered on-line between 0.5 and 60 Hz and was notch filtered at 50 Hz using the IIR digital Butterworth filter built into a modular amplifier.

Continuous data were split into 6 s long trials, starting at t = −3 s and ending at t = 3 s, with respect to an execution cue. Datasets of each patient were decomposed into independent components (IC) (Hyvarinen and Oja 2000) using 'Infomax' algorithm implemented in EEGlab (Delorme and Makeig 2004) under Matlab. IC transformation provides IC components and a transformation matrix. The components were visually inspected and components corresponding to instrumental and biological interferences (line noise, EOG, EMG and ECG) were removed by setting their value to 0, through EEGLab graphical user interface. On average 5–10 components were removed. Following this, signal was back projected from IC into EEG domain by multiplying new IC matrix with the inverse of the transformation matrix. A common average reference was computed for all channels.

Time-frequency analysis was performed in EEGlab based on event-related spectral perturbation (Makeig 1993), the extension of event-related synchronisation/desynchronisation (ERD/ERS) analysis (Pfurtscheller and Lopes da Silva 1999). The baseline period was selected from t = −2 s till t = −1 s. The Morlet wavelet transform was used to perform a time frequency analysis in 3–60 Hz, with a Hanning-tapered window applied; the number of cycles was set to 3 at the lowest frequency.

An average ERD/ERS across patients was calculated in EEGlab. This was used to create the average ERD/ERS scalp maps for a chosen frequency band and time window. A statistical non-parametric method with Holm's correction for multiple comparisons (Holms 1979) was used to test for statistically significant differences between ERD/ERS scalp maps before and after treatments of patients within a group, with a significance level set to p = 0.05.

A SSEP is the response of the central nervous system to an electrical stimulation (Gugino and Chabot 1990, Cruccu et al 2008). The SSEP may infer motor functions on the assumption that an injury severe enough to damage the sensory pathways may also affect the motor pathways. The purpose of SSEP analysis was to detect the latency and the amplitude of the N20 peak that occurs around 20 ms following the electrical stimulation of the upper limbs. The peak has a highly repeatable latency in able-bodied people. The increased delay of N20 is an indicator of the damage of the neural pathways, resulting in axon demyelination. The demyelination in turn causes the reduced propagation velocity of action potentials along the axon, manifested as the increased latency of N20 peak. In more severe cases, when some nerve fibres are severely damaged, the amplitude of N20 is reduced or completely absent (Curt and Dietz 1999). The recovery of the neural pathways, followed by re-myelination, may result in the re-appearance of N20 and in a reduced N20 latency.

In the current study SSEP was measured for the left and right median and ulnar nerves, as these two nerves share innervation of the wrist and fingers. Electrodes were attached on the surface of the skin above the corresponding nerves at the wrist. The nerves were stimulated one at the time using a short pulse electrical stimulation (Model DS7, Digitimer, UK). A stimulation intensity was set so that a small visible twitch could be observed at the thumb for the median nerve and at the little finger for the ulnar nerve. For each nerve, electrical stimulation was delivered 250 times with a frequency of 3 Hz. SSEP of the right hand median and ulnar nerve was measured at electrode location CP3 and of the left hand nerves at the electrode location CP4. EEG was recorded with usbamp, with a sample rate 4800 Hz. The EEG signal was band pass filtered between 2 and 2000 Hz and notch filtered at 50 Hz; individual responses were averaged with respect to the onset of stimulation.

In patients with incomplete tetraplegia who have partially preserved control of movement, the ROM is reduced, as compared to the able-bodied people. The ROM of the right and the left hand wrist, during extension and flexion was measured using a Zebris system (Zebris Medical GmbH, Germany) which measures the travel time of ultrasonic pulses. The pulses are emitted by three stationary transmitters and are recorded by small markers which are ultrasound microphones attached to the hand. The Zebris system markers were placed on the following bony landmarks on the subject's hand: the radius (marker 1), carpometacarpal joint (marker 2) and the carpometacarpal bone (of the index finger or the thumb, marker 3). The ROM was calculated as an angle between the intersecting imaginary lines formed between markers 2–3 and markers 1–2.

The original test has 6 grades, ranging from 0 to 5. The score of 0 equals no contractions in the muscle, 5 equals the ability to hold test position against strong pressure. We used the extended version with subgrades: 0, −1, 1, 1+, −2...5 (Freze et al 1987 adapted) which were converted into integer numbers 0, 1.2, 3....14 for the purpose of a statistical analysis. Due to the nature of the injury, patients initially had better preserved the voluntary control of muscles in the shoulders and upper arms (MMT = 3 to 4, i.e. (3) can move against gravity and (4) can also withstand moderate resistance) than of the forearm, wrist and hand. The initial MMT of most of the forearm/hand muscles was between 0 and 2 (MMT = 2 moves through the ROM through a horizontal plane, i.e. cannot resist gravity). Muscles included in the analysis were muscles tested as the part of a regular clinical practice at the hospital. From the reason of low functionality of hands, we did not test the activities of daily living, as they would be typically tested in e.g. stroke patients. While each muscle has been tested individually, because of the small number of participants, we grouped muscles according to their function, thereby getting larger numbers for statistical analysis. We tested muscles controlling the shoulder, upper arm, forearm/wrist, hand extensor and hand flexor muscles, although we expected that treatment would primarily influence wrist and hand/finger muscles. Muscles were grouped as follows: muscles controlling shoulder (latissimus dorsi, pectoralis major, serratus anterior and deltoid), upper arm muscles (triceps, biceps), lower arm muscles controlling flexion, supination and pronation (supinator, pronator, brachioradialis), extensor muscles of fingers and wrist (extensor digitorum communis, extensor carpi radialis brevis and longus), flexor muscles of fingers and wrist (flexor carpi radialis and flexor digitorum profundis). Nonparametric statistical analyses were performed due to the relatively small number of samples (minimum 10 per group, e.g. 2 muscles in the group of 5 patients). To compare the results of MMT test within a group, a paired Wilcoxon sign rank test was used, while to compare between groups a Wilcoxon rank sum test was applied. Following this, a Holm–Bonifferoni correction for the multiple comparisons was applied. All numerical procedures were performed in Matlab (Mathworks, USA).

At the start of a therapy session, a quick off-line EEG recording was obtained. The EEG recording consisted of 20 trials of MA for each hand and followed the experimental protocol described in 2.2.1. BCI features were based on time-domain parameters (TDP) (Vidaurre et al 2009). The recorded data was separated for each hand and the data for each hand was further split into baseline (from t = −3 to t = 0 s) and MA (from t = 0 to t = 3 s). The aim was to use the data to classify between the MA and the baseline, for each hand separately. Performing such classification should provide a reasonable initial classification accuracy for the on-line session and reduce the need to record large training data. We previously tested the algorithm on able-bodied people (Osuagwu and Vuckovic 2014), showing that a small training data set of about 20 trials results in an initial classification accuracy between 75% and 100%. Computed classifier parameters were then further updated and refined during an on-line BCI session, as described later in section 2.3.2. in order to improve the online performance.

The time domain features of the separated data were calculated for 7–30 Hz EEG frequency band using equation (1)

$$\begin{eqnarray}{\rm{T}}{\rm{D}}{\rm{P}}=\langle {\rm{var}}\left(\displaystyle \frac{{\rm{d}}X{(t)}^{j}}{{\rm{d}}{t}^{j}}\right)\rangle & j=0,\ldots p,\end{eqnarray} \tag{ 1 }$$

where X(t) is a wide-band EEG, t is a current sample, j is a derivative (p = 9), 'var' is a variance operator and 〈.〉 represents a smoothing/averaging operator. The variance operator in this equation acts as a band-power operator since the variance of filtered centred (mean = 0) signal is equal to the band power (Vidaurre et al 2009). The BCI setup showing the computation of TDP is shown in figure 1. The squaring and smoothing, performed over 1 s is the part of a band-power calculation. Following the band-power calculation, the logarithmic transformation of TDP parameters was performed to enforce a normal distributions needed for classifier based on linear discriminant analysis (LDA) (Fukunaga 1990). LDA is a technique used to project data onto a low dimensional space to enable the separation of the data into classes. The classes are separated using a hyperplane that maximizes data separability. LDA can be expressed as

$$\begin{eqnarray}&&h(f)={A}^{T}\cdot {f}_{c}(M,q),\end{eqnarray} \tag{ 2 }$$

where f_c(M, q) denotes an M dimensional feature matrix with q observations of class c (c = 1, 2), A is a transformation and h(f) is a linear discrimination function. The expected value and the measure of scatter for each class in the projection space are obtained from the corresponding values in the original space μ_c and Σ_c using transformations

$$\begin{eqnarray}&&{\tilde{\mu }}_{c}={A}^{T}\cdot {\mu }_{c},\end{eqnarray} \tag{ 3 }$$

$$\begin{eqnarray}&&{\tilde{{\rm{\Sigma }}}}_{c}={A}^{T}\cdot {{\rm{\Sigma }}}_{c},\end{eqnarray} \tag{ 4 }$$

The best line on which to project f is determined by maximising the Fisher's criteria

$$\begin{eqnarray}&&J(A)=\displaystyle \frac{{| {\tilde{\mu }}_{1}-{\tilde{\mu }}_{2}| }^{2}}{{\tilde{{\rm{\Sigma }}}}_{1}+{\tilde{{\rm{\Sigma }}}}_{2}},\end{eqnarray} \tag{ 5 }$$

where the numerator is a between class scatter and the denominator is a within class scatter.

Zoom In Zoom Out Reset image size

The class of the output h depends on which side of the separation plane it lies. The TDP features were used to compute an initial LDA classifier, which classified between a MA and a baseline for each hand. The initial classifiers were stored and later used as initial values for an online classifier.

During a therapy session, a BCI was used on-line with classifiers to discriminate between a hand movement and no movement. To improve the performance of BCI on-line LDA classifier, the mean values of both classes and a within class covariance matrix (equations (3) and (4)) were updated on-line. This was necessary due to the small number of off-line trials. Short off-line training was needed due to a limited availability patients had for the study. Typically patients had 1 h for BCI setup and for training of both hands.

During therapy sessions, patients' EEG was recorded from three pairs of bipolar electrodes, CP3–CF3, CPz–CFz and CP4–CF4, located over the sensory-motor cortex. Bipolar configuration is convenient for supressing the common sources of interferences recorded over both electrodes, such as line noise, EOG or EMG interference. EEG was recorded with usbamp, with a sampling frequency 256 Hz, band-filtered online between 0.5 and 30 Hz (5th order Butterworth filter). The ground electrode was attached to the left ear, monopolar reference to the right ear and the impedance was kept under 5 kΩ.

A difficulty or the threshold of activation of a BCI could be adjusted, so the difficulty could be e.g. increased to reduce a false positive rate or decreased to make the activation easier for a patient who is tired or has a low concentration. The difficulty was adjusted by setting the length of the buffer, i.e. a time sequence containing a classifier output in which a desired class (a left or right hand MA) had to be successfully detected. A classifier made a decision based on the EEG sequence of length b (typically b = 1.5–2 s, while the maximum allowed length was B = 3 s or 768 samples). However, the classifier could make a decision based on a portion period called f. So if a total sequence for a particular training day is b = 2 s, with maximum sequence B = 3 s and f = 75% then difficulty d is 50% (equation (6)). Note that this only affected the length of the buffer containing the outputs of a classifier, and did not in any way affect classifier features or a single outputs of a LDA classifier.

$$\begin{eqnarray}&&d=\tfrac{b\cdot f}{B}=\tfrac{2\,{\rm{s}}\cdot 0.75}{3\,{\rm{s}}}=\mathrm{0.5.}\end{eqnarray} \tag{ 6 }$$

On each therapy session a patient performed 30–40 MA of each hand, separated in sub-sessions, consisting of 20 trials. Each successfully detected movement attempt resulted in the activation of a FES, as described later in the text.

During therapy sessions, a patient sit in front of a computer screen, attempting a movement upon the appearance of a visual cue. The MA was facilitated by a feedback in the form of a gauge (figure 1). Patients were told that during a MA, when a gauge indicator reaches 0, there will be the activation of the set of electrodes attached to their hand muscles in a predefined order. After each trial sub-session, patients got a numerical score (out of 20 trials) on the screen about their performances.

Four bipolar electrodes were attached over the wrist and hand/thumb extensor and flexor muscles. This setup enabled patients to perform a grasp by opening and closing their hand. The electrodes were attached to sequentially stimulate the extensor digitorum, extensor pollicis longus (extensor muscles), flexor digitorum superficialis and flexor policis brevis (flexor muscles). The stimulation of the first two muscles resulted in opening of the hand and four fingers (index finger to pinkie), followed by a thumb abduction; a subsequent stimulation of two flexor muscles resulted in closing of the hand. The whole stimulation sequence, including the opening and closing of a hand, lasted 10 s. The same setup was used for both patient groups. The main difference was that BCI-FES group had to activate FES by attempting to open and close the hand and for FES group stimulator was activated automatically repetitively with 10 s on and 10 s off. A temporal sequence for one MA attempt in BCI-FES group is shown in figure 2.

Zoom In Zoom Out Reset image size

FES was delivered using a multichannel FES device (Rehastim, Hasomed, Germany). The frequency of stimulation was 26 Hz, the pulse width was 200 μs and the current amplitude varied between 15 and 35 mA and was individually chosen for each patient to produce a visible muscle contraction without discomfort.