The feasibility study anticipated isolated wrist flexion and extension. The movements were therefore controlled by the Universal Haptic Device (UHD) robot . The robot opposed the movement to exert muscle forces and also measured forces and wrist positions at a sampling rate of 200 Hz with 12-bits resolution (PCI-6023E, National Instruments Inc., USA). Two arrays of 5 × 13 surface high-density EMG (HDEMG) electrodes (diameter of 1 mm, inter-electrode distance of 8 mm, OT Bioelettronica, Italy) were attached to the flexors and extensor muscles of the dominant forearm (Fig. 1). Electrode columns were approximately perpendicular to the muscle fibers, and covered about three-quarters of the forearm circumference. The recorded HDEMG signals were amplified, band-pass filtered (3 dB, 10–900 Hz) and sampled at 2048 Hz, with 12 bits` resolution (USB EMG 2 amplifier, OT Bioelettronica, Italy). For the off-line analysis only the robotic system generated a trigger signal during the wrist movement to synchronize the force signals with the HDEMG signals at both amplifiers. Later on the recorded force, position and HDEMG signals were synchronized and re-sampled for the off line analysis. If any of the electrodes fails to provide the signal, the signal would be eliminated from the analysis and from the real-time feedback.
Two setups were prepared for the feasibility study; a force feedback displayed online as a graphical feedback (Fig. 1 right) and the surface HDEMG biofeedback. The later provided information on flexors/extensors muscles activities in specific regions. The cumulative contribution of the agonist muscles were presented by a blue vertical bar. However, the target line would move higher, if the cumulative activity of antagonist muscles increases. The vertical green bar presented the time of keeping the agonist active above the target threshold and after the 10 s provided a smiley (Fig. 1 left).
A 66-year old male chronic post-stroke subject (height 175 cm, weight 98 kg, hemiplegia, right-side affected) participated in the feasibility study carried out in the controlled laboratory environment of the University Rehabilitation Institute, Republic of Slovenia - Soča.
The inclusion criteria for the participant were: first unilateral stroke, low to medium hand spasticity level (Ashworth scale < 3), cognitive ability to follow the instructions (MMSE > 24). Exclusion criteria: arthritis, swan neck, nerve disfunction and other neuromuscular disorders not a consequence of stroke.
The feasibility study was conducted in accordance with the Declaration of Helsinki, and was approved by the institutional committee for medical ethics. The participant received a detailed explanation of the study and provided written informed consent before the commencement of the study.
3 Force Tracking
The participant's arm and hand was fixed in the frame of the UHD that limited the movement to the wrist flexion and extension only (Fig. 1 right). The robot provided linear resistance defined by the spring-force mechanism  during the wrist movement. We estimated that the applied wrist force had been adequate to the UHD's resisting force also measured in all three dimensions.
We have developed a trapezoid shape trajectory for force tracking in the sagittal plane of the wrist's axis (Fy). The ramp slope of the signal was adequate to the maximal flexion/extension movement speed that the participant could achieve. The participant was seated, put the arm in the UHD frame and was asked to follow the time-course of the reference force signal displayed on the LCD monitor (Fig. 1 right) by:
The trajectory was designed to require actions in the following order: starting from a position of force equilibrium, followed by a maximal wrist extension, returning to the equilibrium position with 2 s of rest, followed by a maximal wrist flexion. The actions were repeated 10 times, lasting approximately 150 s in total. The participant repeated the procedure 4 times with 10 minutes breaks.
The analysis comprised of computing the tracking signal (TS) error, the cumulative forecast error (CFE) and the mean absolute error (MAE). The CFE is calculated as the total sum of errors (e) - bias: differences between the reference signal and measured wrist force;
4 Emg Biofeedback
The setup for the EMG biofeedback was similar to the force tracking; participant's arm and hand was fixed in the frame of the UHD that limited the movement to the wrist flexion and extension only (Fig. 1 left). But the task design required to keep the force at the specified intensity level and activate only flexors muscle group at flexion or only extensor muscle group at extension. The activities of specific muscle groups were assessed with the HDEMG equipment. The surface HDEMG is highly interferential signal consisting of contribution from several motor units (MU). Thus it is difficult for accurate interpretation. The information on muscle activity depends on the distribution of motor units within the muscle tissue and increases variability of muscle excitation across different muscles. Therefore a method for efficient selection of MU was needed . The major challenge was how to reduce the amount of HDEMG data and present it as a real-time biofeedback. Usually a single information is provided as a feedback , but we found that fusion of all rows and 4 columns of electrodes can efficiently eliminate overestimation of muscle activation . The RMS (root mean square) of the signals, recorded by electrodes in columns 1–4 presented the first circle, columns 4–7 the second, 7–10 the third and the columns from 10–13 the last circle on the display (Fig. 2). A single column was common for the neighboring circles. The desired muscle activities (e.g. m.flexors at wrist flexion) were presented by blue color in circles of the upper row, while the circles of the bottom row became red at the undesired activity (e.g. m. extensors at wrist flexion). Thus the participant was able to follow the muscle activities in 8 circles (4 per row) of different color in addition to the cumulative muscle group activity on the target display. Since the wrist flexor or extensor muscles’ activities were related to the same movement, the feedback information was compliant with the recommendations for biofeedback .
The goal of the task was to move the wrist in the specified direction (flexion/extension) and keep the cumulative agonist muscles activity within the target levels for 10 s. The target levels varied according to the agonists vs antagonists cumulative activity ratio. The cumulative activity of antagonists moved the target threshold for agonists cumulative activity higher. A successful attempt was rewarded with a smiley and a small splashing animation on the display (Fig. 2). The initial target levels were set before the daily assessment by measuring three 3 s long maximum voluntary (MVC) contractions of tested muscles during robot controlled wrist flexion and extension.
The feasibility study anticipated three to four sessions with 10 repetitions of wrist flexions and extensions at ~ 40–60% MVC level in stroke survivals in a single day. The protocol lasted 4 consecutive weeks.
Following the feasibility study protocol the HDEMG signals were examined off-line one by one with computer fault detection algorithm. The algorithm was looking for excessive and zero amplitudes in each particular electrode. The zero amplitude signal was considered a faulty electrode and was removed from further computation. Similarly the excessive amplitudes as a consequence of movement artifacts were also removed .
The HDEMG signal of each electrode was rectified after the signal being detrended. Followed by filtering with the 2nd order Butterworth low-pass filter with a cut-off frequency of 20 Hz  forward and backward, the HDEMG data of the daily sessions were normalized by their maximal values. These values were obtained from the MVC recordings on that day (Fig. 2 bellow), eliminating the effect of the electrodes placement and yielding the excitation in % of maximal excitation, making the recordings comparable across weeks.
The objective of the robot supported EMG feedback was to decrease the activity of antagonist muscles and perform the controlled wrist force generation with agonist muscle groups only, if possible. Thus, the RMS values of HDEMG for each repetition, session and week were calculated separately for flexors and extensor muscle groups for wrist flexion and wrist extension. The envelopes and spatially averaged RMS values were also compared between 1st and 4th week for both movement directions. We expected that the participant would be capable of following the EMG biofeedback and keeping the target within the range activating the agonists muscles only. In addition spatially averaged RMS values were statistically analyzed; normal distribution was rejected in all values, therefore Kruskal-Wallis test (unpaired comparison) was used to check the statistical differences of RMS values across weeks. A correction of significant effects was performed by a Bonferroni test with significance level P < 0.05.
Matlab (MathWorks, Natick MA, USA) was used as a main tool for raw data extraction and filtering and Matlab Statistical Toolbox for statistical analysis. All missing data were interpolated and non-number signals of faulted electrodes were immediately eliminated.