Participants
Patients fulfilling the inclusion/exclusion criteria (see below) were first approached by an occupational therapist at the rehabilitation units of Hospital Esperança or Hospital Vall d’Hebron from Barcelona to determine their interest in participating in a research project on VR-based motor rehabilitation using the Rehabilitation Gaming System. Inclusion criteria were as follows: a) Upper limb hemiparesis secondary to a first-ever ischaemic stroke (MCA territory) or hemorrhagic stroke (intracerebral); b) Proximal upper limb motor deficit (Medical Research Council Scale for proximal muscle strength (MRC) >3); and c) Capable and willing to participate in the RGS therapy (Mini-mental state examination (MMSE) >22). The ethics committee of clinical research of the Parc de Salut Mar and Vall d’Hebron Research Institute approved experimental guidelines. In total 26 hemiparetic stroke patients were identified as potentially eligible participants. We excluded a further six participants showing a previous history of upper-limb motor disability and major cognitive deficits and seizures. The remaining 20 participants (n = 20, age = 62.2 ± 14.3, 11 males, 9 right-sided hemiparesis, MRC = 3.7 ± 0.5, MMSE = 26.8 ± 2.8, 17 ischaemic, 211 ± 390.9 days post stroke) were informed about the aim and procedures of the study, signed informed consent forms and were blinded to the experimental hypotheses.
Design
In order to study the potential of goal-oriented visuomotor amplification for promoting the use of the paretic limb, we use the Rehabilitation Gaming System (RGS) (Fig. 1a), which allows the user to control a virtual body (avatar) seen from a first-person perspective on a computer screen via their own movements that are captured by an imager at 30 Hz (Kinect, Microsoft). Physical execution of goal-directed movement is thus coordinated with the observation of the same movement in VR. RGS includes the Adaptive Biomechanics Controller, which modulates the task difficulty though the amplification of the movement of the virtual limb. Modulation of the movement is achieved by combining two methods: amplifying the amount of movement (i.e. extent amplification) and by attracting the direction of the movement towards the target position (i.e. accuracy amplification) (Fig. 2). Thus range of movement amplification reduces visual errors in movement extent, while accuracy amplification lessens visual directional errors relative to the target. In order to compute the position of the amplified virtual hand at each timeframe, we first extend the vector of the actual hand movement executed by the patient:
$$ \textbf{m}_{e}= \mathbf{m} \cdot G $$
((1))
where m
e
is the vector of the extended hand movement, m is the actual hand movement with respect to the start position, and G is a constant ratio of extent amplification.
Next we project the amplified movement vector m
e
onto the target direction:
$$ \textbf{m}_{p}= (\hat{\textbf{t}} \cdot\textbf{m}_{e}) \hat{\mathbf{t}} $$
((2))
where the operator · denotes a dot product, t is the distance vector from the start position to the target, and \(\hat {\textbf {t}}\) is the unit vector of t.
Finally, the movement amplification at the current frame is defined by
$$ {\textbf{m}_{a}}= \alpha\cdot{\textbf{m}_{p}} + (1-\alpha) {\textbf{m}_{e}} $$
((3))
$$ \textbf{where}~~ \alpha= \frac{|{\textbf{m}_{p}}|} {|\mathbf{t}|} \cdot \frac{1}{G}. $$
((4))
The movement amplification vector m
a
is a weighted combination of two terms: an accuracy amplification vector and an extent amplification vector. The α ratio determines the contribution of each of these two components, and will cancel the amount of amplification of the movement extent when the patient exceeds in distance the desired movement t. Contrarily, if the direction of the executed movement matches the target direction, α will approach 0, thus decreasing the amount of accuracy amplification. After computing the movement amplification vector m
a
and extracting its corresponding hand position (x
′,y
′), we recursively applied an inverse kinematics technique (Cyclic Coordinate Descent) [21] for estimating the angles of elbow and shoulder joints of the avatar. The constant coefficient G was 1.4. The length of the segments of the avatar’s upper-limbs were l
1=0.27, and l
2=0.38. Notice that l
2 denotes the distance from elbow to fingers and therefore exceeds the length of the forearm (Fig. 2).
We developed a new scenario in RGS to quantify and modulate effector selection in stroke patients. Participants were instructed to reach for targets that appeared consecutively in a virtual environment (Fig. 1c, top). At the beginning of each trial, subjects had to position both virtual hands over their corresponding start positions. Start positions were indicated by two green cylinders (7.5 cm diameter) centered 48 cm apart. After the subject maintained the avatar’s hands over the start positions during a variable time interval of 1 ±0.5 ms the two green cylinders disappeared and a target sphere appeared at any of nine possible angles (0°, ±4°, ±8°, ±16°, ±32°) along a semicircular array 65 cm from the projected center of the avatar. Trial time limits (1.75 s) were indicated by continuous changes in the color of the target, which ranged from green to black. Trial time limits were fixed according to the results from a pilot study with stroke patients to guarantee that patients were able to perform a complete reaching movement within this time window. At the end of the trial the target disappeared. The participants were instructed to reach the target as fast as possible with one hand and keep the other hand over the start position. Trials in which the participant moved both hands were automatically invalidated and immediately repeated. Movements in the virtual world were confined to the horizontal plane. Trunk movements in the virtual environment were constrained to ±30° axial rotation. When the center of the virtual hand was placed over the target, participants heard a continuous tone and could observe the increase of their score by 30 points every tenth of a second. These score values were permanently displayed at the top of the screen and accumulated across blocks and phases.
The study was divided into two sessions (Fig. 1b): a familiarization period (S1), and an experimental period (S2). Both sessions were completed during two consecutive days. A session comprised 2 blocks of 14 pointing RW trials each (pre and post phases in Fig. 1b), and 9 blocks of 32 VR-based reaching trials each. VR-based trials were divided in three phases (P1, P2, and P3). In session 2, we refer to these three phases as baseline, intervention and washout. Each of these three phases was divided into 3 blocks of 32 trials each. During the intervention phase we amplified the mapping of the physical movement of the paretic arm to the matched virtual limb. This amplification was progressively and uniformly introduced during the first block of the intervention phase and gradually reduced during the first block of the washout phase. We introduced and suppressed the visuomotor amplification in a gradual fashion to keep participants explicitly unaware of the manipulations.
After each block of trials the patient rested for twenty seconds. In the beginning of each block we included eight forced lateralized trials i.e. four trials with the non-paretic and with the paretic limb respectively, to ensure that participants experience the effect of the kinematic and goal-oriented amplification of the paretic limb. These forced trials where indicated to the subjects by the presentation of only one virtual limb and its corresponding initial position aligned with the position of the corresponding limb of the subject. In the following 24 free-choice trials patients could freely choose their preferred limb. Free choice trials were indicated by the presentation of both virtual limbs at their corresponding initial positions. Within each block of 24 free choice trials, targets appeared at the 0° location on four trials, at the ±4° location on eight trials, at the ±8° location on six trials, at the ±16° location on four trials, and at the ±32° location on two trials. This distribution gave us a higher resolution in the measurement of hand positions closer to the center of the task space therefore reducing ambiguity in assessing hand selection patterns. The sequence of target locations was randomized. The most lateral, ±32° locations, were only reachable by the ipsilateral limb, and were included to decrease the likelihood that participants would use only one hand to reach all the targets. These outer locations were excluded during forced trials.
In order to assess how the effects of visuomotor amplification in VR transfer to real world motor performance, participants performed a pointing task in the real world (Fig. 1c, bottom) before (pre-evaluation) and after (post-evaluation) completing the VR-based reaching task. Each of these evaluations consisted in 14 consecutive trials, the first 7 to be performed using the non-paretic limb and the last 7 using the paretic limb. The trials in which the patient used the non-paretic limb were introduced to guarantee that the task was well understood. In the beginning of each trial participants were instructed to place their index finger at the start position, centered 20 cm in front of them. Next, a colored number from 1 to 7 appeared on the screen and the participant had to point with the whole arm towards the corresponding target until achieving maximum extension. Targets were presented in a semicircular array of cards over the table, displaying colored numbers ordered from 1 to 7 at -32, -16, -8, 0, 8, 16 and 32 degrees, 1 m from the start position. Patients were instructed to self-pace their movements. After completing each pointing movement they returned to the start position. The order of presentation of each of the 7 pointing trials was randomized. Notice that in the real wold task we did not measure hand selection patterns and patients performed one pointing movement per angle and limb within each evaluation phase (pre and post). We took this decision regarding the experimental design in order to shorten its total duration and prevent fatigue.
Outcome measures
In order to evaluate changes in hand selection patterns, we calculate the Point of Subjective Equality (PSE) [16] for each patient over the different phases. The PSE is the theoretical position in space where a subject shows equal probability of spontaneously choosing either limb to reach it. In order to estimate the PSE for each patient and phase we constructed a psychometric function of the probabilities of using the non-paretic/paretic limb to reach towards each target position. Next we fitted the resultant data points using logistic regression as described in [16]. Before estimating the PSEs, we normalized workspaces across subjects by mirroring target directions for those patients with their right arm affected.
To evaluate the participants’ perception of the movement manipulation, a 5-point Likert scale self-report questionnaire was administered at the end of each session (see Additional file 1). The questionnaire consisted of 18 items divided in three different categories referring to the controllability of the virtual arms, subjective competence, and effort, in relation to the paretic and the non-paretic limb. We used reverse polarity on 1/3 of the questions within each category in order to avoid response bias.
Data analysis
To assess the overall within-subject impact of the treatment on effector selection patterns, we performed a 2-tailed Wilcoxon signed-rank test of the PSE values from the baseline, intervention, and washout phase. In order to study the effects of the patients’ reinforcement history on hand selection, we performed a sequential analysis of hand bias by running a two-way repeated-measures ANOVA on the probabilities of selecting the paretic limb during session 2. We used two categorical independent variables: the outcome of the previous trial (success or failure), and the effector selected in the previous trial (paretic or non-paretic).
Performance in the real world pointing task is defined by the patient’s range of motion measured as the average distance covered by the paretic limb for the two targets appearing in the workspace ipsilateral to the paretic arm or the paretic workspace, for the three targets appearing at central locations (at 0°, and ±8°), and for the two targets appearing in the contralateral workspace (i.e. non-paretic workspace). In order to align workspaces across patients, we mirrored target directions for those patients with their right arm affected. We used a 2-tailed Wilcoxon signed-rank test to compare measurements obtained for each workspace and each session, during the pre and post-evaluation. We used Spearman’s rank correlation coefficient to study the dependence between outcome measurements from virtual reality and real world tasks.
To assess the patients’ responses to the questionnaire, we computed the average ratings for those statements belonging to each of the three categories, controllability, performance, and effort, attributed to each limb. Next we analyzed differences within categories, and between sessions, using a 2-tailed Wilcoxon signed-rank test. For all statistical comparisons, the significance level was set to 5 % (p = 0.05). All statistical analysis was done using MATLAB 2013a (The MathWorks, Inc. Natick, MA).
