Design of V2BaT system
The V2BaT system consisted of (1) VR-based task, (2) weight distribution and threshold estimator, (3) WiiBB–VR handshake, (4) heel lift detection, (5) performance evaluation and (6) task switching modules.
VR-based task module
In this study, we have designed a database of VR-based standing balance tasks using Vizard software toolkit (from WorldViz LLC.). The database comprised of 72 VR-based tasks in three contextual settings such as land, water and sky. Each setting had variations. For example, land-based settings projected task environments that had skaters on the road, skiers on ice, etc. For water-based setting, there were swimmers under water, etc. For the sky-based setting, there were flying helicopters, birds, etc.
Figure 5 shows two task environments based on land and water. In each task environment, a context-relevant virtual object, an end goal/target and intermediate milestones were used. For example, in land-based task environments (Fig. 5a), the virtual object was an avatar standing on a road while wearing roller skates on both legs, the end target was a divider at the end of the road, and intermediate milestones were traffic cones located at regular intervals on either side of the road. The end target was meant to be a final milestone to be reached to complete a task. The intermediate milestones were used to help the participant to be able to gauge his/her improvement in CoP maneuvering ability within different tasks. As the participant standing on the two WiiBBs gradually shifted his/her body weight in the anterior direction, the virtual object moved forward toward the end target that was anterior to its initial position in the VR environment (thus reflecting the participant’s weight shifting). As soon as the virtual object crossed the successive intermediate goals (in this case, the traffic cones), these appeared to move backward outside the participant’s view. The idea was to help the participant to gauge his/her ability to shift weight in the anterior direction along with a visual metric in terms of intermediate goals moving backward. In other words, this can offer an individualized metric to monitor the improvement in CoP maneuvering capability in each trial of the tasks. A variation in the degree of task difficulty was also a part of the system design. However, depending on the task difficulty, the scale of displacement of the virtual object with CoP excursion (due to weight shifting) was different for each environment. The degree of task difficulty depended on the conditioned weight distribution (described below in "Weight distribution and threshold estimator module" section) among both the legs of a participant. Here, the tasks were designed to offer visual feedback on one’s directional weight-shifting capability. We chose standing balance tasks that require only the anterior weight-shifting tasks instead of mediolateral weight shifting so that our post-stroke hemiplegic participants do not take advantage of the dominance of one leg compared to the other during shifting weight in the mediolateral direction.
Once a participant completed a task, the V2BaT system provided feedback based on his/her performance. The feedback was given audio-visually through a short audio tone along with one star (*) (akin to the reward system typical of popular commercial computer games). Based on the performance in a task, a participant could receive 1 to 5 stars. Also, V2BaT encouraged the participants by saying either ‘Well done; you are doing great’ (for ‘Adequate’ performance or ‘Keep trying, you can do better’ (for ‘Inadequate’ performance) using pre-recorded audio files.
Weight distribution and threshold estimator module
Our participants were hemiplegic with different weight-bearing abilities on either of their legs and with a spectrum of post-stroke balance disorders. Thus, it was important that the conditioned weight distribution across both the legs of a participant was individualized. Before interacting with V2BaT system, we captured one’s individualized residual weight-bearing ability for each leg. We developed a VR-based weight-shifting task (Fig. 6), specifically designed to estimate one’s residual ability to displace the CoP in the anterior direction. This task was used to compute the individualized threshold for CoP displacement that corresponded to a maximum of 100 points (on a 0–100 scale) in a task.
Figure 6 shows the VR-based task that projected a VR world with a pair of boots, one for left and other for right leg placed in a forest environment. This task required the participant to move the boots in the anterior direction as far as possible inside the forest by shifting weight in anterior direction with each boot being controlled by the CoP displacement due to each leg (measured by two WiiBBs). Then, our system computed the range of movement (representing the CoP displacement) of each boot from their initial position (that is while standing upright without weight shifting). Based on a few trials, the maximum CoP displacement (∆CoPmax_L and ∆CoPmax_R) was computed for both the right and left legs, respectively. Then, the higher CoP displacement among ∆CoPmax_L and ∆CoPmax_R was chosen as overall maximum CoP displacement (∆CoPmax). Subsequently, the individualized threshold (∆CoPTHRESH) was estimated [Eq. (1)]:
$$\Delta {\text{CoP}}_{\text{THRESH}} = \, ( 1+ \delta ) \, \Delta {\text{CoP}}_{\rm{max} } .$$
(1)
The factor \(\delta\) (\(\delta\) = 0.2, chosen as an initial approximation and maintained constant throughout the V2BaT-based exercise session) was introduced to intentionally decide the ∆CoPTHRESH exceeding the participant’s best possible weight-shifting capability at the beginning of the task (in which the participant was asked to displace the virtual boots as far as possible in the forest environment). This was to accommodate the conservative weight shifting of the post-stroke patients while interacting with the VR-based task developed for estimating ∆CoPTHRESH. Subsequently, the participants were asked to interact with the tasks offered by V2BaT system. To achieve the conditioned weight distribution while executing the V2BaT task, we needed to estimate one’s initial weight distribution [WL_ini and WR_ini; Eqs. (2) and (3)] as far as left and right legs of hemiplegic participants were concerned followed by updated weight distribution corresponding to each task:
$${\text{W}}_{{{\text{L}}_{\text{ini}} }} \left( {\text{\%}} \right) = \left( {\frac{{\Delta {\text{CoP}}_{{{ \rm{max} }\_{\text{L}}}} }}{{\Delta {\text{CoP}}_{{{ \rm{max} }\_{\text{L}}}} + \Delta {\text{CoP}}_{{{ \rm{max} }\_{\text{R}}}} }}} \right) \times 100$$
(2)
$${\text{W}}_{{{\text{R}}_{\text{ini}} }} \left( {\text{\%}} \right) = \left( {\frac{{\Delta {\text{CoP}}_{{{ \rm{max} }\_{\text{R}}}} }}{{\Delta {\text{CoP}}_{{{ \rm{max} }\_{\text{L}}}} + \Delta {\text{CoP}}_{{{ \rm{max} }\_{\text{R}}}} }}} \right) \times 100 .$$
(3)
Here, sign ‘×’ represents the scalar multiplication.
WiiBB–VR handshake module
During the balance training, the position of the virtual object in the VR environment ("VR-based task module" section) was controlled by the weighted sum of the CoPs obtained from the two WiiBBs. As the V2BaT tasks progressed, we monitored one’s task performance and accordingly we went for updating the weights WL and WR (with initial weight being WL = WL_ini and WR = WR_ini) across both the left and right legs. Since the task was to shift one’s weight in the anterior direction, we used only the ‘y’ component of the CoP (CoP displacement along the anterior direction) for navigating virtual object for display. However, both ‘x’ and ‘y’ coordinates of CoP were stored for subsequent offline analysis. The raw CoP data acquired at 30 Hz were processed by a 5-point moving average filter. The position of the virtual object was determined from the filtered CoP data by using Eq. (4):
$$\left[ {\text{y}} \right]_{{{\text{VR}}_{\text{OBJ}} }} = {\text{W}}_{\text{L}} \left[ {\text{y}} \right]_{{{\text{CoP}}_{\text{L}} }} + {\text{W}}_{\text{R}} \left[ {\text{y}} \right]_{{{\text{CoP}}_{\text{R}} }}$$
(4)
where WL and WR are the task-specific weight factors; \(\left[ y \right]_{{{\text{CoP}}_{\text{L}} }}\) and \(\left[ y \right]_{{{\text{CoP}}_{\text{R}} }}\) indicate the ‘y’ coordinate of the CoP as measured by the WiiBBs for the left and right leg, respectively.
Heel lift detection module
We wanted to ensure that the participants followed Ankle strategy, an important requirement during standing balance task [43]. Thus, the participants were asked not to lift their heel from the surface of WiiBB while shifting their weight. To identify whether the Ankle strategy was ‘Followed’ or ‘Not Followed,’ we used an ultrasonic sensor-based heel lift detection module (Fig. 7a) that wirelessly communicated the height of the heel above the BoS (surface of WiiBB) at 60 samples/sec to V2BaT system. First, the heel lift detection module was initialized. For this, one was asked to stand upright with his heels touching the surface of WiiBB, and the initial distance (dini (mm)) between the ultrasonic sensor mounted on the participant’s paretic leg (Fig. 7b) and the surface of WiiBB (BoS) was measured. While the participants performed the weight-shifting task, our system continuously measured the distance known as the instantaneous distance (dins) between the ultrasonic sensor and BoS. The two distances, i.e., initial distance (dini) and instantaneous distance (dins), were used to detect one’s heel lift.
The output from the ultrasonic sensor was transmitted wirelessly to the task computer to detect the heel lift via a microcontroller-based circuit. The decision of whether Ankle strategy has been ‘Followed’ or ‘Not Followed’ was taken based on the following equations:
$${\text{~Ankle~strategy}} = \left\{ {\begin{array}{*{20}c} {{\text{Followed;}}\quad ~~{\kern 1pt} \quad if~d_{{{\text{ins}}}} < d_{{{\text{Limit}}}} } \\ {{\text{Not~followed}};\quad ~if~d_{{{\text{ins}}}} \ge d_{{{\text{Limit}}}} } \\ \end{array} } \right.$$
(5)
$$d_{\text{Limit}} = d_{\text{ini}} + d_{\text{th}} ,$$
(6)
where dth = 20 mm = height tolerance. For details on heel lift detection module, please see our companion paper [44]. If the Ankle strategy was ‘Not Followed,’ then a penalty factor was added to the performance score (described below). Otherwise, no penalty factor was considered.
Performance score evaluation module
While the participants performed VR-based tasks, the V2BaT system computed their performance scores. The first performance metric PS1 [Eq. (7)] looked into the CoP displacement:
$${\text{P}}_{{{\text{S}}1}} = 100 - \left( {\frac{{T_{\text{L}} - T_{\text{D}} }}{{T_{\text{L}} }}} \right)100.$$
(7)
Here, TL = length of the straight path between the initial and end target positions; TD = length of one’s CoP displacement (ΔCoP) in the VR environment during weight shifting. The second performance metric PS2 [Eq. (8)] was used to penalize the participant for not following Ankle strategy. The penalty was decided from the duration a participant lifted his heel (TLift) as a percentage of the total task completion time (TCT).
$${\text{P}}_{{{\text{S}}2}} = \left( {\frac{{T_{\text{Lift}} }}{{T_{\text{CT}} }}} \right)100 .$$
(8)
The final percentage performance score (%Pf_Score) was calculated as
$${\text{P}}_{{{\text{f}}\_{\text{Score}}}} = {\text{P}}_{\text{S1}} - {\text{P}}_{\text{S2}} .$$
(9)
The V2BaT system was made adaptive to one’s task performance score. One’s performance was considered as ‘Adequate’ or ‘Inadequate’ based on the percentage performance score. For example, if the score was ≥ 70%, then it was considered as ‘Adequate,’ else ‘Inadequate.’ Please note that the threshold of 70% for the performance score was taken as an initial approximation since a performance score of 70% can be considered as satisfactory during initial sessions of exercise in robot-assisted rehabilitation tasks [45], technology-assisted skill learning [46], etc. This can be adjusted based on the study design.
Task switching module
The rationale of operant conditioning
In this study, we used operant conditioning paradigm for balance training through an implicit and subtle cueing technique, presented subliminally by gradual, individualized and controlled variation of the weight distribution across both the legs during the weight-shifting task. This was achieved by subtly increasing the weight contribution [i.e., weightage WL/WR in Eq. (4)] for the paretic leg so that if a participant increased the usage of the paretic leg to displace virtual object, then the V2BaT system rewarded him/her with a higher displacement in virtual object that in turn resulted in higher performance score in the task. For this, the V2BaT was programmed to offer balance tasks of different difficulty levels coupled with the reward based on the task performance that can be considered as a representative of one’s weight-shifting ability.
Task switching rationale
Post the VR-based task offered for estimating \(\Delta {\text{CoP}}_{\text{THRESH}}\) ("Weight distribution and threshold estimator module" section), the participants were invited to start interacting with VR-based tasks ("VR-based task module" section) offered by V2BaT using the task switching rationale designed with an implicit operant conditioning regime (Fig. 8). The task switching was done using two conditions, namely Condition1 and Condition2:
$$\begin{aligned} & {\text{Condition}}_{1}{:}\,\% P_{{{\text{f}}\_{\text{score}}}} \ge 70\% \,(`{\text{Adequate'}}) \hfill \\ & {\text{Condition}}_{2}{:}\Delta {\text{W}}\left( { = \left| {W_{\text{L}} - W_{\text{R}} } \right|} \right) > \Delta . \hfill \\ \end{aligned}$$
(10)
In Eq. (10), Pf_Score is the final percentage performance score [see Eq. (9)] in a task trial. The quantity \(\Delta W\) is the absolute difference between task-specific weight factors for the left leg (WL) and right leg (WR), respectively. The parameter \(\Delta\) is an arbitrary threshold value for the difference between the weightage allotted to one’s left and right legs. The value of \(\Delta\) (= 5%) was considered as an initial approximation and can be changed based on the study design.
Here, the tasks were of two types, namely (i) catch trial (CT) and (ii) normal trial (NT). In the CT, equal weightage, i.e., WL = WR (Eq. (4), similar to the study by JhonBabič [47]) was allocated to each of the paretic and non-paretic legs of hemiplegic post-stroke participants. In NT, the weightage allocated to each of the paretic and non-paretic legs was not equal. Specifically, WL and WR were updated keeping the operant conditioning in mind. The NT tasks were of different challenge levels (NT_Level) based on the distribution of weights, e.g., values of WL and WR. For a task in the first NT challenge level (NT_Level1), the weightage was WL1 = WL_ini and WR1 = WR_ini. For subsequent NT_Level, the weightage for the paretic and non-paretic legs was increased and decreased by a factor of Δ (5% in this case), respectively. The value of Δ was chosen as an initial approximation, and it can be changed based on the study design. The values of WL and WR were continuously updated as long as the difference between WL and WR (i.e., ΔW = |WL − WR|) was greater than Δ [Condition2 in Eq. (10)]. Also, the participants were switched from one NT_Level to the next only when they scored ‘Adequately’ in the task belonging to an NT_Level (Condition1), and Condition2 was also satisfied. Else, the participant was offered tasks (i.e., task trials) with the same weightage (i.e., without updating WL and WR) until the participant scored ‘Adequately’ (i.e., Pf_Score ≥ 70%). Thus, for a particular NT_Level, there could be ‘n’ task trials and represented as NT_Levelin where ‘i’ represents the challenge level. Also, the V2BaT system offered intermediate CTs (single-task trial) before switching the challenge level of NT. Our idea was to (i) help the participants learn to increasingly use their paretic leg while exercising during NT task trials and (ii) help us to understand the effect of operant conditioning on the weight-bearing capability of the paretic leg under real-life situations when one is expected to use both the legs to a similar extent (i.e., CT task trial).
The total time of balance training (TBT) was 20 min. The task execution started with NT_Levelin tasks (WL = WL1 = WL_ini and WR = WR1 = WR_ini), with i = 1 and n increasing till Condition1 was not satisfied or TBT ≤ 20 min. Once the Condition1 was satisfied, V2BaT offered CTi task of the single trial while storing the weightage factors (WL and WR) used in the completed NT_Levelin task trial. Note that the CTi for i = 1 was considered as CTFirst task. Subsequently, we checked for the Condition2 (Cases 1 and 2) before going ahead with the next NT_Levelin (i > 1).
Case1: If Condition2 was satisfied, then V2BaT system offered a task of next NT_Levelin with i = 2 to the participant. At this NT_Levelin (i = 2), the V2BaT system offered several normal trials (n = 1, 2, 3…) till the participant’s Pf_Score ≥ 70%. Subsequently, next CT task, i.e., CTi with i = 2, was offered by the V2BaT system. This whole process was repeated till the Condition2 failed or TBT > 20 min.
Case2: If the Condition2 was not satisfied, then V2BaT system offered CT task repetitively until TBT > 20 min. In this case, the offered catch trials were considered as CTFinal task trials.
Again, there can be two variations in participation. For example, one variation can be that a participant completed the task execution offered by V2BaT system while staying in Case1 until 20 min was over. Then, the V2BaT system terminated the VR-based training by offering the last task as a CTi task with i = Final. The other variation can be that the participant reached Case 2 before completion of 20 min. In that case, the V2BaT system offered several CTi (with i = Final) task trials until 20 min was over. At the end of 20 min, the V2BaT system offered one additional CTFinal task (for the sake of similarity to that for Case1). Thus, for Case2, there were several CTFinal tasks. We were interested to understand the best performance of a participant at the end of the VR-based balance training. For the participants concluding the task execution while being in Case2, we wanted to understand the best performance achieved by the participant out of the number of CTFinal tasks. Instead of considering the performance for the last task of the CTFinal tasks, we chose the one among the CTFinal tasks for which the participant scored the maximum (i.e., best of final CT task trials [CTB_Final, henceforth)] so as to avoid the effect of any monotony arising out of no variation in the challenge level on the performance. However, if a participant remained in Case1 till TBT > 20 min, then we had no option rather than considering the performance during the last, i.e., CTFinal task (offered just after completion of training duration, i.e., TBT = 20 min) as the CTB_Final.
Participants
The study was carried out in hospital settings after informed consent. Twenty-nine hemiplegic post-stroke survivors (S1–S29) [mean (SD) = 49.55 years (13.89)] volunteered in the study. They had varying residual balance and post-stroke periods (Table 1). The inclusion criteria were (1) ability to follow the instructions and (2) ability to stand and shift weight without orthopedic aids.
Experimental setup
Figure 9a shows the experimental setup that consisted of (i) two WiiBBs, (ii) a pair of slippers, (iii) a heel lift detection module and (iv) a task computer (PC). The two WiiBBs were placed 1 mm apart on the ground. To avoid fluctuation in the CoP values due to the participant’s movement affecting the computation, the WiiBBs were fitted with slippers. This was necessary as the initial position of the virtual object was calibrated to one’s initial position at the start. The position of the slippers was maintained (Fig. 9b) similar to the setup used by Mansfield et al. [17]. A heel lift detection module was used to monitor whether the Ankle strategy was followed or not followed.
Procedure
Our study required a commitment of approximately 45 min from each participant. Once a participant arrived in the experiment room, he/she was asked to sit and relax for 5 min. Then, a physiotherapist in our team assessed the participant’s residual balance using the Berg Balance Scale (BBS) [48] measurement and also ensured that the inclusion criteria were satisfied. Subsequently, the experimenter explained the experimental setup and demonstrated the VR-based tasks to the participant. This was followed by the administration of consent form signing by the participant. Additionally, we also told the participant that he/she was free to quit or take breaks in between the balance training session at any time in case of discomfort.
Once the participant was ready, the experimenter fitted the heel lift detection module to the participant’s paretic leg and asked him/her to stand upright with his/her feet in the slippers attached to the WiiBB (Fig. 9a). Then, the experimenter started the study by exposing the participant to VR-based task designed for estimating ∆CoPTHRESH ("Weight distribution and threshold estimator module" section). In this task, the experimenter asked the participant to stand upright for 10 s so that their baseline CoP due to left and a right leg can be estimated. Also, we recorded the initial distance (dini) between the ultrasonic sensor of the heel lift detection module and the surface of WiiBB. Followed by this step, the participant was asked to shift weight as much as possible from the initial position in the anterior direction while following the Ankle strategy to displace the virtual objects (pair of the boot in Fig. 6) as far as possible in the forest. This process was repeated three times, and the maximum CoP displacement achieved by individual leg has been used to estimate ∆CoPTHRESH as mentioned in "Weight distribution and threshold estimator module" section. Once the threshold CoP displacement was estimated for a participant, he/she was offered the VR-based tasks of different templates (section ‘VR-based Task Module’) for 20 min following the rules of the game engine described in "Task switching rationale" section.
Statistical analysis
While the participants interacted with our VR-based tasks during Stage 2, the V2BaT system offered various NTs of various challenge levels and intermediate CTs. Also, it computed their performance, i.e., %Pf_Score (section ‘Performance Score Evaluation Module’) and recorded displacement in CoP (∆CoP) due to the individual leg. We were interested to understand whether the operant conditioning paradigm using V2BaT system contributed to any statistical improvement in one’s performance and enhanced displacement in CoP from their first CT, i.e., CTFirst task to the best of final CTs (CTB_Final) task. The Shapiro–Wilk test of normality on the participants’ performance and ∆CoP data corresponding to CTFirst and CTB_Final revealed that these were normally distributed. Subsequently, we performed a Student’s t test with a significance level set at p value < 0.05 to check the significance of the improvement.