Short-term learning of the vestibulo-ocular reflex induced by a custom interactive computer game

Abstract

Retinal image slip during head rotation drives motor learning in the rotational vestibulo-ocular reflex (VOR) and forms the basis of gaze-stability exercises that treat vestibular dysfunction. Clinical exercises, however, are unengaging, cannot easily be titrated to the level of impairment, and provide neither direct feedback nor tracking of the patient’s adherence, performance, and progress. To address this, we have developed a custom application for VOR training based on an interactive computer game. In this study, we tested the ability of this game to induce VOR learning in individuals with normal vestibular function, and we compared the efficacy of single-step and incremental learning protocols. Eighteen participants played the game twice on different days. All participants tolerated the game and were able to complete both sessions. The game scenario incorporated a series of brief head rotations, similar to active head impulses, that were paired with a dynamic acuity task and with a visual-vestibular mismatch (VVM) intended to increase VOR gain (single-step: 300 successful trials at ×1.5 viewing; incremental: 100 trials each of ×1.13, ×1.33, and ×1.5 viewing). Overall, VOR gain increased by 15 ± 4.7% (mean ± 95% CI, P < 0.001). Gains increased similarly for active and passive head rotations, and, contrary to our hypothesis, there was little effect of the learning strategy. This study shows that an interactive computer game provides robust VOR training and has the potential to deliver effective, engaging, and trackable gaze-stability exercises to patients with a range of vestibular dysfunctions.

NEW & NOTEWORTHY This study demonstrates the feasibility and efficacy of a customized computer game to induce motor learning in the high-frequency rotational vestibulo-ocular reflex. It provides a physiological basis for the deployment of this technology to clinical vestibular rehabilitation.

INTRODUCTION

When the head turns, the rotational vestibulo-ocular reflex (VOR) maintains gaze stability relative to the external environment by counter rotating the eyes in the orbits. To accomplish this, the map of the semicircular canal (SCC) inputs to extra-ocular muscle (EOM) innervations is calibrated by a precise motor learning process: visual error signals drive synaptic plasticity in the vestibulocerebellum and in flocculus target neurons of the vestibular nuclei to adjust VOR gain and minimize retinal image slip (1–4). Plasticity of the VOR occurs in patients who undergo vestibular therapy but is also induced robustly in healthy individuals via artificial visual-vestibular mismatch (3, 5, 6).

Multiple sensorimotor strategies have been shown to induce VOR plasticity. The reflex gain can be altered by both continuous sinusoidal head motion (7) and by a series of brief head rotations (8, 9). Moreover, not only image motion but also retinal position errors can alter VOR gain (8, 9). It has been shown that the VOR can be adapted differentially for leftward and rightward head rotation (10), which holds promise as a strategy for the rehabilitation of unilateral vestibular hypofunction (11). Schubert et al. (12) have also found that it is more effective to adapt the VOR incrementally in a series of steps with increasing demand rather than attempting to adapt to a large visual-vestibular mismatch all at once. This approach involves gradually changing the amount of target motion relative to the head as the VOR adapts, always working closer to the current level of function.

Peripheral vestibular hypofunction is a common clinical cause of visual and balance impairment that adversely impacts everyday life function (13–15). The mainstay of treatment for these conditions is vestibular rehabilitation exercises that have been shown to be effective for both unilateral and bilateral dysfunction (16–24), central vestibular disorders (25–27), and disorders of higher-order vestibular processing (28). A key component of vestibular therapy is gaze-stability exercises that ask the patient to rotate the head back and forth while trying to maintain fixation on a visual target (19). Although these exercises can be effective, there are several key limitations. First, for large degrees of vestibular impairment, the visual-vestibular mismatch may create an error that is too large for effective motor learning. It is difficult to customize exercises precisely to each patient’s specific level of impairment. Second, there is no reliable way to monitor compliance with therapy and to provide the patient with direct performance feedback. Finally, standard gaze exercises may not be sufficiently interesting to maintain a patient’s attention.

Harnessing new technologies for vestibular rehabilitation has the potential to address these issues. One example is the StabilEyes device that projects a laser-based visual fixation target at a position that can be scaled to head motion, offering the possibility of incremental training (29). Other recent studies have begun to investigate computer gaming and virtual reality as a tool for vestibular rehabilitation (30–35). However, the incremental training approach has not yet been investigated in computerized VOR training, so it is not known if its efficacy translates to video game-based training. Our novel approach was to develop an interactive computer game for gaze stability exercises that can be customized to each individual’s level of vestibular impairment, can train incrementally, and maintains a record of playing and game performance. In this study, we tested the ability of this game to induce rotational VOR adaptation in individuals with normal vestibular function, and we compared the degree of learning in response to incremental and nonincremental training in terms of change in VOR gain, player performance, and dose-dependence of VOR learning.

MATERIALS AND METHODS

Participants

Eighteen healthy adults (ages 18–57, 9 females and 9 males) with no known history of vestibular or other neurological disease participated in this study. None had previously taken part in a VOR adaptation experiment. All participants gave signed informed consent under a protocol that was approved by the Institutional Review Board of the Cleveland VA Medical Center.

General Experimental Procedures

All participants participated in two experimental sessions, each on a different day. The procedures were the same for both sessions, except for the learning paradigm (incremental vs. single-step learning, see VOR Training Strategies). Instantaneous head position was recorded during game playing and VOR testing using a three-axis magnetic-field search-coil system (1.9 × 1.9 × 1.9 m) with a dual scleral coil secured to the forehead. The six raw coil signals were passed through an analog low-pass filter (150 Hz). The signals were digitized at 960 Hz with 16-bit precision and processed online by a custom application written in Simulink Real-Time (SLRT, MathWorks, Natick, MA) that calculated head quaternions and angular velocity and streamed them in real time via UDP to the gaming computer. Eye position during VOR testing was measured with a custom lightweight head-mounted video-oculography system (I-Scan), sampled at 240 Hz. Eye position signals were input to the SLRT application from which they were saved to disk along with raw head position signals for later offline analysis.

VOR Training Game

To induce VOR adaptation, the participant played an interactive computer game, developed with the Unity Game Engine (Unity Technologies, San Francisco, CA), that incorporated visual-vestibular mismatch and a dynamic acuity task. During play, the participant sat with the head unrestrained in the center of the magnetic field coil frame in front of an ultra-wide-screen gaming monitor (Samsung C49HG90DMN, refresh rate 144 Hz, screen dimensions 120 × 34 cm, pixel dimensions 3,840 × 1,080) that was ∼180 cm away (the radius of curvature of the monitor), in an otherwise unilluminated room. The training game was based on a soccer penalty-kick scenario, in which the player was the goalkeeper and had the objective of catching the ball to prevent a goal from being scored. A single scene gave the player a view of the soccer field from the goal. At the center of the scene was the soccer ball that was coming in toward the player. To save the goal, the player was required to predict along which trajectory (of the 8 possible cardinal and diagonal directions) the ball would travel and indicate that direction with a joystick on a game controller. The direction of travel was cued in a dynamic acuity task—a letter “C” (similar to a Landolt-C optotype, although different in its exact proportions) appeared briefly on the center of the soccer ball while the head was turning. The orientation of the C’s opening indicated the direction of ball travel.

The game was structured as a series of individual trials, each representing a single penalty kick. The participant began each trial by centering the head and looking at the soccer ball. An arrow appeared on the ball to indicate the direction the head should be turned (e.g., arrow pointing left cued a leftward head rotation). The required rotation direction was randomized by the game with equal probability to ensure a similar total number of leftward and rightward trials. The game then waited for the participant to rotate the head in the indicated direction. Yaw head speed of 100°/s triggered a flash of the optotype on the middle of the soccer ball for 100 ms (Fig. 1). The participant indicated the perceived optotype orientation with the joystick to complete the trial. If the player selected the direction correctly with the joystick, the gloved hands moved toward the ball and caught it, registering a save; otherwise, the ball was missed and a goal would be scored. The running score was displayed after each trial. When ready, the player recentered the head to initiate the next trial. Three possible conditions led to the premature termination of a trial:

• 1)When the direction arrow appeared, the participant did not begin to turn the head before a predetermined timeout (1.5 s). Because the game waited for the head to be recentered before continuing, this timeout also provided the player the ability to set the pace of the overall game by taking a rest during the play session. Such interruptions, however, occurred infrequently.
• 2)The head was turned in the wrong direction, e.g., the arrow on the ball cued a rightward rotation, but the head was turned to the left.
• 3)The head was turned correctly in the cued direction, but the peak velocity was below the minimum threshold for the optotype to be displayed (100°/s).

Figure 1.

Example trial during game playing. Yaw head velocity from a single representative game trial with leftward head rotation (in accordance with right-hand rule convention, leftward velocities are positive). Display of the Landolt-C optotype was triggered when yaw head speed reached 100°/s. Once illuminated, it remained visible for 100 ms.

In the case of premature trial termination, the participant was given text feedback (e.g., “too slow”) and cued to reset head orientation for a new trial. The duration of each game trial was ∼5 s. Across all participants and sessions, the trial rate during gameplay was 12.6 ± 3.1 trials/min (means ± SD). The total time played depended both on the trial rate and the number of prematurely terminated trials, as the latter did not contribute to the 300-trial total.

VOR Training Strategies

To induce VOR learning, the game incorporated scaled visual-vestibular mismatch: the entire game scene moved on the monitor during head turns by a settable fraction of yaw head velocity. The goal of this study was to elicit an increase in the VOR, analogous to the objective of clinical vestibular exercises, so the visual scene was set to move in the direction opposite to that of the head rotation. This required eye speed to exceed head speed for the retinal image to be stabilized.

The two experimental sessions differed in training strategy. In one case, scene motion was set to -0.5 (×1.5 viewing) for the entire game (300 trials). We term this “single-step learning.” The other experimental session employed an “incremental learning” strategy, increasing scene motion in three equal steps of 100 trials each (scene-motion gains −0.17, −0.33, −0.50; corresponding to viewing conditions ×1.17, ×1.33, ×1.50). Thus, the ultimate learning target for both paradigms was a VOR gain increase of 50%. The order of the two experimental sessions (single-step vs. incremental learning) was pseudorandomized across the group of participants to eliminate a possible confounding order effect.

Optotype size was changed dynamically during game playing. For each set of 10 completed trials, if the percentage of correct responses was less than 60%, the optotype size was increased; if it was greater than 80%, the optotype size was decreased. The goal of adjusting the optotype was to keep the game challenging without making it overly difficult to the point of frustration. For the incremental learning session, the optotype size was reset to mid-range (logMAR ≈ 0.5) at the beginning of each of the three 100-trial blocks.

Experimental Protocol and Data Analysis

Each experimental session followed the same sequence. First, the prelearning VOR was recorded with passive and active yaw head impulses in both directions. VOR testing was performed in complete darkness other than a green fixation LED at the same distance as the monitor used for playing the game (∼180 cm). The LED target flashed every 2 s for 50 ms. For passive head impulses, the examiner centered the head in its range, waited for the LED to flash, and then turned the head, in darkness, either to the right or left, recentering the head before the target reappeared. For active head impulses, the participant was instructed to move the head in the same way. This VOR testing was followed by playing the training game for 300 completed trials, using either the single-step or incremental learning paradigm. The experimental session ended with a second recording of the VOR, this time with active head impulses performed first. After training, no head movement with visual feedback was permitted until post-learning VOR measurements were completed.

Eye and head position data were saved by the SLRT program for later data analysis using custom programs written in MATLAB. Raw signals from the three-axis head coil system were converted to quaternions using standard methods. Horizontal and vertical eye-in-head positions recorded by VOG were calibrated to head position based on fixations of the LED target with the head rotated to different static yaw and pitch head positions. Trial data from game playing were saved separately in a log file by the Unity program. These data included the timing of each trial, the size and orientation of the displayed optotype, and the participant’s response.

The primary outcome measures of this study were the changes in passive and active VOR gains induced by learning, calculated as the ratio of the gains measured after game playing to those measured before the training. VOR velocity gains were determined using an optimization procedure in MATLAB (function fmincon) that incorporated a simple Simulink model of the VOR, with VOR gain and latency as the two free parameters in the model. The error function for the optimization was the squared difference between actual and simulated eye velocity, summed for head impulses for a given condition. Saccades, blinks, and other artifacts were identified using a combination of position (40°), acceleration (7,000°/s²), and jerk (5 × 10⁵°/s³) thresholds (36). For each trial, only the response segment before the first artifact contributed to the optimization. VOR position gain was calculated for each impulse as the ratio of the change in horizontal eye position to change in yaw head position, omitting trials with blink artifacts.

The hypotheses of this study – that playing the video game with visual-vestibular mismatch would lead to an increase in VOR gain and that incremental training would have a larger effect – were tested with the following linear mixed-effect (LME) models using MATLAB® (fitlme function):

$$\text{VOR\hspace{0.17em}gain\hspace{0.17em}∼\hspace{0.17em}Pre\_versus\_Post\hspace{0.17em}+\hspace{0.17em}Incremental\_versus\_SingleStep\hspace{0.17em}+\hspace{0.17em}Active\_versus\_Passive\hspace{0.17em}+\hspace{0.17em}Abducting\_versus\_AdductingEye\hspace{0.17em}+\hspace{0.17em}Head\_Rotation\_Direction\hspace{0.17em}+\hspace{0.17em}Incremental\_versus\_SingleStep\hspace{0.17em}×\hspace{0.17em}Active\_versus\_Passive\hspace{0.17em}+\hspace{0.17em}(1\hspace{0.17em}|\hspace{0.17em}Participant)},$$ (1)

$$\text{VOR\hspace{0.17em}gain\hspace{0.17em}ratio\hspace{0.17em}∼\hspace{0.17em}Incremental\_versus\_SingleStep\hspace{0.17em}+\hspace{0.17em}Active\_versus\_Passive\hspace{0.17em}+\hspace{0.17em}Abducting\_versus\_AdductingEye\hspace{0.17em}+\hspace{0.17em}Head\_Rotation\_Direction\hspace{0.17em}+\hspace{0.17em}Incremental\_versus\_SingleStep\hspace{0.17em}×\hspace{0.17em}Active\_versus\_Passive\hspace{0.17em}+\hspace{0.17em}(1\hspace{0.17em}|\hspace{0.17em}Participant)},$$ (2)

$$\text{VOR\hspace{0.17em}gain\hspace{0.17em}ratio\hspace{0.17em}∼\hspace{0.17em}Incremental\_versus\_SingleStep\hspace{0.17em}+\hspace{0.17em}Abducting\_versus\_AdductingEye\hspace{0.17em}+\hspace{0.17em}Head\_Rotation\_Direction\hspace{0.17em}+\hspace{0.17em}}\left(1\text{\hspace{0.17em}|\hspace{0.17em}Participant}\right),$$ (3)

with Pre_versus_Post, Incremental_versus_SingleStep, Active_versus_Passive, Head_Rotation_Direction, and Abducting_versus_AdductingEye as within-subject categorical variables. Specifically, Model 1 tests the hypothesis that pre- and post-learning gains are different (i.e., that learning occurred). Model 2 tests the hypothesis that VOR learning, as measured by the gain ratio, is greater with incremental training, and it compares the amount of learning in the active and passive VOR as well as between abducting and adducting eyes. An interaction between VOR type and learning type was included to account for possible differences in the effect on active and passive VOR. Model 3 was used to separate results from active and passive VOR testing to investigate further the interaction between VOR type and learning type on VOR gain change. For these models, we used histograms and quantile-quantile (q-q) plots to confirm that VOR gains and gain ratios were approximately normally distributed.

In each of these models, a random intercept (1 | Participant) was included to account for nonindependence within the repeated-measures design in which there may have been differences in the baseline values of the outcome measures (37). For example, in the case of Model 1, the question is whether training produces an increase in VOR gain. The inclusion of the random intercept allows for this effect (VOR gain change) to be detected, even if the baseline VOR gains of individuals vary.

Outside of primary hypothesis testing, individual comparisons in the overall analysis and testing of the secondary outcome measure (dynamic acuity during training) were conducted with paired t tests, as noted specifically in the corresponding text and figure legends where these results are presented.

RESULTS

Head Rotations during VOR Game Playing

Figure 2, A and B show representative head velocity traces and the histogram of peak velocity from one participant during a single training session using the incremental learning paradigm. Individual trial peak yaw head velocities ranged from 105 to 252°/s, but the mean (158°/s) and median (157°/s) fell within the goal range of 140–160°/s. Considering all 36 sessions for the 18 participants, mean peak head speeds ranged from 117 to 176°/s (147 ± 13°/s, means ± SD). Across all participants, the mean was within the goal range for half of the sessions (Fig. 2C). For the group, the mean peak head speed across the entire session did not differ based on the learning method (P = 0.62, paired t test). A difference was observed, however, when data from only the first 100 trials were considered (×1.13 viewing for incremental and ×1.5 viewing for single-step learning). For this initial epoch, the greater visual-vestibular mismatch of single-step learning corresponded to a 7.4% slower mean head speed (Fig. 2D, P = 0.0051).

Figure 2.

In-game head rotations. A: yaw head velocity from a representative subject during game playing. For this and subsequent figures, positive velocities are leftward and negative velocities are rightward. B: histogram of peak head speed (directions combined) for all trials in this example. The median head speed was 158°/s, close to the goal of 150°/s. C: distribution of mean peak velocities for all participants and for both learning sessions (147 ± 13°/s, means ± SD). D: comparison of mean peak head speed during single-step and incremental learning (first 100 trials = ×, last 100 trials = ○). Peak head speed was slower for the first 100 trials of single-step learning (P = 0.005, paired t test), but there was no difference for the final 100 trials (P = 0.13) when visual-vestibular mismatch was the same for both paradigms.

Effect of Training on the VOR

Representative data from passive yaw impulses for one participant are shown in Fig. 3. The increase in eye velocity after incremental training can be seen from both the individual trials (Fig. 3A) and in the two-dimensional velocity plot (Fig. 3B), in which the increased slope after training (red traces) corresponds to a higher VOR gain. Results of the optimization procedure that was used to calculate VOR gain are depicted in Fig. 3C for three head rotations of different velocities, before and after learning. Note that the model fits eye velocity well for all three head impulses. Thus, the increase of VOR gain after learning was similar across this range of head speeds, i.e., VOR learning did not depend on head speed.

Figure 3.

Example vestibulo-ocular reflex (VOR) responses. A: head and left eye velocity for passive VOR trials of one participant before and after incremental training. Eye velocity has been inverted to facilitate comparison with the head velocity of the corresponding direction. Note that after training (post), eye velocity exceeds head velocity, due to an increase in VOR gain evoked by learning. B: 2-D plot of horizontal eye velocity vs. yaw head velocity for the first 173 ms of all trials in A. After training (red traces), the eye speed at a given head speed is greater than at baseline (blue traces). C: representative fits for VOR gain calculation (active head impulses). Each fit was performed on the entire set of head impulses for each direction and condition (active, passive) but only three rotations are shown each for pre- and post-learning. The fit matched the actual eye velocity for the range of head speeds, indicating that gain is independent of speed, even after learning. Gv, velocity gain.

Summary VOR gain data from all participants are shown in Figs. 4 and 5. Most participants’ VOR gain was higher after learning for both active and passive head impulses and for both learning paradigms (Fig. 4). The change in VOR gain was quantified as the ratio of the post-learning VOR gain to the corresponding pre-learning value (VOR gain ratio). Gain ratios were greater than unity for both learning paradigms and for both active and passive head rotations (Fig. 5). Moreover, these gain changes were similar, whether derived from eye and head position or their velocities (Fig. 5), although absolute position gains were on average 4.7% lower than velocity gains (data not shown). Across all four conditions, the gain increase ranged from 11–20% (22–40% of the ideal change of 50%) for velocity gains and 12–16% (24–32% of ideal change) for position gains.

Figure 4.

Comparison of vestibulo-ocular reflex (VOR) gains (from velocity fits, averaged from both eyes and both rotation directions) before and after game playing. The gain increased after learning in most cases, for both active and passive head rotations.

Figure 5.

Vestibulo-ocular reflex (VOR) gain ratios (post/pre, mean ± 95% CI) based on eye velocity and position for each learning paradigm and head rotation type. Gain ratios for each head rotation direction and eye were averaged for each participant. The increase in VOR gain was significant for each condition (P < 0.001 for all cases, paired t tests). CI, confidence interval.

Linear mixed-effect analysis (statistics in Table 1, coefficients in Table 2) confirmed our hypothesis that the vestibular game effectively increased VOR gain for both learning paradigms (P < 0.001 for prelearning vs. postlearning gain). VOR gain was higher in the adducting eye than in the abducting eye (Fig. 6A, Model 1), but there was no difference between eyes with respect to the gain change after learning (Fig. 6B, Model 2). There was also no difference in gain ratios for the two rotational directions (data not shown, Model 2).

Table 1.

P values for LME fits (see materials and methods and results for individual models)

	Model 1 (Velocity Gain)	Model 2 (Velocity Gain Ratio)	Model 1 (Position Gain)	Model 2 (Position Gain Ratio)	Model 3 (Velocity Gain Ratio, Active Rotations)	Model 3 (Velocity Gain Ratio, Passive Rotations)	Model 3 (Position Gain Ratio, Active Rotations)	Model 3 (Position Gain Ratio, Passive Rotations)	Model 4 (NGVE)
Pre vs. post (time)	<0.001	N/A	<0.001	N/A	N/A	N/A	N/A	N/A	<0.001
Adducting vs. abducting eye	<0.001	0.83	<0.001	0.73	0.39	0.40	0.93	0.55	<0.001
Rightward vs. leftward yaw	0.42	0.11	0.41	0.86	0.06	0.90	0.23	0.32	0.916
Active vs. passive VOR (VOR type)	<0.001	0.98	0.765	0.62	N/A	N/A	N/A	N/A	<0.001
Incremental vs. single-step (learn type)	0.54	0.003	0.74	0.19	0.008	0.26	0.16	0.22	0.634
VOR type: learn type	0.003	0.005	0.07	0.06	N/A	N/A	N/A	N/A	0.215

LME, linear mixed-effect; NGVE, normalized gaze velocity error; VOR, vestibulo-ocular reflex.

Table 2.

Coefficients for LME fits (estimate and 95% confidence interval)

	Model 1 (Velocity Gain)	Model 2 (Velocity Gain Ratio)	Model 1 (Position Gain)	Model 2 (Position Gain Ratio)	Model 3 (Velocity Gain Ratio, Active Rotations)	Model 3 (Velocity Gain Ratio, Passive Rotations)	Model 3 (Position Gain Ratio, Active Rotations)	Model 3 (Position Gain Ratio, Passive Rotations)	Model 4 (NGVE)
Pre vs. post (time)	0.141 [0.123, 0.159]	N/A	0.126 [0.110, 0.143]	N/A	N/A	N/A	N/A	N/A	−0.845 [−0.900, −0.791]
Adducting vs. abducting eye	−0.096 [−0.113, −0.077]	0.0031 [−0.024, 0.030]	−0.061 [−0.077, −0.044]	−0.004 [−0.029, 0.020]	0.019 [−0.025, 0.062]	−0.013 [−0.042, 0.017]	0.0014 [−0.030, 0.033]	−0.0093 [−0.040, 0.021]	−0.101 [−0.155, −0.047]
Rightward vs. leftward yaw	−0.0074 [−0.026, 0.011]	0.022 [−0.005, 0.050]	−0.007 [−0.023, 0.009]	0.002 [−0.022, 0.027]	0.042 [−0.001, 0.085]	0.0018 [−0.028, 0.031]	0.019 [−0.012, 0.051]	−0.016 [−0.047, 0.015]	−0.003 [−0.057, 0.052]
Active vs. passive VOR (VOR type)	−0.057 [−0.082, −0.031]	−0.00056 [−0.039, 0.038]	0.004 [−0.020, 0.027]	−0.009 [−0.043, 0.026]	N/A	N/A	N/A	N/A	0.288 [0.211, 0.365]
Incremental vs. single-step (learn type)	−0.0081 [−0.034, 0.018]	0.060 [0.021, 0.098]	−0.004 [−0.027, 0.020]	0.024 [−0.012, 0.059]	0.059 [0.016, 0.102]	−0.017 [−0.047, 0.013]	0.023 [−0.009, 0.055]	−0.020 [−0.051, 0.012]	0.019 [−0.059, 0.096]
VOR type: learn type	0.055 [0.019, 0.092]	−0.079 [−0.13, −0.024]	0.030[−0.003, 0.063]	−0.048 [−0.097, 0.001]	N/A	N/A	N/A	N/A	0.069 [−0.040, 0.178]

LME, linear mixed-effect; NGVE, normalized gaze velocity error; VOR, vestibulo-ocular reflex.

Figure 6.

Comparison of vestibulo-ocular reflex (VOR) velocity gains in adducting and abducting eyes (passive VOR). A: VOR gain before (pre) and after (post) learning, both learning paradigms. In most cases, the gain is larger in the adducting eye. The same was true for active head impulses (data not shown), and considering all data, the difference was highly significant (P < 0.001, from LME of Model 1). B: the gain ratio does not differ between adducting and abducting eyes (P = 0.83, from LME of Model 2). Thus, although the actual gains are different in the two eyes, the degree of learning is equivalent.

Effect of Learning Paradigm on VOR Gain Change

Our hypothesis was that incremental training would produce a greater change in VOR gain than single-step learning for a similar number of training trials. In this study, however, both paradigms increased gain similarly (Figs. 4, 5, and 7). Learning differences between paradigms were small and depended on the type of head rotation and gain calculation method (Table 1). For position gain ratios, there was no clear effect of the training method. For velocity gain ratios, active VOR gain increased more for single-step than for incremental learning (14.1% gain increase for incremental vs. 19.4% for single-step, P = 0.0077), but there was no significant difference for the passive VOR (14.2% gain increase for incremental vs. 11.9% for single-step, P = 0.26).

Figure 7.

Comparison of vestibulo-ocular reflex (VOR) velocity gain ratios. A: incremental vs. single-step learning. B: active vs. passive head impulses.

Effect of Learning Paradigm on Gaze Velocity Errors during Head Rotation

VOR gain is only an approximate measure of VOR performance because it reduces the entire response to a single number. This is most obvious for VOR position gains that indicate only how close the final foveal position is to the point of intended fixation but do not indicate how that gaze error was minimized. To determine the degree to which the gaze is held stable while the head is moving, it is necessary to examine gaze velocity throughout the response. Here, we considered the possibility that equal gain changes by both learning paradigms might obscure a difference in dynamic gaze stability, i.e., incremental learning might lead to an enhanced VOR in which eye and head velocities were more closely matched throughout the duration of the head impulse. To test this, we computed a “normalized gaze velocity error” (NGVE) for the VOR as the sum of the absolute value of gaze velocity relative to the desired fixation goal, normalized to total head velocity to account for variability in the number of head impulses and head velocities across participants and sessions. For the pre-training VOR:

$$\text{NGVE}={\displaystyle \sum }\left|\frac{({\dot{\theta }}_{\text{e}h}+{\dot{\theta }}_h)}{{\dot{\theta }}_h}\right|,$$

and for the post-training VOR:

$$\text{NGVE}={\displaystyle \sum }|\frac{({\dot{\theta }}_{eh}+1.5*{\dot{\theta }}_h)}{{\dot{\theta }}_h}|,$$

where NGVE is the normalized gaze velocity error, ${\dot{\theta }}_{eh}$ is the horizontal angular eye velocity relative to the head, and ${\dot{\theta }}_h$ is head yaw angular velocity. This quantity is minimal when the magnitude of eye-in-head velocity is close to the goal (VOR gain 1.0 before training and 1.5 after training) throughout the entire response (note that the NGVE does not correct for VOR latency, because that is how the VOR is experienced). We tested the effect of learning on NGVE with a fourth linear model:

$$NGVE\text{\hspace{0.17em}∼\hspace{0.17em}PreVsPost+Incremental\_vs\_SingleStep+\hspace{0.17em}Active\_vs\_Passive+\hspace{0.17em}Head\_Rotation\_Direction+\hspace{0.17em}Abducting\_vs\_AdductingEye+\hspace{0.17em}Incremental\_vs\_SingleStep*\hspace{0.17em}Active\_vs\_Passive+}\left(1\text{\hspace{0.17em}|\hspace{0.17em}Participant}\right).$$ (4)

NVGE data were subjected to a Box Cox transformation due to non-normality of error measurements that were based on the absolute value of gaze velocity. There was no difference in NGVE based on the learning strategy, but NGVE was greater for passive than for active rotations and was greater for post-training head rotations (Fig. 8, Table 1, Model 4).

Figure 8.

A: normalized gaze-velocity error (NGVE) during active head rotations after incremental or single-step vestibulo-ocular reflex (VOR) training (active and passive rotations combined). Learning strategy did not impact NGVE (P = 0.90). B: NGVE for passive vs. active head rotations before and after training (data from both training strategies combined, due to lack of effect). The NGVE was greater for passive than for active rotations and was larger for post-training data.

Game Performance during Training

Participants’ performance during game playing can be measured indirectly by the percentage of trials in which they correctly judged the optotype orientation (in the game context, this corresponded to the percentage of goals saved). Were they able to reach, or exceed, the game’s goal of 70% accuracy? Alternatively, game performance can be assessed by the final optotype size achieved during each of the three epochs for both paradigms. The ability to distinguish the orientation of a smaller gap suggests better gaze stability during head rotation.

Response accuracy (Fig. 9A) and dynamic acuity (Fig. 9B) were better for incremental learning only during the first and second epochs when visual-vestibular mismatch was lower than it was during the same epochs of single-step learning. In the final epoch, accuracy decreased for incremental learning as the mismatch increased to match that of single-step learning [P < 0.001, LME fit of %correct ∼ trial_epoch + (1 | Subject)]. Performance improved across epochs of single-step learning (all played with ×1.50 viewing, P < 0.001). Consequently, there was no difference in response accuracy in the third epoch. There was also no effect of response accuracy in the final epoch on the increase in VOR gain for either passive (Fig. 9C) or active (Fig. 9D) head impulses.

Figure 9.

Game performance. A: percentage of correct optotype orientation responses during the three epochs of game playing (mean and 95% confidence interval). Response accuracy differed for the first two epochs but not for the final one (*P < 0.001, **P = 0.019, ***P = 0.14, paired t test). B: equivalent acuity in logMAR based on optotype gap size at the end of each of the three training epochs. This dynamic acuity measure was better for incremental learning at the end of the first two epochs but not after the final one (*P < 0.001, **P = 0.083). C: passive VOR gain ratio (post/pre) as a function of game performance (last 100 trials, P = 0.75). D: active VOR gain ratio (post/pre) as a function of game performance (last 100 trials, P = 0.39). VOR, vestibulo-ocular reflex.

Dose-Dependence of VOR Learning

Finally, we asked whether the degree of VOR learning depended on the amount of exposure to visual-vestibular mismatch. Although the number of successful trials was the same for each game-playing session (300 trials), the total exposure varied due to prematurely terminated trials and a different number of trials during initial pretraining dynamic acuity testing. Independent of total trial number, the amount of exposure also varied, because some participants generated head movements with larger amplitudes and speeds than others (Fig. 2). We found no effect of total head amplitude (Fig. 10) nor the number of total trials nor mean peak head speed (data not shown) on the increase in VOR gain for an individual experimental session. Thus, within the parameters of this experimental design, there was no dose dependence of VOR learning.

Figure 10.

Vestibulo-ocular reflex (VOR) gain ratio as a function of total head turn amplitude during training. There was no effect of total amplitude, and thus of the total amount of visual-vestibular mismatch exposure, on the degree of VOR learning (P = 0.75).

DISCUSSION

In this study, we have shown that a computer game incorporating visual-vestibular mismatch induces robust short-term motor learning in the VOR, increasing VOR velocity and position gains on average by 30% of the visual demand, after a single session of 300 successful trials. Several findings support that this gain increase was due to true motor learning and not simply a cognitive strategy related to game playing. First, the increase was present when the VOR was measured in the dark, outside the gaming context. Second, it affected not only the response to active rotations (as might be expected if the change were a strategy specific to the game) but also, to a similar degree, eye movements evoked by less predictable passive head rotations. Finally, there is evidence that the post-training VOR gain scales linearly with head speed, i.e., the gain increase holds across a range of speeds, even for the passive VOR (Fig. 3C).

Effect of Learning Strategy on VOR Gain Change

Contrary to our hypothesis and different from prior work (12), incremental training was not clearly a more effective driver of VOR gain increase than single-step learning. The only finding was a small effect on the passive VOR, but opposite to what was expected: single-step training produced a larger change in VOR gain than incremental learning.

Comparing the results of our study to those of Schubert et al., should take into account methodological differences that may have influenced the findings. First, studies that showed a benefit of incremental learning employed a small foveal target (12, 29), whereas our game applied the visual-vestibular mismatch to a much larger fraction of the visual field (∼ 37° H × 21° V). Although earlier work found foveal and optokinetic stimuli equally effective in producing VOR learning (7), the same might not be true for rotations of much higher speed and frequency. Indirect support for higher efficacy of larger-field visual stimuli comes from the finding that a foveal target whose contrast is reduced toward that of a typical monitor no longer produces VOR learning with head impulses (38).

Second, the maximal visual-vestibular mismatch applied in our game (50%, ×1.5 viewing) was only half that of the prior studies, in which full error training did not lead to adaptation (12). It could be that we would have observed a similar result for the same large mismatch. Third, the added attention and cognitive effort inherent to an interactive computer game, combined with the motivation of an in-game reward (game score), might boost motor learning beyond that achieved by passive visual following, allowing larger visual-vestibular mismatches to induce learning more effectively. For example, a prior study of robot-assisted motor training after stroke found faster learning in the group that received performance feedback and rewards (39). More study would be required to test the contributions of each of these factors to the efficiency of VOR learning using our game.

Velocity versus position gains.

The effect of learning was similar whether measured by the change in velocity or position gain. Position gain has been used occasionally in quantifying the rotational (8, 40–43) and translational (44) VOR, as well as to quantify ocular counter roll (45, 46). In some cases, VOR “position gain” refers to the ratio of cumulative (integrated) slow-phase eye velocity to the change in head position (40, 41, 43, 47). In essence, this represents average velocity gain, because it eliminates covert saccades that contribute to final eye position. Alternatively, position gain has been calculated simply as the ratio of the change of eye position to the change of head position, either at an early point within the rotation (42) or at the end of the head movement (8). Our study used the latter definition—the ratio of the total change in eye position to the total change in head position, incorporating all elements of the vestibular response throughout the movement. However it is achieved, final eye position is important for rapid head rotations, because it determines whether the point of interest remains foveated when the head rotation is complete. Here, we found a similar effect of training on both velocity and position gains for passive and active rotations. This finding supports a primary vestibular mechanism of VOR motor learning in our study.

Playability of the VOR Game

All participants were able to complete both sessions of the study. None terminated a session early due to nausea or other discomforts, despite the repeated head movements and imposed VVM. The refresh rate of our monitor (144 Hz) was fast enough, and system lags were short enough, for game playing to be well tolerated and for the VVM to produce robust adaptation of both the active and passive VOR. Few of our participants were experienced video game players, yet all learned the game sequence quickly and effectively. Control of a game with fast head rotations was a novel feature, but participants were able to produce the desired head turns consistently with the aid of a brief initial training sequence and feedback from the game while playing (Fig. 2).

From a practical perspective, the fact that robust learning was produced by our game, regardless of the learning strategy, strongly supports the further development of this approach as a tool for clinical vestibular therapy. Although we found no advantage to incremental training for VOR adaptation when vestibular function is normal, this strategy might still be preferred for other reasons. For example, the smaller VVM may be more readily tolerated by patients with vestibular hypofunction. The participants of this study turned their heads more quickly when VVM was lower (comparing the first epoch of incremental learning to that of single-step training). In addition, it may be that patients who do not recover quickly and thus need more aggressive vestibular therapy may be a subset in whom learning is only effective in smaller increments. Moreover, in those with severe VOR hypofunction, the baseline VVM may be larger than the ×1.5 viewing condition imposed in this study. Finally, deploying our game to an immersive virtual reality (VR) headset is part of ongoing work and has the potential to make this intervention more suitable for clinical use. An immersive VR version could increase the likelihood of cyber sickness (48), but the smaller VVM of incremental training could ameliorate this risk. Moreover, our preliminary experience has revealed that cyber sickness could also be mitigated by limiting the field of view inside the headset and reducing the playing period for clinical vestibular therapy to be much shorter than the experimental sessions of this study.

Durability of VOR Adaptation

Here, we did not investigate the retention of VOR changes after learning. A relatively rapid return to the baseline would be expected with re-exposure to the natural visual environment (×1 viewing) (49), especially after gain-increase adaptation (50). Prior studies have shown that more extended retention can be achieved by continuous training for longer periods of time (50, 51); by immobilizing the head in darkness following training (52), thereby eliminating exposure to visual feedback that might detrain the VOR; or by training in a context that is not typical of natural experience (53). Most relevant to the rehabilitation of patients with vestibular dysfunction is the finding that retention of the adapted VOR gain is facilitated when short blocks of training are separated by periods of rest (54). Further study will be required to determine if a similar consolidation effect occurs with our game. Also of note is a key difference between alteration of the intact VOR and gaze stability training in those who have reduced function. In the latter case, the reflex is being trained toward, rather than away from, the gain that is compensatory for everyday experience. Although this might not ensure the retention of learned changes, the new (and lesser) visual-vestibular mismatch would not be expected to drive the VOR back down to the previous lower gain.

Conclusions

This study shows that an interactive computer game in which active head movements control visual scene motion is a highly effective tool for training the VOR in individuals with intact vestibular function. More engaging than traditional approaches, it has the advantages of being readily customized to each individual and of tracking both exercise participation and changes in performance in a way that could ultimately be used to adjust the difficulty and schedule of the training plan to optimize therapeutic benefit. Although this study presented the game on a computer monitor, translation of the software to virtual-reality platforms will provide flexibility for deployment in both clinical and home environments. Future work will adapt this application for standalone immersive VR headsets and will optimize the game for individuals with vestibular hypofunction, providing a flexible platform to test its efficacy for clinical vestibular rehabilitation.

DATA AVAILABILITY

Data will be made available upon reasonable request.

GRANTS

This work was supported by Small Projects in Rehabilitation Research Award I21RX002892 from the United States Department of Veterans Affairs Rehabilitation Research and Development Service; a Steven Garverick Innovation Award from the Advanced Platform Technology Center of the VA Northeast Ohio Healthcare System, Cleveland OH; and by the MetroHealth System, Cleveland, OH (to M.J.F.).

DISCLOSURES

The technology developed and used for this study is part of a pending US patent application (M.J.F. and M.F.W.). None of the other authors has any conflicts of interest, financial or otherwise, to disclose.

AUTHOR CONTRIBUTIONS

Q.L., H.X., W.C., A.S., M.J.F., and M.F.W. conceived and designed research; Q.L., H.X., W.C., A.S., M.J.F., and M.F.W. performed experiments; M.F.W. analyzed data; M.J.F. and M.F.W. interpreted results of experiments; M.F.W. prepared figures; M.F.W. drafted manuscript; M.J.F. and M.F.W. edited and revised manuscript; Q.L., H.X., W.C., A.S., M.J.F., and M.F.W. approved final version of manuscript.