There were two types of reward. The first, a motivational reward, was a coin score directly linked to target and enemy collisions. Subjects won coins if they caught targets, and lost coins if they collided with enemies. When the player caught a target, high contrast gold coins materialized near that object and swept toward a score displayed at the top of the screen. More coins accumulated when targets and enemies were at low contrast, related to a second, contrast sensitivity score (see below). A progress bar on the left side of the screen increased as coins accumulated, and decreased when coins were lost. In addition to coins, auditory feedback and a number of bonus features (e.g., animated helpers, targets, and enemies slowing down) were included to maintain interest in the game.

The second type of reward was a performance-based score displaying contrast sensitivity on a scale of 1–100. This score was calculated from target contrast, which was adjusted continuously using a method described below. Whereas the coin score was derived from the absolute number of target and enemy collisions, the contrast score was based on the relative proportion of collisions, and provided a measure of target visibility. Subjects were told that when they were performing well, the targets would be difficult to see, and that they should try to improve both their coin score and their contrast sensitivity score. At the end of each 90 s level a graph was displayed to subjects showing changes in contrast sensitivity during that level.

Thus, good performance on the game was associated with a high coin score and with a high contrast sensitivity score, and the player's goal was to maximize both scores during each level. Bonus features affected only the coin score and not the contrast score. The top coin scores and contrast scores were saved for each subject and displayed on a “Top 10” page that subjects could access from the menu to view their past performance (Figure 1).

Ten amblyopic children (mean age = 9.3 years; SD = 2.4 years), and 10 amblyopic adults (mean age = 41 years; SD = 8.1 years) participated in the study. All subjects except one had at least a two-line (0.2 logMAR) difference in acuity between the amblyopic and the non-amblyopic (i.e., fellow) eye. One adult subject (SD) was bilaterally amblyopic and had equally poor acuity in both eyes. Six of ten children and six of ten adults had received patching treatment previously. Tables 1, 2 provide clinical details of the subjects. All subjects were informed of the purpose and procedure of the study, which was conducted under ethics approval from the School of Psychology at the University of Nottingham. All subjects provided a detailed ophthalmic history and were refracted by an optometrist before testing. LogMAR acuity (ETDRS) and Stereo acuity (TNO) were measured before and after training. Four adults obtained these measures from their local eye-care specialist before and after playing the game, and sent us the details through email. Two children were tested at Glasgow Caledonian University by a registered optometrist. Subjects who were not tested at the University of Nottingham are marked with an asterisk in Tables 1, 2.

The game was written in HTML5 and JavaScript by Ilixa, a software development company (www.ilixa.com). Ilixa participated in discussions about the design of the game, created all stimuli and the game interface, and devised the algorithm for contrast adjustment based on adaptive staircase procedures (see Section 3.4.3). Subjects were given a demonstration of the game on an Apple G5 iMac computer, on Google Chrome. The monitor was a Trinitron Dell P1130 with a screen width of 40 cm and resolution of 1280 × 1024 pixels. A subset of subjects were also given the demonstration on their laptop computers, which they used to play the game at home. The training sessions were always on subjects' home computers, and included a variety of LCD displays ranging from large-screen television displays to smaller, laptop displays. No iOS devices (e.g., iPhone, iPad) were used. The game was always played with an external computer mouse.

The targets and enemies were broadband, suprathreshold size, gray-scale objects whose contrast varied adaptively against a uniform rectangular gray background (details on contrast follow). All other items in the display besides targets and enemies were colored objects shown at high contrast (see Figure 1, middle panels).

Subjects were refracted and a full ophthalmic history obtained. Subjects were fitted with their best optical correction for the demonstration and an initial practice run on the game, and were instructed to use their best correction when they played the game at home. The game was always played monocularly, with a patch over the fellow eye. Demonstration of the game included subject registration (setting up an ID and password that was used each time the subject played the game), use of the game menu, display calibration, ensuring that target size was set above the acuity limit of the amblyopic eye, and a practice run on the game. During the practice run, the experimenter confirmed that target contrast decreased adaptively from its starting point, indicating that subjects understood how to play the game. Detailed instructions were given on how to set up the game at home. Subjects were told to download the browser Chrome if it did not already exist on their computers, to set their computer monitors to a fixed viewing distance (1 m was suggested, but also see stimulus size section above) for all sessions, and to calibrate their display using the demonstrated method prior to the first session. For children, parents set up the game and confirmed that the instructions were being followed. For subjects who did not attend a demonstration at the University of Nottingham or Glasgow Caledonian University (i.e., four adults), detailed instructions of all steps were provided in a separate document sent through email. Subjects were instructed to play the game every day, for at least 25 min a day. Visual acuity was re-measured after at least 2 weeks of training. A subset of subjects (six children and four adults) continued to play the game after the first acuity re-test, and returned for a second re-test some weeks later.

Each subject had a unique ID, which they used each time they played the game. Their data from each session were stored on a server, which could be accessed by the experimenters at the University of Nottingham. The data for each subject included date of session, duration of session, number of levels played at each session, names of the levels played and the contrast threshold for each level. Hardware and software details were also stored, including the OS, browser name and version, window size (in pixels), and monitor refresh rate. Thus, the experimenter was able to keep track of whether subjects were playing the game regularly, and for a minimum duration each session.

Contrast thresholds from the Trial level, which was obligatory on each run, were used to measure improvement on the game. Contrast thresholds were always measured for the amblyopic eye, with a patch over the fellow eye. For each subject, outlying threshold values (i.e., thresholds more than two standard deviations greater than or less than the mean threshold over all sessions of that subject), were removed from the analyses. Five percent of the children's thresholds and four percent of the adults' thresholds were removed using this criterion. The number of training days between the pre- and post-training acuity tests ranged from twelve to thirty-two (mean = 23.4; SD = 7.21). Thresholds from the very first day (open symbols, Figures 3, 4), were discarded from the analyses as they were measured during the initial demonstration, on a different display than the display used for the remaining sessions. The second day's threshold was considered as the initial threshold value. We measured the amount of improvement as the difference between thresholds on the initial and final training day (i.e., initial threshold minus final threshold). We also calculated learning slopes from linear regression of threshold against day for each subject. According to this measure, improvement is associated with a negative slope significantly different from zero. Improvement on perceptual tasks has also been suggested to increase as a power- or exponential function of the amount of practice (Dosher and Lu, 2007). We compared the fit of a linear model and a decaying exponential model of threshold against session number for each subject using the Akaike Information Criterion (AIC). This criterion produced better exponential fits than linear fits for three of ten adult subjects, and one of ten children (child: DP; adults: BM, IB, JO). For the analyses that follow, we report the linear fits for all subjects.

Performance of amblyopic children and adults on all training days is shown in Figures 3, 4. Figure 5A summarizes the threshold data. The average contrast threshold of children changed from 0.15 (SE = 0.05) on the initial day to 0.09 (SE = 0.02) on the final day. This change in threshold between days was not significant [t₍₉₎ = 1.021, p = 0.33]. As seen in Figure 3, there was substantial variability in children's performance across days, and the learning slope did not differ from zero for any individual child. Contrast thresholds decreased from the initial to final measurement for the adults [Figures 4, 5A, from 0.07 (SE = 0.01) to 0.045 (SE = 0.005)]. This change in threshold was significant [t₍₉₎ = 4.20, p = 0.002]. Figure 4 shows that the improvement in adults was more reliable than in children, with the learning slope for six of ten adults significantly different than zero. Overall these data suggest that for the adults, but not for children, the contrast threshold was a reliable measure of performance on the game. Potential explanations for this difference are discussed later.

LogMAR acuity of the amblyopic eye and the fellow eye of the nineteen subjects is shown in Figure 5B (acuity data were not available for one child participant who was tested outside Nottingham). Acuity of the amblyopic eye improved by 0.12 logMAR both for children [t₍₈₎ = 3.51, = 0.008], and adults [t₍₉₎ = 5.57, p = 0.00034]. Acuity of the fellow eye did not change significantly after training either for children [t₍₈₎ = 1.43, p = 0.19] or adults [t₍₉₎ = 0.6882, p = 0.51]. Figure 5C plots the improvement in acuity of the amblyopic eye against the change in acuity of the fellow eye for all subjects. Improvement in the amblyopic eye outweighed improvement in the fellow eye for all but three subjects, and improvements in the acuity of the amblyopic eye were not correlated with the change in acuity of the fellow eye. Therefore, subjects were not merely improving at reading the letter acuity chart. For four mild amblyopes (children: LG and NJ, initial acuity: 0.14 and 0.20, Table 1; adults: IT and JL, initial acuity: 0.14 and 0.18, Table 2), the difference in visual acuity between the eyes was reduced to less than two lines, rendering them no longer amblyopic according to the criterion of the study. Table 3 gives the correlation coefficient (Pearson's r), and p-values for correlations between absolute amount of logMAR improvement and threshold change, initial logMAR acuity, logMAR improvement in the fellow eye, number of sessions, session duration, total hours played and the age of the participant. Improvements in logMAR acuity were not correlated with most measures, except for a marginally significant positive correlation between initial logMAR acuity and logMAR improvement of the adults (suggesting that poorer starting acuity was associated with larger improvement), and marginally significant associations between improvement and session duration, and improvement and total hours played in children (see below).

A subset of subjects (six children and four adults) played the game for additional days and returned for a second acuity re-test. Performance of these subjects on the game is shown in Figure 6. For these subjects, contrast thresholds did not change significantly between the logMAR pre-test and the first post-test [mean difference = 0.02; t₍₈₎ = 0.82, p = 0.44], or between the first post-test and the second post-test [mean difference = 0.009; t₍₈₎ = 0.52, p = 0.61]. For all subjects except subject WS, learning slopes did not differ from zero. Between the pre-test and first post-test, logMAR acuity of these subjects improved by 1.5 lines [mean = 0.15; t₍₈₎ = 4.6904, p = 0.002]. After additional practice, logMAR acuity improved further by a small amount for children [mean = 0.05, t₍₅₎ = 3.02, p = 0.02; post-test1 minus post-test2], but did not change for adults [mean = −0.02, t₍₅₎ = 2.45, p = 0.09]. The total improvement in thresholds from the initial to the final session of these subjects was not correlated with their improvement in logMAR acuity from pre-test to the second post-test [r = 0.46; t₍₇₎ = 1.40, p = 0.20]. These data suggest that extended play of the game produced slight additional improvements in logMAR acuity only for the children, and that the improvements in logMAR acuity of both groups were maintained over the testing period.

Did the duration of each session affect the amount of improvement? Figures 7A,B show histograms of session duration (in minutes), and and the time of day at which subjects played the game, for all sessions of all subjects. Adults played the game for approximately 24 min on average per session, which was significantly longer than children, who played for an average of 15 min per session [t_(462.249) = 12.05, p < 0.0001]. Therefore, adults, but not children, complied with the suggested duration of practice on the game. Both groups played the game during the latter part of the day, and adults played later than children (median hour: 20:28 vs. 18:52; Wilcoxon rank sum test; W= 32311.5, p = 0.00026). Figure 7C shows the relationship between median session duration for each subject and their logMAR acuity improvement. The correlation between improvement in logMAR acuity and duration was marginally significant for children [r = 0.61, t₍₇₎ = 2.01, p = 0.08], but not adults [r = −0.15, t₍₈₎ = 0.44, p = 0.67]. Note that median duration for adults always exceeded 15 min, and that the range of durations was fairly narrow (Figure 7C).

We examined whether larger amounts of practice, as measured by the total number of hours played by each subject, were associated with larger improvements in logMAR acuity. Figure 7D shows logMAR improvement for all subjects against the total number of hours played. These data include the extra hours played by subjects who trained for additional sessions after their first acuity post-test, and whose acuity was tested more than once (final acuity and number of hours shown for every subject). Figure 7D suggests that larger improvements in logMAR acuity were associated with more hours played; this correlation was marginally significant across all subjects [r = 0.41, t₍₁₇₎ = 1.87, p = 0.08]. When considered for each group separately, the correlation between hours played and logMAR improvement was not significant for adults [r = −0.04, t₍₈₎ = −0.11, p = 0.91], and marginally significant for children [r = 0.66, t₍₇₎ = 2.34, p = 0.05]. Subjects whose acuity was re-tested more than once are indicated in Figure 7D (large symbols), and span the range of logMAR improvements, suggesting that this relationship was not based on acuity re-tests alone. Furthermore, a similar pattern was evident (but not significant) when the number of hours between pre-test and first post-test were considered (see Table 3, Total hours played, Total hours after extended practice), suggesting that the total amount of practice did matter.

Overall, the data suggest that for children, clinically significant improvements in logMAR acuity may depend on a minimum amount of practice per session, and on the total amount of practice across all sessions. The number of sessions was not a good predictor of acuity improvement in children because it was uncorrelated with session duration and total hours played. The relationship between acuity improvement and session duration was less clear in adults, possibly because of the narrow range of durations across subjects. Large improvements in adults also may require far greater amounts of practice than are needed for children, greater than the maximum amount measured in this study. Additionally, the improvement in adults may be limited by other factors than the amount of practice, such as the depth of amblyopia and the properties of the training task. More data including a broader range (and manipulation) of session durations and total amounts of practice are needed to clarify this issue.

Stereoacuity was measured using the TNO test for stereoscopic vision before and after training. Stereoacuity improved from 60 to 30 s of arc for one child, from 120 to 60 s of arc for two children, and from more than 30 min of arc (No Stereo) to 30 min of arc (Gross Stereo) for two children. There was no change in stereoacuity in the other five children. Stereoacuity changed from No Stereo to 480 s of arc for one adult, from No Stereo to Gross Stereo for another adult, and did not change in the other adults. We grouped subjects according to whether their stereo acuity changed (Group 1, N = 7) and did not change (Group 2, N = 11) after game play. There was no difference between these groups in amount of change in logMAR acuity [t_(14.98) = 0.43, p = 0.67], median minutes per session [t_(15.496) = 0.40, p = 0.69], hours played on the game [t_(15.37) = 0.49, p = 0.62], and number of sessions played [t_(10.1) = 0.1.27, p = 0.23]. These results suggest that improvements in stereo acuity could not be predicted from changes in letter acuity, or the amount of practice.

Figure 8A shows average performance on each of the fifteen levels of the game for children and adults. Thresholds were higher for children than adults in all levels, and varied consistently across levels for both groups. These data illustrate the variation in difficulty across levels, which arose from the dynamics of the moving objects on the screen. The level Trial was always played during each session, after which subjects played a subset of the other levels, the choice of which varied across subjects. Figure 8B shows that children's thresholds across all levels were uniformly larger than adult's thresholds, and correlated across levels [r = 0.95, t₍₁₃₎ = 10.99, p < 0.0001]. On average across all levels, children's thresholds were twice adult thresholds [mean = 0.39 vs. 0.18; t_(24.152) = 2.89, p = 0.009].

Figure 8C shows the preference for different levels by adults and children. Each symbol shows the average proportion that a given level was played, of all levels played across all sessions. With fifteen levels in the game, if subjects were choosing all levels equally, the frequency for each level would be approximately 0.07 (i.e., 1/15). However, certain levels were more popular than other levels, and this preference was consistent across adults and children [r = 0.72, t₍₁₃₎ = 3.71, p < 0.002]. The gray symbol shows the level Trial, which was obligatory on each run, and which was played equally often by both groups.

Was level preference predicted by level difficulty? Figure 8D shows the average proportion that each level was played (across all levels and sessions) against the average contrast threshold on that level, separately for adults and children. The frequency that each level was played was uncorrelated with difficulty as measured by the average threshold, both for adults [r = −0.32, t₍₁₃₎ = −1.22, p = 0.24], and children [r = −0.27, t₍₁₃₎ = −1.05, p = 0.31].

An improvement of 1.5–2 lines in logMAR acuity emerges as the standard effect size from a number of studies on perceptual learning in amblyopia (see Levi and Li, 2009; Levi, 2012, for reviews). This effect size is fairly stable despite the considerable variety of tasks and practice durations used. Prolonged practice does confer additional benefits on the trained task in certain cases, (e.g., Li et al., 2007, 2008), but reports of complete resolution of amblyopia (based on the criterion of equivalent visual acuity in both eyes), are confined to cases of mild amblyopia (e.g., Li et al., 2011, and the present study). In the present case, the total number of hours played was positively correlated with logMAR improvement in children but not adults (Figure 7), and visual acuity improved slightly after additional practice only for children and not adults (Figure 6). This difference between children and adults may have arisen due to a longer initial period of procedural learning in children, due to differences in the amount of time played per session (i.e., the distribution of total practice across sessions), or simply due to differences in the maturity of sensory, cognitive and motor skills of the two groups. The maximum duration of game play here was 18.3 h, less than the 50–100 h reported to produce asymptotic performance in amblyopes on a positional discrimination task (Li et al., 2007, 2008). However, the relationship between total hours played and improvement in logMAR acuity was independent of performance on the game, which varied considerably and did not improve for children, but was more reliable, and improved in adults. In Li et al. (2007) as well, asymptotic performance on the trained task after an extended amount of practice was not accompanied by full resolution of the acuity difference between eyes. Hess et al. (2012) also have reported the absence of a correlation between total hours of play on a dichoptic video game and outcome measures including improvement on the game, improvement in logMAR acuity, and stereoacuity. This pattern of results suggests that for severe amblyopia, there may be a ceiling on the functional benefits of practice-based approaches that currently are being tested, and that the sheer amount of practice whilst beneficial for task performance per se, may only go so far toward improving visual acuity.

What are the ingredients for a compelling video game? Features such as the stopping rule, for instance, may influence subjects' engagement over multiple sessions, or within a single session. These features are especially relevant for younger age groups less motivated by the functional benefits of game playing.

Higher contrast thresholds for children than adults have elsewhere been attributed to the immaturity of the sensory system rather than non-visual, attentional factors (Liu et al., 2014). Here however, a number of factors could have contributed to childrens' larger variability in thresholds across sessions. The adult subjects were motivated by the visual benefits of the game and played the game regularly and for longer periods than the children (Figure 7). Keeping the children on task was more challenging. Several parents reported difficulty in motivating their children to play the game after the initial week. As with patching, even video games may carry issues of compliance when prescribed for children. Enhanced game design could address this issue to increase the attractiveness of the game, for instance by varying some other stimulus attribute than contrast, adding narrative elements, and changing the features described above. In the present case although motivational issues may have affected performance on the task, improvement in acuity did not differ between adults and children. Furthermore, although children's thresholds were generally higher than adults', they were strongly correlated with adults' thresholds across the different levels (Figure 8), suggesting that the variability in performance across levels was not due only to differences in motivation or skill. Children also tended to prefer the same levels as adults, suggesting that certain aspects of the game appealed to both age groups equally.

The mechanisms of learning of basic sensory tasks and of video games continue to be investigated. Learning may reside in plasticity of low-level representations (Jehee et al., 2012), decoding or decisional rules (Law and Gold, 2008; Gold and Ding, 2013), attentional sharpening (Otto et al., 2010), or in some combination of these. Perceptual learning has also been attributed to increased sampling efficiency (Gold et al., 1999), reduced internal noise (Lu et al., 2006), or a combination of both (Dosher and Lu, 1998; Lu and Dosher, 2009). Stimulus specific learning, a characteristic of perceptual learning in the normal visual system, was until recently interpreted as evidence for plasticity of sensory representations in primary cortices (see Karni and Bertini, 1997, for an earlier review). It is now clear that stimulus specificity depends on a number of factors including task difficulty (Wang et al., 2013), the axis of generalization (Webb et al., 2007), and attributes of the training regimen (Xiao et al., 2008; Zhang et al., 2010; Hussain et al., 2012a; Hung and Seitz, 2014). Therefore, stimulus specificity (or generalization) in itself cannot isolate the neural mechanism of learning. The broader-than-normal generalization of learning found in amblyopia on contrast sensitivity tasks (Huang et al., 2008; Astle et al., 2010) and the generalization of improvements from the trained tasks to logMAR acuity could reflect higher order learning, but also may reflect the greater capacity for improvement in low-level representations in developmentally impaired visual systems. Overall, the exact mechanisms of improvement after practice of sensory tasks are not yet clear, but the scope for improvement does appear larger in the amblyopic- than in the normal visual system. Improvements in visual function through perceptual learning can be enhanced through pharmacological and environmental interventions (e.g., fluoxetine, dark exposure) that relax the constraints on neural plasticity (Baroncelli et al., 2011; Montey and Quinlan, 2011). Due to the limited practicality of these interventions, other methods of enhancing the functional benefits from perceptual learning remain an area for future work.

We tested the game on a small number of subjects (10 adults and 10 children), and found a positive effect on visual acuity after playing the game for 12 or more sessions. We compared improvements in acuity between the amblyopic and the fellow eye to rule out test-retest effects, and learning on letter-based measures of visual acuity. A more comprehensive study would have included a no-training group, a no-training group that was patched for a similar duration each day as the target group, and groups that played variations of the game to isolate its relevant components. All subjects in this study played the game at home, outside the supervision of the experimenters. Children may or may not have been supervised by their parents. In the above respects, the improvements in visual acuity cannot unequivocally be attributed to game play or to sensory plasticity. We also note a number of recent studies that call into question whether perceptual and cognitive benefits truly arise from practice on video games (Boot et al., 2011; Oei and Patterson, 2014; van Ravenzwaaij et al., 2014). Larger-scale studies, using randomized controlled trial methodology are needed to establish whether video game playing improves perceptual or cognitive skills in the normal population more generally, and in clinical populations specifically.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.