Can artificial intelligence be used to improve the quality of vision in patients with amblyopia? The first digital pill in medicine administered under AI

KEYWORDS:

• AI
• amblyopia
• binocular therapy
• attention
• digital pill
• digital treatment

1. In a nut shell

We live in an ever-increasing digital world, digital applications and diagnostics are part and parcel of everyday clinical practice, particularly in Ophthalmology. All previous digital applications have been directed to improving our diagnostic capabilities but now, for the first time there is a digital pill. This is not a digitally activated/tracked pharmaceutical [1], this is a digital treatment per se, one designed to recover visual function in later life that had been lost in childhood. The neural substrate for this is residual neural plasticity present after the early critical period for visual development. There appear to be two pathways, a bottom-up one that is reliant on the appropriate dichoptic stimulation [2] while the other is a top-down one from higher brain areas that involves attention [3–5]. Treatment efficacy depends on activation of both pathways and there is an important role for AI in ensuring that attention is fully engaged for the duration of the bottom-up stimulation which is administered on a videogaming platform suitable for engaging young children.

2. Two game-changing discoveries

For centuries [6], amblyopia has been thought about in one way and treated in one way. The focus has been on the monocular loss of vision and the treatment has been long-term patching of the fellow sighted eye. The assumption is that the primary problem is the monocular deficit and the loss of integrated binocular function, its consequence. Our views on this are changing as evidence has been presented over the past two decades to suggest that the loss of binocular function is the primary deficit and the loss of monocular vision, its consequence [7]. Further, if the restoration of normal binocular function is the therapeutic goal, then patching one eye for long periods of time is unlikely to achieve it [8]. A binocular problem needs a binocular treatment! The second important discovery in the last two decades is that the basic binocular circuit is intact in amblyopia but simply rendered monocular due to an imbalance in the reciprocal inhibitory connections between the two eyes [9]. Since we know that there is still a degree of brain plasticity even in the adult [9–11], there was hope that with the right(imbalanced) binocular stimulation, it might be possible to reset this interocular balance to restore binocular vision. This was indeed shown to be the case [9] and so out of this new understanding of both the etiology of the condition and its underlying neural substrate, a new approach emerged, one that was ideally suited to the emerging digital revolution.

3. The digital pill - phase 1, supervised administration

In the initial-phase, laboratory studies were carried out on adults [12–18] with amblyopia since it had been believed that this condition could not be treated beyond the age of 12 yrs [19] and there was a concern among the more conservative elements of the profession, particularly those wedded to patching therapy, that to try to restore binocular vision in amblyopes would result in diplopia. Had the binocular circuits not been intact in amblyopia, this is not an unreasonable expectation. Of course, it never occurred for the aforesaid reason. The initial studies demonstrated that binocular vision could be restored and that as a consequence, monocular acuity improved. Laboratory studies on adults benefitted from the fact that the treatment application (1hy/day of dichoptic game playing with suitably contrast imbalanced stimuli) was carefully supervised and that the patients were highly motivated, having suffered from the effects of amblyopia since childhood. What this meant was that there was not only compliance (game play time) but good attentive compliance (stable and controlled attention). At this time, it was not recognized just how important this top-down control was to the successful outcome of the approach. Laboratory studies on children [15,20–23] also showed comparable improvements in binocular vision and monocular acuity after only a few weeks (compared with months for patching) of treatment that were sustained [24] (unlike in 25% of cases after cessation of patching [25]). At-home treatments were also shown to be successful in adults [26] and children [20,23,27–31], however in the latter case considerable time and effort were expended to ensure the treatment was carried out in a way in which the child could allocate their full attention during the 1 hr treatment session.

4. The digital pill - phase 2, unsupervised administration

These promising results from the initial laboratory studies and a more limited number of at-home studies fueled a number of large scale, multi-center clinical trials where the treatment was done at home by children between the ages of 4-13 yrs. These large clinical trials [32–34] did not replicate the laboratory findings and the issue of compliance was highlighted. The treatment protocol of 1 hr per day was not undertaken by a majority of the children for the required time each day (game time as recorded on the digital device) and there was no information on whether the child gave the game their undivided attention even when they were playing. There needed to be an automatic monitoring of what we term ‘attentive compliance’ and a way in which to ensure that it was maintained throughout the 1 hr duration of daily treatment [35]. Parenthetically, a comparable binocular approach using movie viewing [36,37] in a headmounted display does not suffer from the same compliance problems [38] because of its immersive nature. In a collaboration between McGill University and the University of Waterloo, we developed an AI-based approach to monitoring attentive compliance for the gaming approach to treatment and having the ability, via feedback, to either alerting the patient that they are not attending or modifying the game content to re-engage attention, thereby ensuring a regulated and sustained top-down signal to facilitate the bottom-up imbalanced dichoptic stimulation designed to restore the binocular inhibitory circuits to their normal balanced state.

5. The digital pill, phase 3, AI-based monitoring and feedback supervision

Consider a situation in which a clinician is administering a therapy that requires continuous attention and engagement from their patient. The clinician constantly monitors the patient and modifies her communication style or therapeutic approach if she detects that the patient is losing interest. We sought to achieve this combination of monitoring and environmental or therapeutic modification to enable patient re-engagement in a digital form. The solution involves a computer vision AI system that monitors and processes an array of information sources extracted from real-time video of the participant’s face. The information sources including emotional expression, patterns of eye blinks and eye closures, yawning, changes in head angle and gaze direction. By detecting changes across this array of variables, the system can change between three different states: 1) attention-monitoring, 2) feedback to the patient and/or modification of the digital therapy, 3) alerting the caregiver and/or clinician. State 1 occurs when the system variables indicate engagement. A loss of attention triggers a shift to state 2 where the system communicates with the patient and/or modifies the digital therapy content to re-engage the patient. Forward and backward shifts between state 1 and state 2 characterize ‘a smart’ therapeutic system that emulates the behavior of a clinician. State 3 allows the system to seek human support if state 2 behaviors become unsuccessful. As we know more, this monitoring approach would be expected to benefit from a knowledge of the individual clinical features of each patient so that it becomes truly personalized.

6. Conclusion

Ophthalmology has the first digital pill, it has two ingredients. The first is a passive visual stimulation that is designed to rebalance the currently imbalanced binocular inhibitory circuits. It relies on the fact that amblyopia is functionally monocular but structurally binocular. The pathway involved here is essentially feedforward. The second ingredient involves top-down attentional regulation. There is a residual degree of brain plasticity in the circuits controlling ocular balance in the primary visual cortex [39,40] and there is evidence that top-down signals facilitate this [3–5]. This is achieved with an AI-based monitoring/feedback system. With these two ingrediencies, the digital pill should be able to be administered at home with the same efficacy as has been found in the laboratory.

Declaration of Interest

R F Hess and B J Thompson are supported by the Canadian Institute of Health Research, Grant number 156174 and are co-inventors of this AI technology, the patent of which is held by the University of Waterloo and McGill University. The authors have no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.

Reviewer disclosures

One reviewer is an advisor for Luminopia, Inc. The remaining reviewers have no other relevant financial relationships or otherwise to disclose.

Acknowledgments

We would like to acknowledge the work of our co-inventors of this AI application, Dr Otman Basir and Anoop Thazhathumanackal Radhakrishnan from the University of Waterloo.

Additional information

Funding

This paper was funded by the Government of Canada’s Canadian Institutes of Health Research; Institute of Neurosciences, Mental Health and Addiction grant number 156174.