Vision loss is a worrying healthcare concern for the elderly, approximately one in three individuals have some illness impacting their vision by age 65. Age-related macular degeneration (AMD) is the most widespread reason for blindness in developed nations. In Europe, nearly a quarter of individuals aged 60 and more are afflicted by AMD. The ‘dry’ variant is comparatively widespread amongst individuals above the age of 65, and typically causes only marginal sight problems. But, approximately 15% of patients with dry AMD go on to be afflicted with a more serious variant of the illness – exudative AMD, or exAMD – which can have the negative outcome of swift and permanent loss of sight. Luckily, there are treatments that can reduce the speed of subsequent vision loss. Even there exist no preventative treatments currently, these are being looked into through clinical trials. The duration prior to the development of exAMD may thus be indicative of a crucial window to target with regards to clinical interventions: can we forecast which patients will go on to be afflicted with exAMD, and assist in averting sight loss prior to it even happening? 

DeepMind, in their new work, put out in Nature Medicine, worked with Moorfields Eye Hospital and Google Health to go about curating a dataset of imagery of eye retinas, train an AI framework that could forecast the onset of exAMD, and carry out a research to assess their model in contrast with specialist clinicians. They illustrate that their framework is capable of equivalent performance, or better than practitioners, at forecasting whether a patients eye will move on to exAMD within the next six months. Finally, they look into the possible clinical applicability of their system. Their contribution illustrates the prospect of leveraging artificial intelligence in preventative studies for illnesses such as exAMD. 

The Moorfields Eye Hospital AMD Dataset 

They leveraged a dataset of anonymized retinal scans from admitted patients at Moorfields who had been afflicted with exAMD in a single eye, and at risk of getting it in the other eye too. This dataset consists of 2,795 patients throughout seven differing Moorfields sites in and around London, with several age ranges, ethnicities, and genders being indicated. These patients come to the hospital on an ongoing basis to obtain treatment, and go through high resolution three dimensional optical coherence tomography (OCT) imaging of both of their eyes, during every visit. There is usually a delay between when exAMD has onset and when it is detected and treatment is administered. To tackle this, they collaborated with retinal experts to undertake reviews of all scans for every eye and mention the scan when exAMD was first suspected/detected. 

Training a preliminary warning framework for AMD 

The framework consisted of dual deep convolutional networks that leverage as input high-dimensional volumetric eye scans, where every scan is made up of 58 million 3D pixels (voxels) In their prior work, an ongoing collaboration with Google Health, they produced a framework capable of segmenting eye scans into thirteen anatomical classifications. The segmented information was brought together with the raw scan and both were leveraged as inputs to the forecasting model, which received training to predict a patient’s risk of progressing on to exAMD in their other eye over the course of the next six months. 

The advantage of a two phased system is that it provides the artificial intelligence differing perspectives of the eye scans. Anatomical segmentation of the imagery assists the system in learning to model risks on the basis of signs of known anatomical indicators like drusen (miniscule fatty deposits) or reduction of the retinal pigment epithelium (which assists to feed and protect other layers of the retina). Providing the raw eye scans enables the model to go about learning to identify other nuanced changes that could put the patient at risk. Ultimately, the system brings together the data it obtains from these scans to forecast whether and if the eye will move on to exAMD over the course of the next half-a-year. This timeframe was chosen to facilitate the system to forecast at least two follow-up intervals in advance, under the assumption of a maximal follow-up interval of a quarter-of-a-year. 

Clinical specialist benchmark for future forecasting 

It’s critical to setup a benchmark of specialist human performance to contrast how well the framework features performance in contrast to clinical standards. But, forecasting of exAMD is not a typical activity carried out by practitioners, so it’s not obvious if this activity is even doable. To look into this, they carried out a research with six retinal specialists – three optometrists and three ophthalmologists, each one with a minimum of ten years of work experience, to forecast if an eye will go on to be afflicted with exAMD over the course of the next half-a-year. Regardless of the novelty of the activity, the specialists performed better than chance alone, but the activity is tough, and there was considerable variance amongst their evaluations. The framework featured performance as well as, and in some scenarios better than specialists in forecasting whether an eye would go on to be afflicted with exAMD. Simultaneously, demonstrating reduced variability in agreement with every expert, in contrast to experts with one another 

Visualization of illness progression 

It might not be adequate for a framework to merely furnish a prediction, practitioners may also look for data with regards to the anatomic basis for forecasts, which might be of considerable utility for subsequent analysis and interpretation, for instance, for developing research or looking into treatments. An advantage of the system is that it automatically segments every scan into familiar variants of tissue. Extraction of these anatomical and pathological aspects furnishes a structured strategy to visualize the alteration in these tissues over the course of time. The risk scores furnished by the system are in alignment with anatomical modifications over the passage of time, and in conjunction furnish a more nuanced picture of progression into exAMD. 

Foresight instead of hindsight 

It is indeed thrilling times for this domain. The potential to assist practitioners and analysts by generating frameworks that can assist in detection of retinal illnesses in advance and provide clinical insight with regards to their progression. A forecasting framework like this could be leveraged to communicate relevant follow-up intervals to efficiently handle at risk patients. The research adds on to promising preliminary work to produce predictive frameworks for exAMD on the basis of retinal photos and OCT scan results. AI has demonstrated considerable potential in revolutionizing retinal healthcare. 

But there is a lot of work that is pending. This research is not representative of a product that could have implementation in routine work of practitioners. While the model is better at forecasting in comparison to practitioners, there are various other aspects to account for with regards to such systems to be efficient in a clinical scenario. While the framework received training and assessed on a population representative of the biggest eye hospital in Europe, extra work would be needed to assess performance in the context of very differing demographics. One of the latest research pieces looking into the leveraging of a different AI system within a clinical setting illustrated only a few of the sociotechnical problems for such systems, practically speaking. Another tough point of consideration is that any forecasting system will have a specific rate of false positives, that is, when it is discovered that a patient is afflicted with a condition, or forecasted to progress into one, that they don’t currently have. The tradeoff of including an inaccurate AI framework to a preliminary warning loop could be prohibitively expensive to patients who aren’t really at risk, and would require to be thought about carefully in clinical studies of how such frameworks might be leveraged in practical terms. Two system operating points are proposed to go about balancing sensitivity. For instance, for a specificity of 90%, a sensitivity of 34% is accomplished, implying that the system accurately detected progression in one third of scans that did go on to be afflicted with ex-AMD over the course of the next half-a-year. This could detect numerous patients at risk with an accuracy that might be adequate to inform research of novel treatment techniques that might reduce vision loss and enhance patient outcomes.