Insights into the future of eye care

July 29, 2020
Prof. Alastair Denniston
Prof. Alastair Denniston

Volume 16, Issue 6

A UK health data research hub containing vast numbers of fundus and OCT images opens the door to improved screening and diagnoses of eye diseases and even ‘oculomic’ services for systemic health.

The year is 2025, and I have just had my annual health check … at my local optometrist’s. Vision and eye health are fine given my age and I am now on my way to my family doctor as my oculomic profile flagged some concerns that need to be checked.1 This is the second year I have opted into this free screening service as part of my annual eye check and it now feels normal.

The Oculomic Health Evaluation software provides a systemic health profile based on my fundus photographs and optical coherence tomography (OCT) scans. Within five minutes of sitting down for my scans I have been alerted to the fact that I am mildly anaemic with an estimated haemoglobin of 10.2g/L. On the plus side, my retina-estimated biochemical profile is fine and my total cholesterol assessed at less than 200g/dL.

The biggest surprise was that my estimated systolic blood pressure (BP) is up at 140mmHg. I quickly checked against my smart watch, which agrees. The diastolic reading is always less accurate from the fundal images so they still do not tend to report that, and there are easier ways of measuring your BP!

The rest of the report includes a raft of things that they have not found – no early signs of dementia or other neurodegenerative diseases; no atrial fibrillation; estimated normal cardiac and renal function; no signs of systemic inflammation; and no metabolic disturbances.

From where I am sitting in 2020, this all seems possible – and likely. We are increasingly recognising the power of the information encoded within our eyes, captured on both two-dimensional fundus images and three-dimensional OCT scans.

Eyes as windows to our health

The range of systemic variables we can predict will only increase but let us reflect on where we are now. In one of my favourite papers of all time, Poplin and colleagues developed an algorithm that could predict age, smoking status, body mass index, diabetes status and systolic and diastolic blood pressure from colour fundus images.2

External validation of this algorithm showed impressive accuracy for most of these variables, although systolic BP was strikingly more accurate than diastolic BP. A similar approach has recently shown good performance in predicting haemoglobin levels,3 and I would expect that we will see an increasing number of associations for other blood and tissue constituents being published over the next few years.

But for me, the most exciting thing about the paper by Poplin et al was none of these things. Yes, the performance was impressive, but it was reasonable to predict that a machine learning approach could detect the subtle pathophysiological changes that we would expect to be present with increasing age: higher BP and so on. To a large extent this is doing the same as humans, but better.

What really blew me away was that it could predict the biological sex of an individual from their fundus photograph with a very high accuracy (AUC of 0.97). This was totally unexpected.

It may not be as useful as some of the other metrics, but it tells us something more profound: that there are more aspects of our health and biological status that are encoded within our eyes than we ever dreamed of. We just need to look for them.

These kinds of discoveries from routine ophthalmic imaging are only possible because of three things. Firstly, advances in machine learning with specific techniques such as deep learning that can manage the high-dimensional data contained within images such as fundus images and OCT scans.

Secondly, the computational power that enables these to operate on the scale of datasets required. Thirdly, and most fundamentally, we need the datasets of sufficient quality and scale to train and test them on. And this is, in fact, our greatest limiting factor to innovation in this space: the availability of data.

Where big data comes in

A health data research hub focused on eye health called aims to address this challenge. Launched in September 2019, INSIGHT4 forms part of the Health Data Research UK (HDRUK)5 initiative, which aims to “unite the UK’s health data to enable discoveries that improve people’s lives.” It is a non-profit organisation supported by ten of the UK’s major health and science funders.

INSIGHT seeks to make health data available to researchers and innovators so that together we can better understand diseases and find ways to prevent, treat and cure them. The programme is designed to enable discovery and innovation in the field of oculomics, through enabling safe and appropriate access to deidentified data at scale. We think of it as a ‘biobank for ocular imaging’, with the ability to link to other relevant clinical data.

The datasets for INSIGHT initially came from the two founding NHS partners – Moorfields Eye Hospital NHS Foundation Trust and University Hospitals Birmingham NHS Foundation Trust. The nature of the populations across London and Birmingham makes this data exceptionally diverse in terms of ethnicity and socio-economic status, and with the NHS being available to everyone in the UK, using such data means some of the selection biases that may occur in other health systems are minimised.

Our existing resource is around ten million retinal images, but this is increasing by thousands of scans across multiple modalities every day across our existing network of hospitals. Building from our initial hospitals, we are looking forward to welcoming other NHS trusts and community eye health providers who have expressed their support for this initiative and their desire to scale INSIGHT across the UK.

Not only does this increase the scale of data—something which is particularly important for rarer conditions—but also ensures that all parts of the UK are represented when it comes to testing hypotheses or the performance of AI systems prior to implementation at a national level, for example in diabetic retinopathy screening or community monitoring of age-related macular degeneration.

Logistics

We are working with NHS patients and the wider public to establish exactly how the system will operate with two areas of central importance: data access and the sustainability framework. Our charity partner Action Against AMD (AAAMD) has been leading on the development of a data trust advisory board, with strong patient and public representation that will form the framework for evaluation, assess incoming applications and will provide the ongoing scrutiny of data access through the programme.

Along with HDRUK nationally, we are also engaging with patients and the public to agree a sustainability model that is considered fair and returns value to the NHS. INSIGHT and other health data hubs are funded for 3 years by the UK government and various industry partners, but in order to continue, it will need to charge for its services in order to be sustainable and return value to the NHS.

In addition to the NHS Partners, and AAAMD, the health hub also includes both an academic partner (University of Birmingham) and two industry partners—Roche and Google Health—who are contributing their expertise and experience in a range of areas including pharmaceutical development, real world evaluation, technology data architecture and data security.

Principles of participation for all partners of INSIGHT and for HDRUK ensure that the datasets created are not exclusive to partners, and that there is no preferential access – all applications for data access will undergo the same rigorous and independent level of scrutiny against patient benefit and data safety. Additionally, all data accessed through the system will be deidentified, and applications will be evaluated against the ‘five safes’ of data security and international standards.

The aim, ultimately, is to improve people’s lives through facilitating and accelerating research that makes a difference in the real world. Building databases are costly in terms of time and resources, and they depend on the support and engagement of our patient communities.

However, we see this as an investment to enable others to do amazing things. We hope that the programme will help others to discover new insights into sight-threatening disease, to automate and improve the quality of screening and diagnosis, and even open up the possibility of ‘oculomic’ screening services for systemic health.

References

  1. Wagner SK, Fu DJ, Faes L, et al. Insights into systemic disease through retinal image-based oculomics. Translational Vision Science & Technology. 2020;9. doi:https://doi.org/10.1167/tvst.9.2.6
  2. Poplin R, Varadarajan AV, Blumer K, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nature Biomedical Engineering. 2018;2:158-164.
  3. Mitani A, Huang A, Venugopalan S, et al. Detection of anaemia from retinal fundus images via deep learning. Nature Biomedical Engineering. 2020;4:18-27.
  4. www.insight.hdrhub.org
  5. www.hdruk.ac.uk

Prof. Alastair Denniston

E: a.denniston@bham.ac.uk

Prof. Denniston is a consultant ophthalmologist at University Hospitals Birmingham NHS Foundation Trust, UK, honorary consultant ophthalmologist at Moorfields Eye Hospital, honorary professor at the University of Birmingham and director of INSIGHT, the health data research hub for eye health. He has no financial disclosures.

download issueDownload Issue : Ophthalmology Times Europe July/August 2020