Oculomics: The New Diagnostic Revolution

Pioneering scientists at UCL and Moorfields Eye Hospital have engineered a new AI system which uses simple scans of the eye to uncover a wide range of health issues, such as heart problems and even Parkinson’s disease, with greater accuracy than previous models. This is a milestone for the emerging field of oculomics, which aims to aid diagnosis with the help of ophthalmological biomarkers.

More than windows to the soul

The retinas are the only part of the human body through which the network of capillaries (the smallest blood vessels) can be observed directly and in detail. This means that the state of an individual’s retinas can be utilised as a brief CV of their overall health, providing a powerful tool to efficiently diagnose systemic problems. The optic nerve and inner retinal layers offer insights of the central nervous system’s condition. Furthermore, the structure and integrity of the retinal vascular network often reflects the state of blood vessels in other organs, such as the heart and kidneys.

Clinical retinal imaging

Retinal imaging is an important diagnostic tool, which involves taking photos of the fundus (inner, back surface of the eye). The 2 main methods are:

● Colour fundus photography. Specialised fundus cameras are used to take pictures of the inside of the eye, displaying the pattern of blood vessels.

● Optical Computerised Tomography. Offering cross-sectional views of the layers of the retina and measuring how thick each of them are, this method is helpful for diagnosing conditions like diabetes related macular edema (when blood vessels leak fluid into the retina).

Images produced by these methods require specialist expertise to interpret, which is a problem for developing countries with a lack of ophthalmological expertise.

Pushing frontiers with AI

In a recent paper, researchers have introduced RETFound, an open-source AI tool with the ability to diagnose and predict multiple diseases with a single scan of the retina (Zhou et al., 2023). Researchers had to overcome a major difficulty in order to develop RETFound - the cost and difficulty of obtaining a large quantity of well-classified eye scans (eg, as ‘normal’ or ‘not normal’) to train the AI. This problem was avoided by using a model called Self-Supervised Learning (SSL), which involved a pre-training phase where the AI was shown 1.6 million retinal scans, in order for it to learn the normal features of a retina. The AI could then be shown a small number of labelled images, allowing it to learn how certain diseases affected the characteristics of the retina.

Finally, RETFound was tested on the accuracy of its classification. Detection was best for ocular diseases, such as diabetic retinopathy (characterised by abnormal capillary growth and bleeding, and potentially leading to blindness). On a scale where 0.5 represents a model that performs no better than a random prediction and 1 represents a model with perfect accuracy, RETFound scored up to 0.943 (depending on the data set used) for diabetic retinopathy. Although accuracy was lower for systemic diseases, RETFound still outperformed other machine-learning tools; for heart failure, stroke and Parkinson’s disease, it scored 0.794, 0.754 and 0.669 respectively.

A further advantage of this pioneering tool is that it addresses one of the main shortcomings of other AI models by accurately labelling scans from ethnically diverse populations, and those with rare diseases. Eventually, tools like this may be utilised by areas with a shortage of trained doctors to improve diagnostic accuracy.

Beyond hospitals

Personalised medicine has been developed significantly in recent years, thanks in part to the advancement of artificial intelligence’s capabilities and the proliferation of Internet of Things technologies. One notable example is smart mirrors, innovative products which are being designed to, among other things, notice deviations in a person’s normal appearance that may be symptoms of a disease (Miotto et al., 2018).

Currently, smart mirror projects have rather limited diagnostic capabilities. One such project, the Wize Mirror, developed by the European research project SEMEOTICONS, can detect mental health issues, such as stress and anxiety, as well as weight gain, which may suggest an increase in cardio-metabolic risk. This is achieved by detecting changes in factors such as eyelid motion and mouth activity (Andreu et al., 2016). However, such products are in an early developmental stage, and will likely take decades to reach the accuracy of medical imaging devices.

In the future, smart mirror devices may employ oculomics, providing more biomarkers to allow AI programs to better inform users of potential health issues. Many further challenges must be overcome before this can be achieved, including the current need for an additional lens to take accurate retinal scans.

However, the future of oculomics looks bright, and machine-learning systems, like RETFound, are likely to continue their breathtaking pace of improvement, paving the way for a revolution in medical diagnostics.