Causal AI to discover mechanisms and surrogate markers of retinal and brain disease from observational datasets

Background:

Microvascular diseases are important causes of illness and death. Examples include diabetes mellitus, stroke, and vascular dementia. We need to understand how these diseases damage small blood vessels in order to improve diagnosis, prognostication, and predict response to treatments. Measuring microvascular disease within living human organs such as the brain is difficult, but retinal vessels are similar to the brain, and can be measured very precisely. Measuring the impact of systemic disease on retinal vessels, with reference to the impact of disease on brain vessels, could help us predict, diagnose, and understand brain disease mechanisms.

Studies assessing retinal biomarkers of brain disease have often relied on regression models associating a retinal variable (X) with some brain endpoint (Y). A significant X~Y association is taken as evidence that a retinal biomarker might be a “surrogate” of brain disease. However, this is known to be inadequate for assessing surrogate endpoints in clinical trials, and sophisticated methodology has been developed to assess the amount of information shared by a prospective surrogate X, and a true endpoint Y, about a randomized treatment (A)¹.

Retinal biomarkers are typically studied in observational datasets rather than clinical trials. In this setting, prospective longitudinal data are the most useful for understanding disease effects on eye and brain over time. Several large cohorts now combine eye and brain data, and there is a strong motivation to use these to develop retinal biomarkers of brain disease. However, at present we do not have sophisticated statistical approaches to assess retinal biomarkers in longitudinal datasets, analogous to those available for analysing surrogate endpoints in randomised clinical trials. These tools would describe the information shared by change in a retinal biomarker (delta_X), and a true endpoint (delta_Y), about a change in disease exposure over time (delta_A). Adapting surrogate methods to handle longitudinal data about disease trajectory could allow us to understand the causes of microvascular disease in the eye and other organs, and from this, develop prognostic and surrogate markers for clinical use. This is because we can use information about vascular biology to justify key topological elements of structural causal models (relating left eye, right eye, brain and/or other organs). By this we can infer which systemic variables are plausible causes of visible microvascular damage, and assess retinal variables as disease markers using adapted surrogate methodology.

One such longitudinal dataset is the Scottish Collaborative Optometry-Ophthalmology Network e-research (SCONe) – which combines ocular and systemic health data from the general population and is a unique resource to help understand very early causes of eye and systemic health problems in people as they age. Other longitudinal datasets in Edinburgh combine retinal and brain imaging data, such as the Mild Stroke Study 3 (MSS3).

New methodology can be applied to these to understand and prognosticate risk of stroke and vascular dementia, help demonstrate proof of concept, and facilitate external validation of prospective disease markers.