A deep learning algorithm based on fundus photographs to measure retinal vascular parameters and their additional value beyond the CAIDE risk score for predicting 14-year dementia risk
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Retinal photography is a valuable non-invasive tool for assessing the nature of vessel changes. It is of interest whether retinal vascular parameters can improve the ability to predict dementia risk of the widely used Cardiovascular Risk Factors, Aging, and Incidence of Dementia (CAIDE) model.
Methods
A fully automated artificial intelligence algorithm providing measures of seven meaningful parameters of the retinal vascular was developed and evaluated. Applying this algorithm to analyze the fundus images collected by the UK Biobank (UKB) study and the Beijing Research on Ageing and Vessel (BRAVE) study, we further explore the associations between retinal vascular parameters with arteriosclerosis and cognitive function across two countries. Finally, we fitted a published algorithm-estimated CAIDE model based on fundus images developed by our research group to the UKB study and evaluated the additional value of retinal vascular parameters beyond the algorithm-estimated CAIDE model for predicting 14-year dementia risk.
Findings
In the UKB cohort (n=35,838), the algorithm-estimated CAIDE model achieved an area under the curve (AUC) of 0.697 (95% confidence interval [CI]: 0.678-0.717) for 14-year all-cause dementia, which was comparable to that of the actual CAIDE model (AUC=0.683, 95% confidence interval: 0.663-0.703, p=0.272). Furthermore, adding retinal vascular parameters to the algorithm-estimated CAIDE model increased the AUC statistically significantly from 0.692 to 0.711 (p<0.001) in the derivation set (n=38,384), and from 0.682 to 0.706 (p=0.018) in the internal validation (n=9594) for all-cause dementia.
Interpretation
The integration of automatically extracted retinal vascular parameters into the algorithm-estimated CAIDE model improves the predictive ability of 14-year dementia risk. Compared to the original CAIDE model, the retinal vascular parameters-enhanced algorithm-estimated CAIDE model may provide a more accurate dementia risk assessment with just a single fundus photograph.
Funding
The Capital’s Funds for Health Improvement and Research, and the National Natural Science Foundation of China.
