Predicting aortic stenosis from colour fundus photographs from the AlzEye Study using a deep learning model.

Abstract

Purpose : Signs of systemic health conditions, such as cardiovascular and neurodegenerative disorders, can be discerned in retinal images. This study aimed to assess the effectiveness of colour fundus photographs (CFP) in the analysis and prediction of aortic stenosis by employing a deep learning model.

Methods : Data for this study was collected from the Moorfields AlzEye project, which is a retrospective cohort study linking ophthalmic records of 353,157 patients who visited the Moorfields Eye Hospital from 2008 to 2018 with their systemic health records from nationwide hospital admissions. Patients diagnosed with aortic stenosis, identified using ICD codes I35, were included, with 6,325 patients having undergone retinal imaging. The dataset was divided into training, validation, and test groups at a 55%:15%:30% ratio. Colour fundus photographs underwent preprocessing with AutoMorph. The model development process involved fine-tuning our foundational model, RetFound, with the training dataset, followed by evaluation on the test dataset. Bootstrap methods were employed to ascertain a 95% confidence interval (CI).

Results : The model demonstrated an AUROC of 0.717 (95% CI: 0.709-0.725) and an AUPR of 0.710 (95% CI: 0.701-0.720). It achieved an F1-score of 0.650 (95% CI: 0.611-0.688) for the prediction of aortic stenosis, with a sensitivity rate of 0.668 (95% CI: 0.541-0.794).

Conclusions : Our findings show that retinal images are indicative of systemic health conditions and that our deep learning model can predict aortic stenosis presence from colour fundus photographs. This suggests a promising avenue for non-invasive, cost-effective, and efficient prediction and management of aortic stenosis using retinal imaging.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.