Ocular Biomarkers: Useful Incidental Findings by Deep Learning Algorithms in Retinal Photographs

Abstract

Abstract Background/Objectives: Ocular biomarkers can provide immediate and non-invasive indications of ocular and systemic health but are underutilised due to the paucity and inequitable distribution of eyecare professionals. Deep learning analysis of colour fundus photographs has task shifting potential to efficiently differentiate ocular biomarkers, as well as providing earlier diagnosis, additional reach via telehealth, and ultimately improving population health. The study aimed to explore the clinical implications arising from deep learning detection of non-target retinal biomarkers in colour fundus photographs. Subjects/Methods: Patients referred for treatment-resistant hypertension were imaged between 2016 and 2022 at a specialty clinic in Perth, Australia. The same 45° colour retinal photograph selected for each of the 433 participants imaged was processed by three deep learning algorithms. All positive results for diabetic retinopathy in non-diabetic participants were graded by two expert retinal specialists. Results: A total of 29 non-diabetic participants were flagged as positive for diabetic retinopathy by deep learning algorithms. Of these, 28 (96.6%) had clinically significant non-target retinal pathology likely to benefit from early intervention. The algorithms with more target diseases captured less incidental disease. All three algorithms demonstrated a correlation between false positive diabetic retinopathy results and severity of hypertensive retinopathy. Conclusions: The findings indicate that existing deep learning models can identify additional pathologies likely to benefit from early intervention within an at-risk, hypertensive cohort, and have potential for immediate clinical application in other populations. The findings also support a pivotal pathway toward autonomous comprehensive screening.