Deep Learning-Derived Retinal Vascular Morphology and Incident Treatment for Sight-Threatening Diabetic Retinopathy.

Objective: Retinal vascular morphology (RVM) features could be useful biomarkers for risk stratification for incident sight-threatening diabetic retinopathy (STDR). We evaluated whether deep learning-derived macula-centered RVM, extracted at scale, are associated with incident treatment for STDR including proliferative diabetic retinopathy (PDR) and clinically significant macula edema (CSME). Design: Retrospective longitudinal cohort study (2000-2022). Participants: We included 20 855 patients (33 861 eyes) with gradable, treatment-naïve eyes with a macula-centered fundus image at baseline. Methods: Arterial and venous RVM features of complexity, density, caliber, and tortuosity were extracted using an automated deep learning tool (AutoMorph). Multivariable Cox proportional hazards models were performed to quantify the association between RVM and incident treatment for STDR, adjusting for age, sex, ethnicity, baseline diabetic retinopathy (DR) and diabetic macula edema (DME) grade, and intereye correlation using a cluster term. Main Outcome Measures: Incident treatment for STDR including PDR or CSME. Results: Over a mean 1.74 (standard deviation 2.15) years of follow-up, 3446 eyes (10.2%) received treatment for STDR (1387 PDR; 2059 CSME). Higher arterial complexity and density were strongly protective for incident STDR treatment (arterial complexity Q4 vs. Q1 hazard ratio [HR] 0.52 [95% confidence interval 0.45-0.60]; arterial density HR 0.47 [0.41-0.54]). Larger venous caliber was associated with an increased risk of incident STDR treatment (HR 2.53 [2.23-2.86]), while increasing arterial caliber was associated with a reduced risk of incident PDR treatment (HR 0.64 [0.53-0.77]). There were weak associations between increasing arterial and venous tortuosity and an increased risk of incident treatment for PDR and CSME, respectively (arterial tortuosity HR 1.11 [95% confidence interval 1.01-1.22]; venous tortuosity HR 1.16 [1.03-1.31]). Retinal vascular morphology feature associations were generally stronger for incident treatment for PDR than CSME. Associations between RVM and incident treatment for STDR remained consistent after a sensitivity analysis which only included patients with no or mild DR and no DME at baseline. Conclusions: Macula-centered RVM features show clinically plausible associations with incident treatment for STDR. Retinal vascular morphology features extracted using an automated deep learning tool are potential prognostic biomarkers that could be useful for DR risk prediction models alongside other retinal imaging biomarkers and clinical risk factors. Financial Disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.