Deep Learning Derived Quantification of Reticular Pseudodrusen for Survival Modeling of Progression to Age-Related Macular Degeneration in a Large Real-World Cohort

Abstract

Purpose: Reticular pseudodrusen (RPD) are an important phenotype in age-related macular degeneration (AMD), but their natural history remains poorly characterized. Most prior studies were limited by small cohorts, short follow-up, and/or binary assessments. In this work, we used an automated deep-learning (DL) algorithm to detect and quantify RPD, enabling cross-sectional characterization of their burden and distribution as well as longitudinal evaluation of their association with progression to late AMD over 10 years. Methods: Retrospective cohort study including patients 50 years with non-advanced AMD (as determined by an International Classification of Diseases-10 [ICD-10] code) seen between 2015-2025 at the Massachusetts Eye and Ear. Inclusion required availability of a Spectralis Optical Coherence Tomography (OCT) scan (Heidelberg Engineering) within 7 days of first AMD diagnosis, no other vitreoretinal diseases, and 3 months of follow-up. A validated DL model identified and segmented RPD, allowing quantification of presence (defined as 5 instances of RPD in more than 1 OCT B-scan) as well as area and volume (square-root and cube-root transformed, respectively, for analyses) on baseline OCT scans. A second validated DL model identified and segmented classic drusen, another well-established phenotype in AMD, to quantify their volume (cube-root transformed) on baseline OCT scans and adjust for baseline AMD severity during survival modeling. Cox proportional hazards (CoxPH) models accounting for age, smoking, and Age-Related Eye Disease Study (AREDS) medication use were used to assess associations between baseline RPD metrics and incident geographic atrophy (GA) or choroidal neovascularization (CNV). The likelihood ratio test (LRT) was applied to nested CoxPH models to evaluate the incremental fraction of new information provided by each RPD metric. The preprocessed RPD area and volume were both categorized by quartile and transformed via restricted cubic splines (RCS) to account for non-linear effects. Results: We included 3846 eyes (n=2254 patients, mean age 74.6±9.1 years), among which 46% (n=1756) had RPD at baseline. Over a median of 2.78 years (IQR: [1.21,5.37]) of follow-up, 7.2% (n=278) progressed to GA over a median duration of 1.72 years (IQR: [1.02, 3.24]) and 10.2% (n=394) progressed to CNV over a median duration of 1.73 years (IQR: [0.87, 3.18]). Stratified by baseline RPD status, progression to GA occurred in 12.8% (n=225) of eyes with RPD vs 2.5% (n=53) without, and progression to CNV occurred in 14.9% (n=261) vs 6.4% (n=133), respectively. RPD presence was found to correlate with the presence of classic drusen; however, RPD was more prevalent superior to the fovea whereas classic drusen were clustered around the central foveal region. The LRT revealed that inclusion of each RPD metric individually significantly improved prediction (p.05) of both outcomes and GA alone; however, only RPD presence and RPD cube-rooted volume significantly improved CNV prediction (p.05). Baseline RPD presence was associated with a higher risk of progression to CNV (HR=1.53, p.05) and GA (HR=2.18, p.01). Categorization by quartile demonstrated that greater RPD area (Q4 - both outcomes: HR=2.04, p.01; GA: HR=3.99, p.05; CNV: HR=1.66, p>0.05) and volume (Q4 - both outcomes: HR=1.95, p.05; GA: HR=3.82, p.05; CNV: HR=1.58, p>0.05) were also associated with an increased risk for both outcomes combined and GA, but not for CNV alone. RCS confirmed a nonlinear relationship between both RPD area and volume with all outcome events and graphically demonstrated that the most informative increase occurred within the first two quartiles. After calculating RPD presence, area, and volume, as stratified by spatial location in the retina according to the Early Treatment Diabetic Retinopathy Study (ETDRS) grid, only RPD presence in a few key locations was significantly associated with progression to GA or both outcomes combined. Conclusion: In the largest study to date quantifying RPD metrics, their presence increased the risk of progression to late AMD even after adjusting for other well-established demographic and clinical risk factors, such as classic drusen. Though RPD presence offered the most consistent signal, RPD area and volume also improved predictive performance of progression to study outcomes. This supports the prognostic value for deep learning and segmentation and quantification of RPD in AMD risk stratification and treatment optimization.