Justin C. Muste, Timothy T. Xu, Erik Massenzio, Collin J. Richards
Wills Eye Hospital
Retinal Aging Clocks Based on Deep Learning Predictions of Age From Retinal Images are Associated With Systemic Health and Evolutionarily Conserved – Glenn Yiu, MD, PhD
Dr. Yiu trained AI models to predict age from images of humans and non-human primates. Models performed well when predicting age within the species they were trained on, but performance declined substantially when applied across species. These results suggested AI age-prediction models capture both conserved aging features and species-specific visual characteristics, but age clocks are not fully transferable between species.
Association of Deep Learning Derived Retinal Age With Cognitive Impairment – Benjamin Kim, MD
Dr. Kim used an AI-derived retinal vascular age model trained on multiple cohorts using the AI-READI dataset. The Montreal Cognitive Assessment was used as a surrogate measure of cognitive impairment. Retinal age showed a stronger association with cognitive performance than chronological age. Similar findings were observed in a separate cohort with Alzheimer disease and frontotemporal dementia. Overall, these findings suggest that AI-derived retinal age from a routine fundus photograph may provide a readily accessible biomarker associated with cognitive impairment and neurodegenerative disease, particularly through retinal vascular changes.
Pathology-Aware Retinal Aging Clocks From OCT and Fundus Imaging Predict Systemic Comorbidity Burden – Chase Ludwig, MD
Dr. Ludwig evaluated whether retinal/OCT-derived biological age provides clinically meaningful information beyond chronological age. Biological age increased with disease and diagnosis, suggesting that retinal structural features may capture aspects of systemic or biological aging, and biological-age models generally outperformed chronological age. Importantly, models trained only on healthy individuals generalized poorly to diseased populations, while incorporating diagnostic information during training improved biological-age prediction across all disease groups. OCT-derived structural measurements were among the most informative retinal features, although distinguishing normal aging from disease-related changes remains challenging.
OCT-F: AI-Enabled Retinal Function From Structural OCT – Roger Goldberg, MD, MBA
Dr. Goldberg tackled 68-point microperimetry. Despite its unique ability to assess retinal function, microperimetry is time-consuming and provides relatively sparse sampling of the macula. In contrast, OCT functional mapping (OCT-F) uses deep learning to combine OCT volume scans with microperimetry data to infer retinal sensitivity across the entire macula without requiring microperimetry testing. Applying this to OAKS/DERBY, OCT-F also showed about a one-third reduction in the risk of developing an absolute scotoma while maintaining some inferred sensitivity. OCT-F may offer a more comprehensive and practical functional biomarker of GA progression, with preservation of central retinal sensitivity closely associated with preservation of vision.
AI-Assisted OCT Biomarkers Quantification of Diabetic Macular Edema Untreated Eyes Identifies Four Different Phenotypes – Edoardo Midena, MD, PhD
Dr. Midena discussed a precision-medicine approach to DME. He showed an AI-based algorithm developed specifically for DME was used to analyze more than 2,000 eyes from the Italian Retinal Society cohort, including >400 untreated eyes. The analysis identified at least four distinct DME phenotypes, with the more severe phenotypes comprising roughly half of untreated eyes and showing greater neuroinflammatory features including subretinal fluid, hyperreflective inflammatory signs, and outer-retinal disruption. AI-based OCT phenotyping may allow DME to be classified according to underlying pathophysiology, potentially enabling treatment to be individualized rather than applying a uniform standard-of-care strategy to all patients.
AI-Enabled Noninvasive Detection and Quantification of Retinal Inflammation in Uveitis – Parisa Emami, MD, MPH
Dr. Emani leveraged an AI-based approach to noninvasive monitoring of retinal inflammation and vascular leakage in noninfectious uveitis, addressing the limitations of fluorescein angiography (FA). Researchers created a dataset of patients who underwent same-day OCT, fundus photography, and FA, using FA as the reference standard to train and validate models. The model performed well internally (AUC 0.98 for OCT and 0.90 for fundus photographs) and retained performance in an external University of Nebraska cohort (AUC 0.83 and 0.70, respectively). They then developed the ODYSSEY Index, which correlated strongly with FA findings (r = 0.90), and longitudinal changes in the index tracked improvement or worsening on FA. An associated app can analyze multiple OCT scans within seconds, potentially providing a noninvasive, quantitative biomarker to assess and monitor retinal inflammation without repeated fluorescein angiography.
Artificial Intelligence in GA: Learnings From the Pegcetacoplan Clinical Program – Murtaza Adam, MD
Dr. Adam applied AI-based OCT and FAF analysis to measure GA progression. In addition to a more detailed measure of progression, novel markers like EZ loss and RPE loss were shown as potentially earlier markers of disease progression and treatment response. Pegcetacoplan was associated with approximately a 50% reduction in EZ loss over two years, with greater preservation in patients treated monthly or every other month, and treatment effects were particularly pronounced in eyes with faster baseline progression.
One-Week Response to Anti-Vascular Endothelial Growth Factor Therapy as a Marker of Long-Term Therapeutic Response – Hasenin Al-khersan, MD
Dr. Al-khersan evaluated if earlier response to anti-VEGF therapy could predict long-term visual outcomes in patients with DME, nAMD, and RVO. One-week visual acuity improvement correlated significantly with 24-month visual outcomes across all three disease groups, with the correlation strengthening over the first three months. In contrast, OCT anatomy alone was found to be a less reliable early predictor of eventual visual acuity.
Explainable Deep Learning-Guided Fluorescein Angiography Morphometry in CADASIL Mouse Models – Leo Kim, MD, PhD
Dr. Kim applied AI models to vascular changes in mouse models of small vessel disease. Even when FA appeared largely normal on visual inspection, the AI model distinguished subtle vascular abnormalities. Using Grad-CAM and occlusion-sensitivity saliency mapping, the investigators further refined the AI to identify minor changes in vessels such as beading. The findings suggest that AI-assisted image analysis may help uncover otherwise subtle retinal vascular biomarkers.
Automated Pipeline for Pre-Screening and Treatment Determination in Ophthalmology Clinical Trials – Vlad Diaconita, MD, FRCSC
Dr. Diaconita discussed the role of AI to improve the efficiency of retinal clinical trials. Two major applications were discussed: pre-screening, where AI combines retinal imaging with EMR/clinical information to identify patients likely to meet trial eligibility criteria and reduce costly screen failures, and trial monitoring, where AI analyzes imaging and clinical data to flag potential protocol deviations or treatment-readiness issues for review by coordinators and investigators. The broader goal is a workflow in which AI performs repetitive quantitative assessments and integrates multimodal data, while the research coordinator or PI makes the final clinical decision. Overall, AI could help identify the right patients before enrollment and ensure protocol adherence during the trial.
