Adult patients with diabetes (n = 3745) seen at Johns Hopkins Medicine primary care sites were referred to the Wilmer Eye Institute either based on a primary care provider referral or autonomous AI diagnostic result (referral was made after a positive or non-diagnostic result). An inverse-probability-weighted regression, which incorporated propensity score matching on social determinants of health and relevant clinical variables, showed that implementation of an autonomous AI-assisted diabetic screening program in a primary care clinic was associated with increased presentation to eye care specialist care by African-Americans (p = 0.02). This is significant because African-Americans have traditionally been less likely to undergo annual screening exams and more likely to present with more severe forms of diabetic retinopathy (DR). The results suggest a potential association between office-based, AI-assisted DR screening and improved downstream ophthalmic access for African-American patients. However, given that the analysis was exploratory, this association should be interpreted cautiously and further validated.

PURPOSE. Genetic studies implicate the matrix metalloproteinase-9 (MMP9) locus in neovascular age-related macular degeneration (nvAMD) risk but genotype-phenotype associations of MMP9 with nvAMD are lacking. This study aimed to investigate the influence of MMP9 genotype and T-cell subset frequency on structural and functional treatment outcomes in nvAMD. METHODS. We reanalyzed single-cell RNA sequencing data and used ELISA and flow cytometry in THP-1-derived monocytes to measure immune cell expression of MMP9 within human choroids. In a clinical nvAMD cohort of 38 patients, we genotyped the nvAMD risk single nucleotide polymorphism (SNP; rs4810482) and quantified retinal fluid using deep-learning-based optical coherence tomography (OCT) image analysis. On a subset of nine patients, we performed high-dimensional immunophenotyping. RESULTS. MMP9 is predominantly expressed in mature THP-1-derived dendritic-like cells (ELISA, P = 0.009; flow cytometry, P = 0.001). Patients with the TC genotype of MMP9 exhibited greater disease severity compared to CC or TT genotypes with a significantly higher total retinal fluid volume (P = 0.009). Immunophenotyping revealed that higher proportions of circulating CD8+ effector memory T cells re-expressing CD45RA (TEMRA) were associated with increased residual subretinal fluid (P = 0.03), indicating persistent disease activity. CONCLUSIONS. MMP9 genotype affects structural and functional outcomes in patients with nvAMD. Along with the observed systemic immune dysregulation, these findings support the role of a MMP9-dendritic-T-cell axis in nvAMD immunopathogenesis and highlight this as a potential therapeutic target.

The population structure of an inbred population of 781 people on Norfolk Island in the Pacific, 318 of which are descendants of the original Mutineers of the Bounty, is analyzed phenotypically using shape from stereo retinal fundus photographs. Three-dimensional optic nerve head (ONH) shape is reconstructed from stereo pairs by a multi-scale stereo matching algorithm. Using deep neural network, the shape of ONH, which is under genetic control, is decomposed into a set of hierarchical features through self-taught learning. Features captured at different levels are selected according to their discriminant power in identifying the two populations. The prediction accuracy is evaluated with stratified cross validation. Given the selected feature set, individuals are grouped into k hierarchical clusters and cluster membership fractions are determined for k=2,3,4,5,6,7. Population structure analysis on the basis of phenotypes through image analysis allows heritability and linkage analysis, including founder effects from English and Polynesian ancestors, potentially leading to new genetic risk factors for glaucoma and other ONH-related eye diseases.

Importance:Diabetic retinal neurodegeneration precedes vascular changes associated with diabetic retinal disease (DRD). Studies in adults with type 1 diabetes (T1D) show there is retinal layer thinning with DRD, yet there are limited data in youth with T1D. Objectives:To determine if retinal layer thickness changes on optical coherence tomography (OCT) imaging were associated with glycemic outcomes and DRD in youth. Design, Setting, and Participants:This prospective cohort study was conducted at an academic pediatric diabetes center among youth with T1D aged 9 to 21 years participating in the ACCESS2 (AI for Pediatric Diabetic Eye Exams Study 2) study. Participants were enrolled and data were collected July 11, 2022, and April 30, 2025. Data analysis was performed from June 2025 through October 2025. Exposure:OCT imaging. Main Outcomes and Measures:The primary outcome was macular OCT volumes, which were segmented by the Topcon Maestro camera software and reviewed by the Wisconsin Reading Center for 3 neuroretinal layers: (1) retinal nerve fiber layer (RNFL) thickness, (2) ganglion cell and inner plexiform layer (GCL+IPL) thickness, and (3) GCL+IPL+RNFL thickness, as well as total retinal thickness. Layer thicknesses were analyzed for associations with glycemic outcomes and DRD and for potential covariates. Results:A total of 294 youth with T1D (n = 578 eyes), among whom mean (SD) age was 15.8 (2.8) years, 153 participants (52.0%) were female, and 108 participants (36.7%) had public insurance, were included. Participants had a median (IQR) duration of diabetes of 7.0 (4.6-10.1) years and a median (IQR) hemoglobin A1c (HbA1c) of 8.5% (7.5%-9.9%); 210 participants (71.4%) used an insulin pump. Of the total 578 eyes, 65 eyes (11.2%) had mild DRD and 10 eyes (1.73%) had moderate DRD. In adjusted analyses, moderate DRD vs no DRD was associated with RNFL thickness of -1.2 µm (95% CI, -2.9 to 0.5; P = .20), GCL+IPL thickness of -1.2 µm (95% CI, -2.8 to 0.4; P = .19), and outer retinal layer thickness of -0.8 µm (95% CI, -3.9 to 2.2; P = .80). In multivariable models, GCL+IPL and outer retinal layer thickness were associated with HbA1c (β = -0.39; 95% CI, -0.78 to -0.01; P = .04; and β = -0.81; 95% CI, -1.49 to -0.12; P = .02, respectively). Conclusions and Relevance:In this prospective cohort study, neuroretinal layer thinning was observed in youth with T1D without clinically apparent DRD and was associated with higher HbA1c. These findings support elucidating the development of diabetic retinal neurodegeneration and its potential role as a biomarker of retinal vascular disease in youth.

PURPOSE:Artificial intelligence (AI)-based screening models hold promise for identifying individuals with undiagnosed age-related macular degeneration (AMD) in nonspecialist settings. A standardized reference framework for image labeling is needed to enable consistent training, validation, and deployment of AI-based screening algorithms. The goal of the present study was to establish expert consensus on an image-based reference standard for labeling AMD. DESIGN:Modified Delphi consensus study. PARTICIPANTS:Fellowship-trained retina specialists, ophthalmologists, AI specialists, and imaging specialists. METHODS:A prespecified Delphi process was conducted using structured surveys. Over 2 rounds, panelists assessed opinions on existing reference standards, including the Age-Related Eye Disease Study scale and Beckman scale, as well as imaging methods such as color, OCT, and autofluorescence. The surveys also evaluated imaging features of AMD, including drusen, pseudodrusen, and pigment changes, as well as referral criteria. Consensus was defined using a 9-point Likert scale, with predefined statistical thresholds for agreement. MAIN OUTCOME MEASURES:Agreement on key elements of a reference standard. RESULTS:Consensus was reached on adopting the Beckman classification as the level 1 reference standard (median score, 8; agreement). OCT use for identifying key AMD features, including drusen, geographic atrophy (GA), and choroidal neovascularization, also reached consensus (median scores, 8.5-9; agreement). Pigment change detection did not reach consensus (median, 7.5; uncertain), and screening age thresholds showed nonconsensus (median, 8; uncertain). Referral thresholds reached consensus, including urgent referral for neovascular AMD and nonurgent referral for GA and intermediate AMD (median, 9; agreement). CONCLUSIONS:This study defined a consensus-based reference standard for labeling AMD from images for AI-based screening. These recommendations are intended to support consistent AI model development and evaluation, while remaining distinct from clinical practice guidelines. FINANCIAL DISCLOSURE(S):Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Teleophthalmology and artificial intelligence (AI) -based retinal screening have emerged as scalable approaches to improve detection of diabetic retinal disease (DRD), yet their economic value within U.S. health systems is not well defined. We developed a 5-year Markov model to evaluate the cost-effectiveness of these alternative strategies against screening by an eye care professional (ECP) across two scenarios: a small primary care network and a large integrated health system with Willingness-to-Pay (WTP) ranges, respectively, of $400–1500 and $800–3500 per patient over 5 years. AI-based strategies yield 3 times more screenings completed, 3.6–3.8 times more true positives, and 7.5–8.0 times more patients who initiate treatment than ECP. Teleophthalmology strategies are moderately effective. The most cost-effective strategy for each scenario depends on health system scale and WTP. For a primary care network with $400 WTP, teleretinal via handheld camera is preferred for sites screening ≤ 2282 patients; handheld AI otherwise. At $1500 WTP, teleretinal via handheld camera is preferred for sites with ≤ 884 patients, handheld AI for 885–1476, and stationary AI otherwise. In an integrated health system at $800 WTP, ECP dominates across volumes of 250–14,000 patients. At $3500 WTP, teleretinal via handheld is preferred for ≤ 1393 patients per site; stationary AI otherwise. AI-based strategies, particularly handheld AI, are cost-effective with larger patient volumes and higher WTP thresholds. Their cost-effectiveness is driven by real-time diagnostic feedback that enables earlier detection and reduces costs associated with managing advanced forms of DRD.

The "Diabetic Retinal Disease (DRD) Cure Accelerator," a joint initiative by the Mary Tyler Moore Vision Initiative and the Collaborative Community on Ophthalmic Innovation, aims to modernize clinically meaningful staging systems and endpoints for DRD. During a June 2025 workshop involving over 100 international experts, participants emphasized the urgent need for validated structural and functional endpoints grounded in the retinal neurovascular unit. Current DRD staging limitations were highlighted alongside the importance of early biomarkers, including retinal nonperfusion, inner retinal thinning, and disorganization of retinal layers. Functional priorities included contrast sensitivity, electrophysiology, and performance-based mobility parameters that reflect real-world visual impairment. Beyond clinical staging, this approach facilitates precision phenotyping to isolate specific molecular pathogenic pathways driving disease in individual patients. Mechanistic granularity provides a foundation for discovery science, enabling targeted drug re-purposing and the development of curative therapeutics. Key next steps include prospective multicenter studies, harmonized protocols, and reference standards and endpoints for AI based on patient relevant outcomes. By integrating scientific, clinical, and patient-centered perspectives, this Accelerator seeks to establish an international consensus on robust endpoints. Ultimately, this precision-based approach aims to transform DRD diagnosis and treatment, accelerating the access of patients worldwide to therapies that reduce diabetes-related vision loss.

Autonomous artificial intelligence (AI) for pediatric diabetic retinal disease (DRD) screening has demonstrated safety, effectiveness, and the potential to enhance health equity and clinician productivity. We examined the cost-effectiveness of an autonomous AI strategy versus a traditional eye care provider (ECP) strategy during the initial year of implementation from a health system perspective. The incremental cost-effectiveness ratio (ICER) was the main outcome measure. Compared to the ECP strategy, the base-case analysis shows that the AI strategy results in an additional cost of $242 per patient screened to a cost saving of $140 per patient screened, depending on health system size and patient volume. Notably, the AI screening strategy breaks even and demonstrates cost savings when a pediatric endocrine site screens 241 or more patients annually. Autonomous AI-based screening consistently results in more patients screened with greater cost savings in most health system scenarios.

Sample size calculations for power analysis are critical for clinical research and trial design, yet their complexity and reliance on statistical expertise create barriers for many researchers. We introduce PowerGPT, an AI-powered system integrating large language models (LLMs) with statistical engines to automate test selection and sample size estimation in trial design. In a randomized trial to evaluate its effectiveness, PowerGPT significantly improved task completion rates (99.3

Introduction and Objective: Population Achieved Sensitivity (PAS) assumes that the primary goal of a diagnostic process is to identify patients who can benefit from intervention. We used PAS to analyze adoption bias — characterized by inequitable adoption of AI technologies — for an autonomous AI for diabetic eye exams. Methods: We compared two autonomous AI algorithms paired with a desktop fundus camera and a handheld retina camera. Sensitivity for the preregistered clinical trials were reported (NCT02963441 and NCT05808699) and the PAS formula was derived from an ethical framework as presented in npj Digital Medicine - Nature. Access was estimated from the numbers of desktop fundus cameras and handheld retina cameras deployed in US primary care settings. The heatmap presents PAS values for any level of access (0-100% penetrance) and sensitivity ≥ 60%. Results: PAS increases with increasing access and/or sensitivity. The top-right quadrant shows the highest PAS value. Though the handheld retina camera has slightly lower sensitivity (82%) compared with the desktop fundus camera (87%), its greater potential for adoption (estimated at 10X) shows increased detection of diabetic retinal disease in real-world settings. Conclusion: Mitigating adoption bias requires balancing accuracy and access, quantifiable through PAS. This example illustrates how to achieve this balance, empowering clinicians to focus on both diagnostic accuracy and access. R. Channa: None. C. Joyce: Consultant; Digital Diagnostics. M.D. Abràmoff: Stock/Shareholder; Digital Diagnostics. Board Member; Digital Diagnostics. Consultant; Digital Diagnostics.

ImportanceSafe integration of artificial intelligence (AI) into clinical settings often requires randomized clinical trials (RCT) to compare AI efficacy with conventional care. Diabetic retinopathy (DR) screening is at the forefront of clinical AI applications, marked by the first US Food and Drug Administration (FDA) De Novo authorization for an autonomous AI for such use.ObjectiveTo determine the generalizability of the 7 ethical research principles for clinical trials endorsed by the National Institute of Health (NIH), and identify ethical concerns unique to clinical trials of AI.Design, Setting, and ParticipantsThis qualitative study included semistructured interviews conducted with 11 investigators engaged in the design and implementation of clinical trials of AI for DR screening from November 11, 2022, to February 20, 2023. The study was a collaboration with the ACCESS (AI for Children’s Diabetic Eye Exams) trial, the first clinical trial of autonomous AI in pediatrics. Participant recruitment initially utilized purposeful sampling, and later expanded with snowball sampling. Study methodology for analysis combined a deductive approach to explore investigators’ perspectives of the 7 ethical principles for clinical research endorsed by the NIH and an inductive approach to uncover the broader ethical considerations implementing clinical trials of AI within care delivery.ResultsA total of 11 participants (mean [SD] age, 47.5 [12.0] years; 7 male [64%], 4 female [36%]; 3 Asian [27%], 8 White [73%]) were included, with diverse expertise in ethics, ophthalmology, translational medicine, biostatistics, and AI development. Key themes revealed several ethical challenges unique to clinical trials of AI. These themes included difficulties in measuring social value, establishing scientific validity, ensuring fair participant selection, evaluating risk-benefit ratios across various patient subgroups, and addressing the complexities inherent in the data use terms of informed consent.Conclusions and RelevanceThis qualitative study identified practical ethical challenges that investigators need to consider and negotiate when conducting AI clinical trials, exemplified by the DR screening use-case. These considerations call for further guidance on where to focus empirical and normative ethical efforts to best support conduct clinical trials of AI and minimize unintended harm to trial participants.

Diabetic eye disease (DED) is a leading cause of blindness in the world. Annual DED testing is recommended for adults with diabetes, but adherence to this guideline has historically been low. In 2020, Johns Hopkins Medicine (JHM) began deploying autonomous AI for DED testing. In this study, we aimed to determine whether autonomous AI implementation was associated with increased adherence to annual DED testing, and how this differed across patient populations. JHM primary care sites were categorized as "non-AI" (no autonomous AI deployment) or "AI-switched" (autonomous AI deployment by 2021). We conducted a propensity score weighting analysis to compare change in adherence rates from 2019 to 2021 between non-AI and AI-switched sites. Our study included all adult patients with diabetes (>17,000) managed within JHM and has three major findings. First, AI-switched sites experienced a 7.6 percentage point greater increase in DED testing than non-AI sites from 2019 to 2021 (p < 0.001). Second, the adherence rate for Black/African Americans increased by 12.2 percentage points within AI-switched sites but decreased by 0.6% points within non-AI sites (p < 0.001), suggesting that autonomous AI deployment improved access to retinal evaluation for historically disadvantaged populations. Third, autonomous AI is associated with improved health equity, e.g. the adherence rate gap between Asian Americans and Black/African Americans shrank from 15.6% in 2019 to 3.5% in 2021. In summary, our results from real-world deployment in a large integrated healthcare system suggest that autonomous AI is associated with improvement in overall DED testing adherence, patient access, and health equity.

Topic: The goal of this review was to summarize the current level of evidence on biomarkers to quantify diabetic retinal neurodegeneration (DRN) and diabetic macular edema (DME). Clinical relevance: With advances in retinal diagnostics, we have more data on patients with diabetes than ever before. However, the staging system for diabetic retinal disease is still based only on color fundus photographs and we do not have clear guidelines on how to incorporate data from the relatively newer modalities into clinical practice. Methods: In this review, we use a Delphi process with experts to identify the most promising modalities to identify DRN and DME. These included microperimetry, full-field flash electroretinogram, spectral-domain OCT, adaptive optics, and OCT angiography. We then used a previously published method of determining the evidence level to complete detailed evidence grids for each modality. Results: Our results showed that among the modalities evaluated, the level of evidence to quantify DRN and DME was highest for OCT (level 1) and lowest for adaptive optics (level 4). Conclusion: For most of the modalities evaluated, prospective studies are needed to elucidate their role in the management and outcomes of diabetic retinal diseases. Financial Disclosure(s): Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Where adopted, Autonomous artificial Intelligence (AI) for Diabetic Retinal Disease (DRD) resolves longstanding racial, ethnic, and socioeconomic disparities, but AI adoption bias persists. This preregistered trial determined sensitivity and specificity of a previously FDA authorized AI, improved to compensate for lower contrast and smaller imaged area of a widely adopted, lower cost, handheld fundus camera (RetinaVue700, Baxter Healthcare, Deerfield, IL) to identify DRD in participants with diabetes without known DRD, in primary care. In 626 participants (1252 eyes) 50.8% male, 45.7% Hispanic, 17.3% Black, DRD prevalence was 29.0%, all prespecified non-inferiority endpoints were met and no racial, ethnic or sex bias was identified, against a Wisconsin Reading Center level I prognostic standard using widefield stereoscopic photography and macular Optical Coherence Tomography. Results suggest this improved autonomous AI system can mitigate AI adoption bias, while preserving safety and efficacy, potentially contributing to rapid scaling of health access equity. ClinicalTrials.gov NCT05808699 (3/29/2023).

Purpose Patients with non-proliferative macular telangiectasia type 2 (MacTel) have ganglion cell layer (GCL) and nerve fibre layer (NFL) loss, but it is unclear whether the thinning is progressive. We quantified the change in retinal layer thickness over time in MacTel with and without diabetes. Methods In this retrospective, multicentre, comparative case series, subjects with MacTel with at least two optical coherence tomographic (OCT) scans separated by >9 months OCTs were segmented using the Iowa Reference Algorithms. Mean NFL and GCL thickness was computed across the total area of the early treatment diabetic retinopathy study grid and for the inner temporal region to determine the rate of thinning over time. Mixed effects models were fit to each layer and region to determine retinal thinning for each sublayer over time. Results 115 patients with MacTel were included; 57 patients (50%) had diabetes and 21 (18%) had a history of carbonic anhydrase inhibitor (CAI) treatment. MacTel patients with and without diabetes had similar rates of thinning. In patients without diabetes and untreated with CAIs, the temporal parafoveal NFL thinned at a rate of -0.25 +/- 0.09 mu m/year (95% CI [-0.42 to -0.09]; p=0.003). The GCL in subfield 4 thinned faster in the eyes treated with CAI (-1.23 +/- 0.21 mu m/year; 95% CI [-1.64 to -0.82]) than in untreated eyes (-0.19 +/- 0.16; 95% CI [-0.50, 0.11]; p<0.001), an effect also seen for the inner nuclear layer. Progressive outer retinal thinning was observed. Conclusions Patients with MacTel sustain progressive inner retinal neurodegeneration similar to those with diabetes without diabetic retinopathy. Further research is needed to understand the consequences of retinal thinning in MacTel.

The purpose of this study was to evaluate the feasibility of a generalizable deep-learning (DL) based system with no a priori knowledge of fundus photographs to generate monocular depth map information about optic disc structures from this imaging modality. Images of 30 stereo pairs of fundus photographs centered on the optic disc of 30 subjects were analyzed with this DL system to generate monocular depth maps using zero-shot cross-dataset transfer. These maps were registered onto reference standard depth maps derived from Optical Coherence Tomography. Accuracy of the DL system was assessed by the root of mean squared error (RMSE) between the estimate and reference standard. 47% of the total images from the dataset were successfully processed, with mean RMSE of 0.081. Our findings demonstrate that single image, monocular depth estimation with a generalizable DL system using zero-shot cross-dataset transfer applied to retinal color fundus photographs is feasible and has potential. Received: 24 July 2024 | Revised: 30 September 2024 | Accepted: 10 October 2024 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement The data that support the findings of this study are openly available in Iowa Carver College of Medicine at https://medicine.uiowa.edu/eye/inspire-datasets, reference number [15]; in Github at https://github.com/isl-org/MiDaS, reference number [18]; in PyTorch at https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html, reference number [19]; in ImageJ Docs at https://imagej.net/software/imagej2/, reference number [23]. Author Contribution Statement Rony Gelman: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing - original draft, Writing - review & editing, Visualization, Supervision, Project administration. Michael D. Abràmoff: Writing - original draft, Writing - review & editing, Visualization, Supervision, Project administration.

Diabetic retinopathy can be prevented with screening and early detection. We hypothesized that autonomous artificial intelligence (AI) diabetic eye exams at the point-of-care would increase diabetic eye exam completion rates in a racially and ethnically diverse youth population. AI for Children’s diabetiC Eye ExamS (NCT05131451) is a parallel randomized controlled trial that randomized youth (ages 8-21 years) with type 1 and type 2 diabetes to intervention (autonomous artificial intelligence diabetic eye exam at the point of care), or control (scripted eye care provider referral and education) in an academic pediatric diabetes center. The primary outcome was diabetic eye exam completion rate within 6 months. The secondary outcome was the proportion of participants who completed follow-through with an eye care provider if deemed appropriate. Diabetic eye exam completion rate was significantly higher (100%, 95%CI: 95.5%, 100%) in the intervention group ( n = 81) than the control group ( n = 83) (22%, 95%CI: 14.2%, 32.4%)(p < 0.001). In the intervention arm, 25/81 participants had an abnormal result, of whom 64% (16/25) completed follow-through with an eye care provider, compared to 22% in the control arm (p < 0.001). Autonomous AI increases diabetic eye exam completion rates in youth with diabetes.

We examined which subgroups of patients benefit the most from deployment of autonomous artificial intelligence (AI) for diabetic eye disease (DED) testing at primary care clinics through improved patient access to ophthalmic care. Patients (n = 3,745) were referred to ophthalmology either via standard of care (primary care provider placed a referral) or AI (referral was made after a positive or non-diagnostic autonomous AI result). Both groups presented with good vision (median best-corrected visual acuity BCVA of worse-seeing eye was Snellen 20/25), without significant difference in the presenting BCVA between both groups. BCVA was not associated with the referral pathway in multivariable regression analysis. However, patients from the AI referral pathway were more likely to be Black (p < 0.001) and have hypertension (p = 0.001), suggesting that deployment of autonomous AI is associated with improved ophthalmic access for patients with a higher baseline risk for poor DED outcome before vision loss has occurred.