In this week’s issue
- Associations have been noted between primary open angle/normal tension glaucoma and dementia but do not seem to be correlated with the APOE genotype, the strongest known genetic risk factor for Alzheimer’s.
- Retinal imaging may be an important player in the future diagnosis of bronchopulmonary dysplasia and pulmonary hypertension in neonates.
- Perimenopause is associated with a clear second peak of uveitis activity after pubertal peak.
- Large language models (LLMs) continue to have difficulty reading retinal fundus photos with 65% accuracy at best across various LLMs.
Spotify Link:
https://creators.spotify.com/pod/show/5xxCPPPznB0yONJbNOIC4I/episode/690v3mr4Eo9o4dkpQXxt1X/wizard
Ophthalmology
Something for “All of Us”: Association between glaucoma and dementia
Cup to disc and cognitive risk. Both considered neurodegenerative disorders, glaucoma and dementia have been inconsistently associated for nearly two decades. Given this fraught association, additional research to tease apart modifying risk factors has been needed. Investigators in Pennsylvania leveraged the NIH “All of Us” dataset to explore association between the two conditions. In this retrospective, longitudinal analysis, more than 45,000 subjects were included over a median follow-up of approximately 6.5 years to examine associations with additional analysis of subtypes and the strongest genetic risk factor for Alzheimer's, the APOE genotype. After matching controls and excluding those with pre-existing dementia, glaucoma was associated with a significantly increased risk of all-cause dementia. When stratified by subtype, primary open-angle glaucoma and normal-tension glaucoma demonstrated particularly strong associations with Alzheimer’s, whereas angle-closure glaucoma did not retain a statistically significant link after adjustment. Finally, elevated dementia risk persisted across common variants (ε2ε3, ε3ε3, ε3ε4), suggesting that the glaucoma-dementia association is not confined to individuals with high genetic risk for Alzheimer’s disease. This study reinforces a growing body of literature positioning glaucoma as more than an isolated optic neuropathy, supporting the concept of shared neurodegenerative or microvascular mechanisms between the eye and brain. While causality cannot be inferred, glaucoma may represent either a marker of systemic neurodegeneration or a condition sharing overlapping pathogenic pathways.
JAMA Ophthalmology
Deep learning diagnoses pulmonary disease using retinal photos
The eyes are a window to cardiopulmonary disease? Bronchopulmonary dysplasia (BPD) and pulmonary hypertension (PH) are related causes of morbidity and mortality in premature infants, yet their diagnosis remains challenging. A BPD diagnosis primarily relies on non-specific demographic data, whereas a PH diagnosis requires invasive procedures, including cardiac catheterization. To address these limitations, this proof-of-concept study developed deep-learning algorithms trained on retinal fundus photos (RFPs) obtained from routine retinopathy of prematurity (ROP) screening at ≤34 weeks of postmenstrual age from across 7 NICUs to identify biomarkers for predicting BPH and PH. Model performance was assessed using the area under the receiver operating characteristic curve (AUC). For BPD, a multimodal approach combining demographic and imaging data improved predictive performance compared with demographics-only (AUC = 0.72) and imaging-only models. For PH, retinal imaging alone was diagnostic, outperforming the demographics-only model and provided no added predictive value in an integrated model. This study demonstrates that retinal images are highly diagnostic of cardiopulmonary disease, supporting their use for earlier, non-invasive risk stratification of BPD and PH in preterm infants.
American Journal of Ophthalmology (AJO)
Reproductive life stages shape noninfectious uveitis activity in women
From the first to final period, your eyes stay “period-ically” involved. Noninfectious uveitis is a major cause of vision loss, with unpredictable flares of intraocular inflammation. Because hormonal states are known to influence immune function, researchers have begun to question whether different stages of a woman’s reproductive life may modify uveitis activity. In this study, investigators retrospectively analyzed data from 700 patients in the AIDA Network Uveitis Registry (2020-2025), comparing uveitis flare frequency and inflammation severity across age-defined reproductive stages (pre-pubertal, early puberty, late puberty, reproductive, perimenopause, and postmenopausal), with males included as controls. Uveitis activity peaked during early puberty for both sexes, with a clear second peak observed in perimenopausal women. Children and early adolescents had more severe anterior segment inflammation and were more likely to require systemic immunosuppressive therapy, while posterior segment inflammation markers remained stable across stages. Reproductive life stage appears to modify uveitis activity, suggesting clinicians should consider hormonal milestones in clinical management and treatment planning.
British Journal of Ophthalmology
How well can large language models interpret retinal imaging?
Teaching artificial intelligence (AI) to read the retina is harder than it looks. As AI tools rapidly evolve, careful evaluation is needed before they are integrated into clinical care. In this study, researchers assessed the performance of eight leading multimodal large language models (LLMs) including GPT-4.5, Gemini 2.0 Pro, Claude 3.7 Sonnet, and others using 100 fundus images representing a broad spectrum of ophthalmic diseases. Across models, rhegmatogenous pathologies were identified most accurately, while myopic maculopathy proved particularly challenging, with mean accuracy below 25%. GPT-4.5 demonstrated the highest overall diagnostic accuracy (65.0%), closely followed by Gemini 2.0 Pro (63.0%). Lower-performing models not only showed reduced accuracy but also frequently declined to provide diagnoses. Overall, these findings suggest that while current LLMs show promise in interpreting fundus images, their limitations highlight the need for further development before routine clinical deployment.
