Sally L. Baxter

Background: Interpretation of optical coherence tomography (OCT) imaging requires clinicians to translate complex imaging findings into clinical narrative reports, contributing substantially to documentation burden and physician burnout. Our goal is to develop and evaluate a fine-tuned multimodal large language model (LLM) for automated generation of clinician-style OCT reports to document retinal disease. Methods: We curated 80,102 macula OCT B-scan images paired with corresponding clinician-authored reports from 2,208 patients at a tertiary retina center (2011–2024). A multimodal LLM (Llama 3.2 Vision-Instruct) was fine-tuned to generate clinician-style reports from OCT B-scans. Performance was evaluated based on assessments by three masked expert graders. A set of 250 AI-generated and 250 clinician-authored reports were assessed for predicted source (AI vs. clinician) and three quality metrics (correctness, completeness, and conciseness) using a 5-point Likert scale. Intra- and inter-grader agreements as well as quantitative text similarity metrics (BLEU, ROUGE, METEOR, BERTScore) were also computed. Findings: In masked expert evaluations, graders demonstrated near-chance accuracy (49–55%) in distinguishing AI-generated from clinician-authored reports. Experts graded AI-generated reports as less accurate than clinician-authored reports (3.75 vs. 3.99, p=0.0002), but there were no significant differences between AI and clinician reports for completeness (4.13 vs. 4.23, p=0.08) or conciseness (4.21 vs. 4.13, p=0.14). Experts also graded reports that they perceived as clinician-authored significantly higher regardless of true source (p<0.0001). Inter-grader rater agreement was highest for conciseness (weighted 𝝹 0.52–0.58) and followed by correctness (weighted 𝝹 0.47–0.51) and completeness (weighted 𝝹 0.33–0.49).The AI-generated reports achieved high semantic similarity with clinician-authored reports (BERTScore F1 0.89) despite lower lexical overlap (BLEU 0.09, ROUGE-1 F1 0.36). Interpretation: A fine‑tuned multimodal LLM can generate retinal OCT reports that are indistinguishable from clinician-authored reports and comparable in correctness, completeness and conciseness to those written by clinicians. These findings suggest that LLMs can assist with clinical documentation, reduce reporting burden, and improve workflow efficiency in high-volume ophthalmic imaging clinics.

Background: Rural US communities experience disproportionately high rates of visual disability yet have limited access to ophthalmologists. Teleophthalmology may help address these gaps, but its effectiveness depends on broadband connectivity. The relationship between broadband access and ophthalmologist density has not been well characterized. Objective: The aim of this study is to quantify the association between household broadband access—defined as subscription rates or connection prevalence—and county-level ophthalmologist density and to identify sociodemographic predictors of access. Methods: We conducted an ecological study of all 3141 US counties using 2019 data from the American Community Survey, Area Health Resources File, and National Center for Health Statistics (NCHS). Broadband access was the primary exposure; ophthalmologist count with county population as an offset was the outcome. The primary analysis used negative binomial regression, adjusting for urbanicity, income, education, age, sex, race/ethnicity, unemployment, and insurance status. Sensitivity analyses included population-weighted linear regression and state fixed effects models. County-level heatmaps illustrated geographic patterns. Results: Median household broadband access was 56.6%, ranging from 72.2% in the most urban counties (NCHS category 1) to 49.1% in the most rural (NCHS category 6). In unadjusted negative binomial regression, each 10-percentage-point increase in broadband access was associated with a 68% higher ophthalmologist rate (incidence rate ratio=1.68, 95% CI 1.61‐1.76; P <.001). After adjustment, each 10-percentage-point increase was associated with a 46% higher rate (incidence rate ratio=1.46, 95% CI 1.37‐1.56; P <.001). Sensitivity analyses were consistent with primary analysis. Regions with both low broadband access and zero ophthalmologist density were concentrated in the South, Mountain West region, and Alaska. Conclusions: Broadband access is strongly associated with ophthalmologist availability across US counties, independent of sociodemographic factors. Areas lacking ophthalmologists also tend to lack broadband adoption, creating compounded barriers to both in-person and teleophthalmic care. Efforts to expand broadband may support more equitable access to vision services in underserved regions.

The ability to understand and affect the course of complex, multi-system diseases like diabetes has been limited by a lack of well-designed, high-quality and large multimodal datasets. The NIH Bridge2AI AI-READI project ( aireadi.org ) aims to address this shortfall by generating an AI-ready dataset to support AI discoveries in type 2 diabetes mellitus (T2DM). This manual of procedures provides a detailed description of the AI-READI protocol.

This study aims to identify new genetic loci associated with primary open-angle glaucoma (POAG) and explore shared genetic risk factors across African, European, and Admixed American/Latino populations. Genome-wide Association Study (GWAS) utilizing data from the All of Us Research Program. The study included 374,254 participants, with 4,305 individuals diagnosed with POAG and 369,949 controls. Participants were categorized by ancestry: European, African, and Admixed American/Latino. We used short-read sequencing data and applied strict quality control measures (MAF > 0.01, INFO > 0.8). GWAS were conducted for each ancestry group using a logistic mixed model, adjusting for age, sex, and the top 11 principal components. A fixed-effect meta-analysis combined the results across ancestries. Genome-wide significance was set at p < 5 × 10 ^− 8 . The primary outcome measures were the identification of genetic loci associated with POAG, and the analysis of transcription factors linked to these loci in relevant tissues. In the European cohort, we identified four novel loci associated with POAG, linked to the TUT4 , RYK , MOXD1 , and UBAP2 genes, as well as the previously known TMCO1 locus. In the African cohort, we found five new loci, including TSPAN17 , SLC16A7 , LOC100506869 , LINC02388 , and LOC107984606 . For the Admixed American/Latino cohort, we identified GATA5 , FAM135B , and LINC00871 genes as novel loci. Our analysis identified three novel loci in individuals of European ancestry, mapped to the genes TUT4 , RYK , and MOXD1 . We identified 56 genome-wide significant variants, including six putative novel loci, and found that most ancestry-specific signals replicated in the cross-ancestry meta-analysis, with the exception of several attenuated associations in the smaller Admixed American/Latino cohort. These findings indicate that the genetic determinants contributing to POAG may differ across populations, underscoring the importance of accounting for population-specific genetic architectures in the study of complex traits. Given the substantial variation in POAG prevalence among ancestries, it is plausible that certain genetic variants exert ancestry-specific effects. Consequently, conducting ancestry-stratified GWAS is essential for elucidating these unique genetic contributions.