Updates

  • Reliable, reproducible research is critical to improving patient care and accelerating scientific discovery. Excited to share that our work has been recognized with The National Institutes of Health Replication Prize for advancing reproducible clinical research. Through LATCH (LLM-Assisted Testing of Clinical Hypotheses), our team developed a framework that helps make EHR-based research more transparent, reproducible and scalable using AI-assisted workflows. In testing, the framework successfully replicated findings from previously published studies in just minutes while also enabling new avenues for discovery using real-world clinical data. Grateful to collaborate with Cecilia S. Lee, MD, MS, Yu Jiang, Yuka Kihara, Nayoon Gim, In Gim, Marian Blazes, M.D., and Yue Wu on this effort. https://lnkd.in/g3Z94K53
  • OCTCube-M: A 3D multimodal optical coherence tomography foundation model for retinal and systemic diseases with cross-cohort and cross-device validation, The eye can provide important clues about a person’s overall health. We developed OCTCube-M, a next-generation AI model trained on millions of retinal images that analyzes the retina in 3D and combines information from multiple imaging technologies. The system can detect eye diseases, identify signs of conditions such as diabetes and hypertension, and predict how certain retinal diseases may progress over time. By extracting more information from routine eye scans, OCTCube-M highlights the potential of retinal imaging as a window into both eye and systemic health.
  • Machine learning derived retinal pigment score from ophthalmic imaging shows ethnicity is not biology - People are often grouped by ethnicity when researchers study diversity in medical imaging datasets, but biology does not fit neatly into these categories. In our study, we developed a machine learning tool that measures retinal pigmentation directly from eye photographs, creating a continuous Retinal Pigment Score (RPS). Analysis of large international datasets showed substantial overlap in retinal pigmentation across ethnic groups, demonstrating that ethnicity is an imperfect proxy for biological variation. This new approach could help researchers build more representative datasets and develop AI systems that work more reliably across diverse populations.
  • Improving Residency Matching through Computational Optimization, showed that the process used to match medical students with residency programs could be improved using modern computational optimization. Using historical ophthalmology residency and fellowship match data, we compared the long-used Gale-Shapley matching algorithm with a new “residency optimizer” designed to improve outcomes for both applicants and programs. The optimized approach matched more applicants to one of their top choices, reduced unfilled positions, and performed better when rank lists were shortened or when couples were matching together. Our findings suggest that updating residency matching algorithms could make the process more efficient, fairer, and less burdensome for applicants and training programs.
  • Multi-omic spatial effects on high-resolution AI-derived retinal thickness}, Nature Communications: A routine eye scan contains far more information than previously recognized. In this study, researchers used AI to measure retinal thickness at tens of thousands of locations across the retina, creating one of the most detailed maps of retinal structure ever generated. By combining these measurements with genetic, molecular, and health data from large population studies, we discovered links between retinal changes and a wide range of systemic diseases. The work highlights the retina as a window into whole-body health and shows how AI can uncover biological signals that may help enable earlier disease detection and improved risk assessment.
  • Convolutional neural network-based classification of glaucoma using optic radiation tissue properties - The retina may provide a window into how quickly a person is aging. We used artificial intelligence to estimate biological age directly from retinal photographs and found that people with an “older-looking” retina were more likely to experience age-related health problems. Because retinal imaging is fast, noninvasive, and widely available, this approach could provide a new way to measure healthy aging and identify individuals at higher risk of future disease. The work demonstrates how routine eye images can reveal information about health far beyond the eye itself.
  • Estimating uncertainty of geographic atrophy segmentations with Bayesian deep learning - Artificial intelligence can help doctors identify and measure retinal damage, but clinicians also need to know when an AI model may be uncertain. To address this issue, we developed an AI system that automatically identifies geographic atrophy lesions in retinal scans while simultaneously estimating its own confidence in each prediction. The approach highlights areas where the model is less certain, helping users interpret results and identify cases that may require closer review. This added transparency could improve trust in AI-assisted eye care and support the safe adoption of automated image analysis tools in clinical practice.
  • Training deep learning models to work on multiple devices by unsupervised cross domain learning with no additional annotations: Medical AI systems are often trained using images from a single camera or scanner, which can limit their usefulness in other clinics that use different equipment. We developed a technique that allows AI models trained on one OCT device to work accurately on images from other devices without requiring additional expert annotations. The approach helps overcome one of the biggest barriers to deploying AI in healthcare: differences in imaging hardware across institutions. By making AI tools more adaptable and robust, this work moves retinal image analysis closer to widespread clinical use.