JAMA+ AI Conversations Real-World Performance of AI in Screening for Diabetic Retinopathy
Apr 18, 2025
Arthur Brant, Chief Resident in Ophthalmology at Stanford, and Sunny Virmani, Group Product Manager at Google, dive into the potential of AI in diabetic retinopathy screening. They discuss a groundbreaking study comparing AI's effectiveness to human evaluations, highlighting both opportunities and disparities in screening access. The conversation addresses challenges such as model drift and accountability in AI use. They stress the importance of ongoing monitoring and innovative solutions to improve patient care, especially in underserved areas.
AI Snips
Chapters
Transcript
Episode notes
AI's Role in Diabetic Screening
- Diabetic retinopathy screening is underperformed globally, especially in India where millions lack routine screening.
- AI tools deployed in large networks can aid by screening patients closer to home, improving early detection.
Challenges of AI Model Drift
- AI models may face performance drift due to changes in patient demographics, camera hardware, and technician skill levels.
- Continuous testing on diverse, real-world data is essential to ensure AI reliability post-deployment.
Proactive Monitoring of AI Systems
- Developers should proactively monitor AI model performance after clinical deployment using real-time data.
- Ensuring patient safety and best care is a shared priority among doctors and AI manufacturers.