Chinese team advances AI vision model for eye care

An international research collaboration effort led by Chinese scientists has unveiled a vision-based foundation model that promises to transform eye care worldwide.

It has demonstrated, through an international clinical trial, how artificial intelligence (AI) will soon be able to assist doctors in settings ranging from primary care clinics to specialist centers.

Published in Nature Medicine, the study, led by Tsinghua University and Shanghai Jiao Tong University, shared details concerning EyeFM, an AI system trained on 14.5 million ocular images and paired clinical texts from global, multi-ethnic datasets. The system was created by the universities in collaboration with international partners.

Forty-four ophthalmologists from China, India, Malaysia, Denmark, Equatorial Guinea and the United States, working in primary and specialty care clinics, validated its efficacy, highlighting EyeFM's utility as a clinical co-pilot.

The team also added a "doctor feedback" feature, a quick loop that makes the model suitable for use in both low-resource clinics and high-complexity specialty hospitals.

Earlier, similar AI tools tended to learned from just one type of data, so they could not process varied information in the way a doctor could. They usually checked only on old records and were not tested in advance in various treatment scenarios and facilities. Hardly any were put through randomized trials, and few studies examined how doctors and AI could collaborate.

In a single-center, double-masked trial involving 668 high-risk patients in China, 16 ophthalmologists were randomized to use either EyeFM or standard care alone during retinal-disease screening. In a primary analysis, EyeFM support raised diagnostic accuracy to 92.2 percent, compared with 75.4 percent in the control group.

The study provides high-level evidence that large medical AI models can power both primary and specialty care, offering a ready-to-reuse playbook for integrating AI into an effective, everyday clinical practice tool.