Researchers turned a simple eye scan into a rapid AI-powered screening tool that detects multiple chronic diseases in seconds, potentially expanding access to early diagnosis worldwide.
Study: AI framework for multidisease detection via retinal imaging. Image credit: muratart/Shutterstock.com
In a recent study published in Nature Medicine, researchers developed Reti-Pioneer, an artificial intelligence (AI)-based system that analyzes retinal images for rapid, non-invasive, and low-cost detection of chronic diseases. The system could identify type 2 diabetes, hypertension, gout, and osteoporosis from retinal images in approximately 30 seconds, highlighting its potential to screen systemic diseases through a simple eye scan.
AI Eye Scans Target Chronic Disease Screening
Metabolic and endocrine diseases are becoming increasingly common worldwide. These conditions place considerable strain on healthcare systems, warranting the development of scalable screening methods. Current screening approaches primarily analyze blood samples. Although effective, these methods are invasive, expensive, and time-consuming to repeat frequently for long-term, population-wide health monitoring.
Prior studies have shown that retinal changes may represent conditions such as diabetes, hypertension, or thyroid disorders. Most systems, however, can detect only one disease at a time, require high-quality images, and have not been verified in different population subgroups.
Reti-Pioneer Combines Three AI Vision Models
In the present study, researchers developed Reti-Pioneer to detect multiple diseases using retinal scans. They used 107,730 color fundus images for development. These images were obtained from 53,865 participants in the UK Biobank (UKBB) and multiple hospital-based cohorts across China. They were captured under various imaging conditions in diverse clinical settings.
The team combined three pre-trained vision foundation models with quality-aware modules, enabling the system to retain and analyze lower-quality retinal images rather than exclude them. The pre-trained models were Vision Mamba, RETFound, and Swin Transformer. The team trained and fine-tuned the model to screen for six metabolic and endocrine disorders. These included hypertension, type 2 diabetes, hyperlipidemia, osteoporosis, thyroid disease, and gout.
The researchers then validated the framework using external datasets from resource-limited and high-resource regions in China. They also included the Singapore Epidemiology of Eye Diseases (SEED) study cohort for validation. In total, the validation dataset comprised 23,232 retinal images obtained from 11,616 individuals.
Furthermore, the researchers examined long-term predictive performance by testing the model on a longitudinal UKBB subset with up to 15 years of follow-up. They generated saliency maps to identify retinal regions most strongly associated with disease prediction. They then investigated whether the retinal patterns were associated with specific blood proteins and genetic risk markers linked to those diseases.
Subsequently, the investigators conducted reader studies in which ophthalmologists reviewed retinal images with and without AI assistance to compare diagnostic performance. They also conducted a prospective “silent” trial, allowing researchers to evaluate its accuracy and speed in routine practice without influencing clinical decision-making. Lastly, they performed a pilot study with 606 participants to assess whether the system could fit smoothly into clinical workflows.
AI System Detects Six Diseases Rapidly
Reti-Pioneer detected all six diseases, with internal-test area under the receiver-operating characteristic curve (AUROC) values ranging from 0.699 to 0.833. The system showed consistent performance across external validation cohorts in Singapore and China, covering both well-resourced and resource-constrained settings, although performance varied by disease and cohort. Reti-Pioneer was more effective at identifying individuals with multiple diseases than models that rely solely on clinical features.
The system could not only detect existing diseases, but also estimate an individual’s risk of developing them years later. Reti-Pioneer attained 10-year AUC values ranging from 0.662 to 0.813 for disease onset prediction, with the strongest predictive ability for osteoporosis and gout. Interestingly, the system’s ‘quality-aware’ component enabled analysis of retinal scans even when image quality was suboptimal, highlighting its applicability in busy primary care centers, which often produce blurry images.
In the prospective silent trial, Reti-Pioneer produced screening results in approximately 30 seconds per person, considerably faster than conventional laboratory workflows, which required nearly eight hours. The system achieved an image acquisition rate of nearly 99 %, and the AI framework successfully processed all acquired images during inference.
The model also performed better than the Finnish Diabetes Risk Score (FINDRISC), a commonly used questionnaire-based screening tool for diabetes. Its higher AUROC values compared with FINDRISC (0.776 vs. 0.565) indicate stronger diagnostic accuracy. A negative predictive value (NPV) of 0.966 indicated strong accuracy in identifying individuals unlikely to have type 2 diabetes.
Reti-Pioneer also improved diagnostic accuracy among clinicians by 10-20 %. Over 80 % of clinicians and patients reported that the system was easy to use, efficient, and acceptable for clinical practice.
AI Retinal Screening Shows Real-World Promise
The findings suggest that the Reti-Pioneer system could eventually support routine medical screening to simultaneously detect endocrine and metabolic diseases from retinal images in clinical practice. The speed, affordability, consistency across diverse populations, and diagnostic accuracy, even using low-quality images, make the system especially useful in real-world settings.
However, the authors noted that the system’s current diagnostic and predictive accuracy remains below the threshold required for broad clinical adoption. Larger studies are needed to validate the findings. Further investigations must determine whether AI-guided retinal screening can improve healthcare efficiency, accessibility, and early disease detection for at-risk populations and whether it can improve long-term clinical outcomes and resource allocation in routine care.
Download your PDF copy by clicking here.
New Genetic Pathways Linked to Cardiometabolic Disease Identified