A new study shows how retinal images could help flag diabetes, hypertension, gout, osteoporosis, thyroid disease, and high cholesterol, but researchers caution that broader clinical use will require larger trials and stronger evidence of patient benefit.
Study: AI framework for multidisease detection via retinal imaging. Image Credit: spaxiax / Shutterstock
In a recent study published in the journal Nature Medicine, researchers developed a retinal imaging framework for multi-disease detection.
Retinal Imaging and Oculomics Background
The rise in metabolic and endocrine diseases worldwide poses a significant challenge to healthcare systems and highlights the need for early detection methods. Current screening approaches heavily rely on blood biomarkers, whose collection involves patient discomfort, high costs, and logistical challenges, limiting the feasibility for frequent longitudinal screening.
Oculomics, the use of ocular imaging and artificial intelligence (AI) for systemic health monitoring, has been promising in identifying preclinical cardiovascular, neurodegenerative, and renal conditions. However, oculomics research depends on high-quality images and is largely limited to single-disease frameworks, limitations that recent medical foundation models may address.
Reti-Pioneer Framework and Study Design
In the present study, researchers introduced Reti-Pioneer, a multi-task framework based on retinal imaging for multi-disease detection. First, a multimodal dataset of color fundus photographs (CFPs) was curated, comprising 107,730 CFPs from more than 53,000 individuals in the United Kingdom Biobank (UKB) and Chinese hospital registries. A multimodal learning model was developed that incorporated CFPs of heterogeneous quality with structured clinical metadata.
Further, this architecture employed an ensemble of large, pretrained vision foundation models, such as RETFound, Swin Transformer, and Vision Mamba, to leverage their capabilities. In the internal dataset, Reti-Pioneer showed an area under the receiver operating characteristic curve (AUROC) of 0.83 for gout and type 2 diabetes mellitus (T2DM), 0.79 for osteoporosis, 0.74 for hyperlipidemia and hypertension, and 0.7 for thyroid disease.
Reti-Pioneer Performance Across Cohorts
Reti-Pioneer was also externally tested across Chinese datasets representing resource-limited and high-resource settings. In a dataset derived from regions with limited healthcare resources, Reti-Pioneer showed moderate-to-strong discriminative ability, with AUROCs of 0.9 for osteoporosis, 0.82 for thyroid disease and T2DM, 0.81 for hypertension, 0.73 for gout, and 0.63 for hyperlipidemia. In the analysis combining datasets from high- and limited-resource settings, the model maintained robust performance.
Furthermore, Reti-Pioneer was validated in a multi-ethnic Singaporean cohort, achieving AUROCs of 0.69 for T2DM, 0.75 for hypertension, and 0.62 for hyperlipidemia. Stratified analyses by ethnicity produced consistent AUROCs; for hypertension, the AUROC was 0.73, 0.75, and 0.77 for Indian, Chinese, and Malay individuals, respectively.
The corresponding AUROCs for T2DM were 0.65, 0.67, and 0.69, respectively; these data suggest broadly consistent performance across ethnic groups, although the authors noted that further work is needed to improve representativeness and generalizability.
Next, the researchers evaluated Reti-Pioneer on a withheld subset of the UKB to examine its predictive performance for six diseases at five- and 10-year intervals.
In this analysis of 15,704 participants without preexisting disease, Reti-Pioneer achieved AUROCs of 0.76 and 0.74 for five- and 10-year incident T2DM, respectively. The corresponding AUROCs were 0.76 and 0.72 for hypertension and 0.75 and 0.74 for hyperlipidemia, respectively.
The authors noted that longer-term prediction was more challenging than cross-sectional screening and that future studies should use formal time-to-event methods.
Clinical Validation and Real-World Use of Reti-Pioneer
Further, Reti-Pioneer’s diagnostic performance was benchmarked against that of retinal specialists. Specialists interpreted CFPs from 200 patients per disease, and re-evaluated the same images one or two weeks later, assisted by Reti-Pioneer. Retinal specialists had a mean accuracy of 88% for T2DM, 70% for thyroid disease, and 79% for gout when assisted by Reti-Pioneer, compared to 71%, 63%, and 51%, respectively, without assistance.
Next, a prospective silent trial was conducted in a primary care setting. Fundus images were acquired in real time from the point of care for immediate testing with Reti-Pioneer. This workflow showed greater throughput efficiency, with significantly shorter time from image acquisition to report generation than physical examination or laboratory reports.
Finally, a prospective real-world study was conducted at physical examination and community health centers with 606 participants.
All participants undergoing physical examinations were offered testing with Reti-Pioneer. Reti-Pioneer analyzed participants’ retinal images and clinical metadata. It demonstrated greater discriminative ability for T2DM (AUROC: 0.78) than the Finnish Diabetes Risk Score. AUROCs for hypertension, hyperlipidemia, gout, osteoporosis, and thyroid disease were 0.84, 0.7, 0.8, 0.88, and 0.65, respectively.
Satisfaction surveys completed by clinicians and participants indicated high usability and acceptance of Reti-Pioneer. Specifically, 80% of participants were very satisfied across evaluation domains. Moreover, 52% expressed willingness for payment, while 4.9% were unwilling.
Clinicians rated deployment factors and identified patient acceptance, regulatory issues, and system integration as the most important considerations, while rating high on all decision-support dimensions.
The authors also explored biological interpretability by linking retinal latent features to disease-related plasma protein signatures, although genetic risk associations were limited. Saliency maps were used to highlight fundus regions relevant to model predictions.
Retinal AI Screening Implications
In summary, the study developed a retinal imaging-based framework, Reti-Pioneer, to detect and predict six metabolic and endocrine diseases.
Reti-Pioneer demonstrated overall moderate-to-strong discriminative performance that varied by disease and cohort, with consistent performance across datasets, ethnicities, and resource settings. It also demonstrated a promising, but still preliminary, ability to stratify future disease risk. However, the authors cautioned that its current diagnostic and predictive accuracy remains below the threshold required for broad clinical adoption, and that larger, multicenter, randomized studies are needed to test patient-level benefits and real-world clinical impact.
Overall, Reti-Pioneer could provide a low-cost, scalable, and translatable pathway from oculomics to actionable clinical screening.
Download your PDF copy by clicking here.
Wegovy linked to higher risk of eye stroke and vision loss