AI framework for multidisease detection via retinal imaging

The rising burden of endocrine and metabolic diseases demands scalable and accessible screening tools. Here we developed Reti-Pioneer, a multitask retinal imaging framework that integrates quality-aware modules with pre-trained foundation models for efficient, multidisease detection. In general, the framework was developed using 107,730 color fundus photographs from both community-based and hospital-based cohorts and achieved area under the receiver operating characteristic curve values on internal test data of 0.833 (95% confidence interval 0.810-0.856) for type 2 diabetes mellitus, 0.832 (0.799-0.866) for gout, 0.787 (0.742-0.833) for osteoporosis, 0.740 (0.726-0.755) for hypertension, 0.736 (0.721-0.751) for hyperlipidemia and 0.699 (0.667-0.730) for thyroid disease. The framework generalized well to six external cohorts from both resource-limited and high-resource settings, and showed biological interpretability via plasma proteomic correlations. In a primary care silent trial, it completed screening in 30.6 ± 6.0 s per case, notably faster than standard laboratory workflows. A subsequent clinical pilot for type 2 diabetes mellitus yielded an area under the receiver operating characteristic curve of 0.776 (0.710-0.842) and negative predictive value of 0.966 (0.946-0.983), surpassing the Finnish Diabetes Risk Score, with high acceptance from clinicians and patients. Overall, Reti-Pioneer could provide a translatable, low-cost pathway from oculomics to actionable clinical screening.