Story perspectives

UK Researchers Predict Six Metabolic Diseases via Retina, Blood Proteins

5/10/2026

38 6

Summary

UK Researchers Predict

UK researchers combined retinal images with plasma proteomics to predict six endocrine/metabolic diseases.
Saliency mapping and integrated gradients identified image pixels driving predictions.
Elastic net selected retinal features for an autoencoder-derived OPLS-DA model separating cases from controls.
The study measured 2,923 proteins on Olink Explore 3072, kept 2,920 after dropping >30% missing data, and validated the model on 15,000 images from Chinese and Singaporean in SEED cohort with IRB approval.

Summary

Reti-Pioneer Trials

The Reti-Pioneer framework incorporates a quality-aware module that retains low-quality fundus photos, using probability scores for good, usable, and bad classifications to guide early processing.
A prospective silent trial ran from September 2025 to October 2025, processing 30–50 cases per day in a primary-care setting while keeping AI predictions blinded from clinicians and patients.
An open-label pilot conducted between September 2025 and November 2025 provided AI-generated risk reports to both clinicians and participants, allowing assessment of diagnostic consistency with laboratory standards.
Participant and clinician satisfaction were measured with 5-point Likert-scale questionnaires covering ease of use, report clarity, workflow compatibility, trust in AI, and willingness to pay.
System throughput was tracked as the elapsed time from image capture to report generation, with images transmitted in real time to a centralized GPU cluster for immediate analysis.
The AI model was trained on an RTX 3080 GPU (10 GB memory) using PyTorch v2.8.0, the AdamW optimizer, a warm-up strategy, and a cosine-annealing learning-rate scheduler.

1 of 2

UK Researchers Predict

UK researchers combined retinal images with plasma proteomics to predict six endocrine/metabolic diseases.
Saliency mapping and integrated gradients identified image pixels driving predictions.
Elastic net selected retinal features for an autoencoder-derived OPLS-DA model separating cases from controls.
The study measured 2,923 proteins on Olink Explore 3072, kept 2,920 after dropping >30% missing data, and validated the model on 15,000 images from Chinese and Singaporean in SEED cohort with IRB approval.

1 / 2