One-Year Prediction of Vision-Threatening Diabetic Retinopathy Complications Using Multimodal Artificial Intelligence: A Multicenter Real-World Study

Current diabetic retinopathy management relies primarily on contemporaneous severity grading and does not directly quantify the short-term risk of vision-threatening complications. We developed and internally evaluated image-based artificial intelligence models for one-year prediction of diabetic retinopathy severity, proliferative diabetic retinopathy (PDR) and center-involved diabetic macular edema (ci-DME) using retinal imaging acquired during routine care.The prospective multicenter EviRed cohort included 3,183 patients with diabetes, 8,520 visits and 16,509 eye-level observations collected across 14 French hospitals. Device-specific ensembles were developed for ultra-widefield color fundus photography (CLARUS500 and OPTOS) and OCT angiography (CIRRUS5000 and PLEX Elite9000). Predictions from individual images and both eyes were aggregated at the patient-visit level. Participants were partitioned at the patient level into adevelopment set (65\%) and a held-out test set (35\%).For prediction of diabetic retinopathy severity at the subsequent annual visit, OPTOS and CLARUS500 achieved quadratic weighted kappa values of 0.866 (95\% confidence interval, 0.842--0.878) and 0.852 (0.830--0.877), respectively. For incident ci-DME, CIRRUS5000 achieved the highest discrimination (AUC, 0.826; 0.770--0.878), whereas PLEX Elite9000 yielded the highest AUC point estimate for incident PDR (0.877; 0.807--0.950). Late fusion did not consistently outperform the strongest device-specific model. In secondary complete-case analyses, observed ci-DME incidence increased across quartiles of the image-derived risk score, reaching 19.7--30.4\% in the highest-risk quartile. AI models also achieved higher AUC point estimates than unassisted expert risk assessments for both complications, although these comparisons were exploratory.Retinal imaging acquired at a single routine-care visit therefore contains information associated with diabetic retinopathy severity and vision-threatening complications one year later. These findings support image-based short-term risk stratification, while external validation, calibration and prospective clinical utility assessment remain necessary before deployment.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-18
DOI
https://doi.org/10.21203/rs.3.rs-11433556/v1
Primary Topic
Retinal Imaging and Analysis
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

One-Year Prediction of Vision-Threatening Diabetic Retinopathy Complications Using Multimodal Artificial Intelligence: A Multicenter Real-World Study

Abir Zureik, Mohsen Tabejamaat, Gwenolé Quellec, Mathias Gallardo et al.
2 links
Research Square
Retinal Imaging and Analysis
preprint

One-Year Prediction of Vision-Threatening Diabetic Retinopathy Complications Using Multimodal Artificial Intelligence: A Multicenter Real-World Study

Abir Zureik, Mohsen Tabejamaat, Gwenolé Quellec, Mathias Gallardo, Pascale Massin, Deborah Cosette, Philippe Zhang, Aurélien Ducournau, Mathieu Lamard, Alireza Rezaei, Francois Potevin, Pierre Deman, Laurent Borderie, Sejal Ghate, Éric Vicaut, Aude Couturier, Capucine Lepicard, Marine De Bony, Ali Salehi, Hugo LE BOITE, Alexandre LE GUILCHER, Mostafa EL HABIB DAHO, EviRed Investigators
preprint en

Abstract

Current diabetic retinopathy management relies primarily on contemporaneous severity grading and does not directly quantify the short-term risk of vision-threatening complications. We developed and internally evaluated image-based artificial intelligence models for one-year prediction of diabetic retinopathy severity, proliferative diabetic retinopathy (PDR) and center-involved diabetic macular edema (ci-DME) using retinal imaging acquired during routine care.The prospective multicenter EviRed cohort included 3,183 patients with diabetes, 8,520 visits and 16,509 eye-level observations collected across 14 French hospitals. Device-specific ensembles were developed for ultra-widefield color fundus photography (CLARUS500 and OPTOS) and OCT angiography (CIRRUS5000 and PLEX Elite9000). Predictions from individual images and both eyes were aggregated at the patient-visit level. Participants were partitioned at the patient level into adevelopment set (65\%) and a held-out test set (35\%).For prediction of diabetic retinopathy severity at the subsequent annual visit, OPTOS and CLARUS500 achieved quadratic weighted kappa values of 0.866 (95\% confidence interval, 0.842--0.878) and 0.852 (0.830--0.877), respectively. For incident ci-DME, CIRRUS5000 achieved the highest discrimination (AUC, 0.826; 0.770--0.878), whereas PLEX Elite9000 yielded the highest AUC point estimate for incident PDR (0.877; 0.807--0.950). Late fusion did not consistently outperform the strongest device-specific model. In secondary complete-case analyses, observed ci-DME incidence increased across quartiles of the image-derived risk score, reaching 19.7--30.4\% in the highest-risk quartile. AI models also achieved higher AUC point estimates than unassisted expert risk assessments for both complications, although these comparisons were exploratory.Retinal imaging acquired at a single routine-care visit therefore contains information associated with diabetic retinopathy severity and vision-threatening complications one year later. These findings support image-based short-term risk stratification, while external validation, calibration and prospective clinical utility assessment remain necessary before deployment.

Research Square
Carl Zeiss (Germany) (DE), Université de Bretagne Occidentale (FR), Université Paris Cité (FR), Hôpital Lariboisière (FR), Laboratoire de Traitement de l'Information Médicale (FR)
Agence Nationale de la Recherche
Good health and well-being
Retinal Imaging and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon