Publication
Purpose : To evaluate whether ethnicity-diverse machine learning (ML) models improve the diagnostic performance for detecting more-than-mild diabetic retinopathy (mtmDR) on ultrawide-field (UWF) images compared with models trained on ethnicity-specific datasets.
Methods : A large-scale vision transformer architecture (SwinV2) was used to develop ML algorithms trained jointly (ethnic-diverse) and separately (ethnic-specific) on UWF images from White (W, N=4959), Non-White/Non-Black (NWB, N=12843), and Black (B, N=492) patients obtained from the Joslin Vision Network. Ground-truth grading of DR severity was determined by reading center evaluations following standardized ETDRS-based protocols. mtmDR was defined as moderate nonproliferative DR or worse, or the presence of diabetic macular edema. The ethnicity-diverse model was evaluated on held-out data from each ethnicity group, whereas each ethnicity-specific model was tested only on its matched held-out ethnicity subset. Performance metrics included sensitivity, specificity and area under the receiver operating characteristic curve (AUC).
Results : Prevalence of mtmDR in the W, NWB, and B datasets was 13.2%, 11.7%, and 20.7%, respectively. When evaluated on ethnicity-matched test sets, the performance of the W-trained model was sensitivity 0.919 specificity, 0.882, AUC 0.957; NWB-trained model achieved sensitivity 0.817, specificity 0.932, AUC 0.95; and the B-trained model achieved sensitivity 0.643, specificity 0.704, AUC 0.688. The combined multi-ethnic model achieved either equivalent or improved performance for each ethnic-specific tests: For W, sensitivity/specificity/AUC 0.924/0.883/0.965; for NWB, 0.853/0.908/0.95; for B, 0.857/0.87/0.948.
Conclusions : Ethnicity-diverse ML models demonstrated equivalent or superior diagnostic accuracy for detecting mtmDR on UWF images compared with non-diverse. These findings underscore the importance of population diversity and stratified training approaches in mitigating algorithmic bias and improving the equity, safety, and reliability of AI systems for diabetic eye disease detection.
