Purpose: To develop a machine learning (ML)-driven polygenic risk score (PRS) for diabetic retinopathy (DR) and evaluate the extent to which lifestyle may mitigate genetic risk.
Design: Multicenter, multiethnic cohort study.
Participants: 9 1691 participants with DR-free prediabetes/diabetes from the UK Biobank (UKB); and 1119 participants with DR-free diabetes from the Guangzhou Diabetic Eye Study (GDES).
Methods: An ML-driven PRS for DR was constructed on an a priori basis by systematically integrating 182 literature-derived single nucleotide polymorphisms, which was subsequently applied to 2 multiethnic, prospective cohorts with genotyped data for unbiased stratification of genetic susceptibility. Lifestyle adherence was determined using a scoring system based on 4 behavioral factors (no smoking, optimal weight control, regular physical activity, and healthy sleep) that were categorized as favorable (3-4 factors), intermediate (2 factors), and unfavorable (0-1 factor). The independent and joint associations of genetic risk and lifestyle with incident DR were estimated using Cox proportional hazards models, with risk reduction analyses stratified by genetic risk category.
Main outcome measures: Hazard ratios (HR) on incident DR.
Results: Participants at high genetic risk had a 37% higher risk of developing incident DR than those at low genetic risk, regardless of lifestyle (HR, 1.37, 95% confidence interval [CI], 1.18-1.60; P < 0.001). An unfavorable lifestyle was associated with a 49% higher risk than a favorable lifestyle, regardless of genetic risk (HR, 1.49, 95% CI, 1.34-1.65; P < 0.001). High genetic risk combined with an unfavorable lifestyle more than doubled the DR risk (HR, 2.09, 95% CI, 1.67-2.60; P < 0.001). Among participants at high genetic risk, a favorable lifestyle was associated with a 44% lower DR risk (HR, 0.56, 95% CI, 0.42-0.75; P < 0.001), corresponding to a reduction in the standardized DR rates from 7.5% (95% CI, 7.4-7.6) to 4.6% (95% CI, 4.5-4.7). The gene-lifestyle synergistic associations were confirmed by replication in the GDES: compared with participants at low genetic risk, those at intermediate (HR, 1.47, 95% CI, 1.13-1.93; P = 0.004) and high (HR, 1.68, 95% CI, 1.30-2.17; P < 0.001) genetic risk exhibited a significantly greater risk of incident DR. Adherence to a favorable lifestyle resulted in an relative risk reduction (RRR) of 47.4% and an absolute risk reduction (ARR) of 5380.0 standardized DR events for those at high genetic risk, and an RRR of 16.6% and an ARR of 988.3 standardized DR events for those at intermediate genetic risk, per 100,000 person‑years of follow‑up.
Conclusions: Both genetic risk and modifiable lifestyles demonstrated independent and joint impacts on DR risk. Although risk reduction was greatest among those with high genetic risk, our findings support universal behavioral interventions for DR prevention, regardless of genetic background.
Financial disclosure(s): The author(s) have no proprietary or commercial interest in any materials discussed in this article.
Keywords: Diabetic retinopathy (DR); Joint effect; Lifestyle; Machine learning (ML); Polygenic risk.
Copyright © 2026 American Academy of Ophthalmology, Inc. Published by Elsevier Inc. All rights reserved.