Diabetic retinopathy (DR), a leading cause of vision loss in working-age adults, affects approximately one in three individuals with diabetes. The global prevalence of diabetes is projected to increase significantly, resulting in a corresponding rise in DR cases. Current diabetic eye screening programs, such as the UK's Diabetic Eye Screening Programme (DESP), aim to detect referable DR through color fundal photographs (CFPs) graded by trained professionals. However, this approach is resource-intensive, with a large proportion of screenings showing no referable DR. This study addresses the need for a more efficient and targeted screening approach by developing and validating deep learning systems (DLS) to predict the emergence of referable DR and maculopathy over 1, 2, and 3-year periods. The ability to accurately predict future disease progression would allow for individualized screening intervals, leading to earlier intervention for high-risk individuals and reduced burden for those at low risk. This would result in optimized resource allocation and improved patient outcomes, minimizing unnecessary appointments and maximizing the timely detection and treatment of sight-threatening conditions.

Existing DR progression prediction models often rely on systemic risk factors (such as HbA1c and lipid levels) or historical retinopathy grades. Systemic risk factor models require invasive blood tests, complicating implementation within existing screening services. Retinopathy-grade-based approaches, while readily available, often lack specificity, leading to a higher rate of unnecessary referrals. Furthermore, these methods frequently necessitate two sequential visits, increasing resource consumption and being reliant on patient attendance. Previous studies have explored the use of deep learning systems (DLS) for predicting DR progression using CFPs. However, limitations include small sample sizes, non-representative datasets, and exclusion of certain patient groups. Notably, previous DLS models mostly focused on predicting DR progression generally and not independently predicting referable maculopathy. This study aims to overcome these limitations by using a larger, more diverse dataset, incorporating both CFPs and risk factor characteristics, and separately predicting referable DR and maculopathy, both critical vision-threatening outcomes.

This study utilized data from two geographically distinct UK DESPs: the south-east London DESP (SEL-DESP) and the Birmingham, Solihull, and Black Country DESP (BSBC-DESP). A total of 162,339 eyes were initially included, with 110,837 having eligible longitudinal data used for model training and validation. The remaining data were used for model pretraining. Three types of DLS were developed: tabular DLS (using risk factor characteristics), image DLS (using CFPs), and multimodal DLS (combining both). The risk factors included age, sex, ethnicity, diabetes type, DM duration, visual acuity, and index of multiple deprivation rank. The image DLS comprised two EfficientNet-V2-s models, one for each eye field, with feature map concatenation. The tabular DLS utilized TabNet. Both image and tabular DLS underwent pretraining on a DR severity classification task before being trained to predict emergent referable DR, maculopathy, or either within 1, 2, or 3 years. An auxiliary task of detecting mild-moderate DR at baseline was also included. The multimodal DLS combined image and tabular DLS predictions via a mean ensemble. Performance was assessed using AUROC, specificity at 80% sensitivity, positive and negative predictive values, and stratified by age, sex, and ethnicity. Ablation and attribution studies were conducted to investigate the importance of specific CFP regions and risk factors.

The multimodal DLS achieved high AUROC values for predicting emergent referable DR, maculopathy, or either within 1, 2, and 3 years. For instance, the internal test multimodal DLS AUROC values were 0.95 (95% CI: 0.92–0.98) for 1-year emergent referable DR, 0.84 (0.82–0.86) for 1-year emergent referable maculopathy, and 0.85 (0.83–0.87) for 1-year emergent referable DR or maculopathy. Similar high performance was observed in the external test set, demonstrating good generalizability. Image and multimodal DLS outperformed the tabular DLS significantly in both internal and external testing. Specificity at 80% sensitivity was generally high for DR prediction, but lower for maculopathy. Ablation studies indicated the central macula field's importance for image DLS predictions. Attribution maps showed that the image DLS primarily focused on the central macula area for both DR and maculopathy prediction. Tabular DLS attributions highlighted risk factors such as lower visual acuity, longer diabetes duration, younger age, male gender, and Black, mixed, or unspecified ethnicity as associated with higher risk predictions. The study demonstrated that very few eyes progressed to proliferative DR undetected (0% in external testing and 0.3% in internal testing at 2 years).

The results demonstrate the potential of DLS for accurate prediction of emergent referable DR and maculopathy using readily available data from a single screening visit. The superior performance of image and multimodal DLS compared to tabular DLS highlights the value of integrating CFPs into prognostic models. The high AUROC values and acceptable specificity at 80% sensitivity suggest that these DLS could significantly improve the efficiency and effectiveness of diabetic retinopathy screening. The findings from ablation and attribution studies provide valuable insights into the features that are most relevant to the models predictions. The successful generalization of the model to an independent external test set indicates that the model is robust and can be applied to other populations. The study addressed limitations of existing models by using a large, diverse dataset, analyzing referable DR and maculopathy independently, and employing non-invasive data acquisition.

This study successfully developed and validated deep learning systems for predicting the emergence of referable DR and maculopathy over 1, 2, and 3-year periods. The high accuracy and generalizability of the models suggest a potential application in individualizing diabetic eye screening, leading to more efficient resource allocation and earlier interventions for high-risk individuals. Future research should focus on prospective validation, refining operating thresholds, and exploring integration into existing clinical workflows.

The study acknowledges limitations such as potential bias in self-reported data (ethnicity), potential grading variability, and the need for larger validation populations for more robust subgroup analysis. Further research is needed to evaluate the DLS's performance with other retinal imaging devices and to explore the use of different DLS architectures and training techniques. Determining clinically appropriate sensitivity and specificity thresholds for various population subsets will also be crucial for successful implementation in clinical practice.