Predicting 1, 2 and 3 year emergent referable diabetic retinopathy and maculopathy using deep learning

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy (DR) progression prediction is crucial for personalized screening and timely treatment.
• Deep learning systems (DLS) were developed to predict emergent referable DR and maculopathy.
• Systems utilized risk factor data, fundal photographs, or both (multimodal).

Purpose of the Study:

• To develop and validate deep learning systems (DLS) for predicting 1, 2, and 3-year emergent referable diabetic retinopathy (DR) and maculopathy.
• To compare the performance of DLS using risk factor data, fundal images, or a combination of both.
• To assess the potential of DLS for individualised, risk-based screening in diabetic eye care.

Main Methods:

• Developed and validated deep learning systems (DLS) using a large dataset of eyes from UK diabetic eye screening programmes.
• Trained models on 110,837 eyes with longitudinal data and pre-trained on 51,502 eyes.
• Tested DLS performance on internal (27,996 eyes) and external (6928 eyes) datasets.

Main Results:

• Multimodal DLS demonstrated high predictive accuracy for emergent referable DR and maculopathy at 1, 2, and 3 years, with AUROCs ranging from 0.79 to 0.95.
• Performance was significantly better for multimodal and image-based DLS compared to tabular DLS.
• External validation confirmed the robust performance of the multimodal DLS across different datasets.

Conclusions:

• Deep learning systems, particularly multimodal and image-based approaches, accurately predict diabetic retinopathy and maculopathy progression.
• These DLS offer a promising tool for individualised, risk-based screening, optimizing resource allocation and patient care.
• AI-assisted screening can help identify high-risk individuals for prompt intervention while reducing unnecessary screening for low-risk patients.