Predicting the risk of developing diabetic retinopathy using deep learning.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic retinopathy (DR) screening is crucial for preventing blindness.
• Increasing diabetes prevalence poses challenges to scaling DR screening efforts.
• Developing automated systems is essential for efficient patient risk stratification.

Purpose of the Study:

• To develop and validate a deep-learning (DL) system for predicting the 2-year risk of DR development.
• To assess the performance of DL models using different fundus photograph inputs (three-field vs. one-field).
• To compare the predictive power of DL systems against traditional risk factors.

Main Methods:

• Two DL systems were created using teleretinal screening data from primary care.
• Models were trained on a large dataset of fundus photographs, augmented with multitask learning.
• Validation was performed on internal (USA) and external (Thailand) datasets.

Main Results:

• The three-field DL system achieved an AUC of 0.79 in internal validation.
• The one-field DL system achieved an AUC of 0.70 in external validation.
• Combining DL systems with risk factors significantly improved AUC in both validation sets (0.81 internal, 0.71 external).

Conclusions:

• DL systems can predict DR development from fundus photographs.
• These systems are independent of and more informative than existing risk factors.
• Risk stratification tools using DL can optimize screening and improve patient outcomes.