Digital solution for detection of undiagnosed diabetes using machine learning-based retinal image analysis.

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Undiagnosed diabetes is a significant global health concern, contributing to severe complications and high healthcare costs.
• Current diagnostic methods based on metabolic indices have limitations due to variability.
• Retinal imaging offers a stable, non-invasive marker for assessing glycemic status.

Purpose of the Study:

• To develop and validate a machine learning-based classification model for diabetes detection using retinal images.
• To assess the efficacy of this digital approach across diverse populations (Asian and Caucasian).
• To establish retinal image analysis as a tool for prescreening undiagnosed diabetes.

Main Methods:

• A classification model was developed using retinal images from 2221 subjects (945 with diabetes, 1276 controls).
• A 10-fold cross-validation with a support vector machine approach was employed for model training and internal validation.
• External validation was performed on separate datasets from Hong Kong and the UK.

Main Results:

• The model achieved high performance metrics, including 92% sensitivity and 96.2% specificity in cross-validation.
• External validation demonstrated excellent sensitivity (99.5% in Hong Kong, 98.0% in UK).
• The accuracy was comparable across Asian and Caucasian retinal images, indicating global applicability.

Conclusions:

• A validated digital machine learning method for diabetes risk estimation using retinal images has been developed.
• Retinal image analysis is a fast, convenient, and non-invasive technique suitable for community health applications.
• This digital method serves as an ideal solution for prescreening undiagnosed diabetes.