UBC researchers used AI to identify sex differences in diabetic retinopathy that could be essential for precision diagnostics and care.
Sex differences could play a key role in how vision complications are diagnosed and managed in people with diabetes, according to a recent study led by UBC researcher Dr. Ipek Oruc.
The study, published in PLOS One, used AI to identify potential differences in how diabetic retinopathy — the most common cause of blindness associated with diabetes — progresses depending on whether a patient is born male or female.
“Regular monitoring is crucial for detecting early signs of diabetic retinopathy and to prevent its progression to moderate or severe disease,” said Dr. Oruc, an associate professor in the Department of Ophthalmology and Visual Sciences at UBC. “However, aside from considerations during pregnancy when females are at higher risk for progression, sex is not presently factored into disease diagnosis, management or treatment.”
Diabetic retinopathy affects more than one million Canadians. The condition is characterized by changes to tiny blood vessels at the back of the eye, an area known as the fundus. In the early stages, symptoms of the disease often go unnoticed. However, once the condition advances, various vision-threatening complications may arise including the risk of vision loss.
“If left untreated, around half of high-risk non-proliferative patients will experience progression to proliferative diabetic retinopathy within a year — making it particularly crucial to identify and manage the disease early,” said Dr. Oruc, who is also a researcher with the Vancouver Coastal Health Research Institute and Djavad Mowafaghian Centre for Brain Health.
“Sex-specific disease indicators could have significant implications for the diagnosis, management and treatment of diabetic retinopathy.”
Dr. Ipek Oruc
Using AI to detect subtle differences
For the study, researchers trained and tested an AI model using nearly 3,000 fundus images from patients with diabetic retinopathy. The images came from roughly equal numbers of female and male patients.
The AI was trained to determine a patient’s sex based solely on the images. The team used a type of AI called a convolutional neural network, or CNN, which is designed to recognize patterns in images. This builds on their earlier work showing that AI can detect sex-based differences in healthy retinas.
“CNNs can find patterns in images not readily apparent to the human eye, even among experts in the field,” said Dr. Oruc.
To understand if there were differences in how diabetic retinopathy manifests in females and males, the researchers used an explainable-AI technique called Guided Grad-CAM. This method is able to peer into the mysterious ‘black box’ of AI models to highlight which areas of an image the AI focuses, helping researchers interpret the results.
The analysis showed that the AI focused on different parts of the retina depending on sex. In images from female patients, the model emphasized the macula, the area responsible for sharp central vision. In images from male patients, it focused more on the optic disc and surrounding blood vessels.
“This pattern differed noticeably from the saliency maps generated by CNNs trained on healthy eyes, which did not highlight these particular regions, indicating that diabetic retinopathy may manifest differently by sex,” said Dr. Oruc.
Sex-specific diagnostics could enhance care
The locations in the fundus emphasized by the model raise the hypothesis that women may be at higher risk for developing macular edema — a leading cause of blindness in diabetic retinopathy — while men may be at greater risk for more severe proliferative diabetic retinopathy.
“Further research is needed to validate this hypothesis,” said Dr. Oruc, which is why her team is presently conducting a follow-up study. “If confirmed, sex-specific disease indicators could have significant implications for the diagnosis, management and treatment of diabetic retinopathy.”
The research team’s application of AI was able to achieve a high level of accuracy using a smaller dataset for training the algorithm than is typical.
“This is particularly important, as large datasets are more costly and time-consuming,” said Dr. Oruc. “The success of this approach with a small dataset puts studies like ours within reach of other research teams with limited resources.”
A version of this story was originally published on the Vancouver Coastal Health Research Institute website.