Dr. Paolo S. Silva
Credit: Joslyn Diabetes Center
A new study demonstrates the accuracy and feasibility of using machine learning models to identify the progression of diabetic retinopathy using ultra-widefield images.1
Presented at the 83rd Annual Scientific Sessions of the American Diabetes Association (ADA 2023), the results show that artificial intelligence (AI) predicts correct labels for approximately 91% of images. Alternatively, the label showed greater progression than the original label.
“Currently, estimating the risk of developing diabetic retinopathy is one of the most important yet challenging challenges for physicians treating patients with diabetic eye disease,” said Joslyn’s Beetham Eye Institute Telemedicine. Co-Director Paolo S. Silva, M.D., said. Harvard Medical School Diabetes Center and Associate Professor of Ophthalmology said in a statement.1 “Our findings suggest that the use of machine learning algorithms may further tailor disease progression risk and individualize patient screening intervals, reducing costs and improving vision-related outcomes.” It shows that it is possible.”
The number of people with diabetic retinopathy is expected to nearly double by 2050, affecting more than 14 million people in the United States.2 Because the medical knowledge and clinical experience required to estimate the risk of progression of diabetic retinopathy varies among clinicians, it may prove difficult to determine the risk of progression of diabetic retinopathy in the clinical setting. There is a nature. The current severity scale for diabetic retinopathy may inform clinicians of progression risk and provide updated recommendations for follow-up and treatment.
The current analysis evaluated how the use of AI algorithms could improve the process of estimating the risk of diabetic retinopathy progression. To that end, the research team created and validated a machine learning model for the progression of diabetic retinopathy from ultra-wide-field retinal images. The research team determined baseline diabetic retinopathy severity and progression based on clinician review of images and 3-year long-term follow-up using the Early Diabetic Retinopathy Study (ETDRS) severity scale. We labeled each image for degrees.
Data showed eight classes: no DR nonprogression (14.62%), mild nonproliferative DR (NPDR) progression (10.16%)/nonprogression (10.73%), moderate NPDR progression (10.1%) /non-progression (15.85%), severe NPDR progression (11.27%)/non-progression (10.68%), and proliferative DR (16.55%). For analysis, 9970 unique images were split into train, validation, and test datasets based on a 60-20-20 ratio.
The investigators noted that class imbalance was addressed during model building using data augmentation. We also noted that the ResNet model trained on the dataset had an accuracy of 81% for the classification test and an area under the curve (AUC) of 0.967 for the test dataset. The goal of the model is to reduce false negatives, which refers to predicting classes that are less progressive than true labels.
Upon analysis, the researchers found that for 91% of the images, the predicted labels were either correct or showed significant improvement over the original labels. As a result, this analysis demonstrated the accuracy and feasibility of using a machine learning model for identifying the progression of DR developed using ultra-widefield images.
The researchers suggest that the use of machine learning algorithms could further tailor the risk of disease progression and personalize screening intervals to reduce costs and improve vision-related outcomes.
References
- The American Diabetes Association highlights innovations in diabetes technology for glycemic control and diabetic eye disease. American Diabetes Association Highlights Innovations in Diabetes Technology for Glycemic Control and Diabetic Eye Disease | Ada. June 23, 2023. Accessed June 24, 2023. https://diabetes.org/newsroom/press-releases/2023/american-diabetes-association-highlight-innovations-diabetes-technology-glucose-management-diabetic-eye-condition.
- Lundeen EA, Burke-Conte Z, Rein DB, et al. Prevalence of diabetic retinopathy in the United States in 2021. JAMA Physical mol. Published online June 15, 2023. doi:10.1001/jamaophysicalmol.2023.2289
