OR WAIT null SECS
A new analysis presented at ADA 2023 suggests a machine learning algorithm may refine the risk of diabetic retinopathy progression and personalize screening intervals for patients.
New research indicates the accuracy and feasibility of using machine learning models to identify the progression of diabetic retinopathy using ultra-widefield images.1
The results, presented at the 83rd American Diabetes Association Scientific Sessions (ADA 2023), show the artificial intelligence (AI) prediction for approximately 91% of the images were either correct labels; or the labels showed greater progression than the original labels.
“Currently, estimating the risk of diabetic retinopathy progression is one of the most important, yet difficult tasks for physicians when caring for patients with diabetic eye disease,” Paolo S. Silva, MD, co-chief of telemedicine at Beetham Eye Institute, Joslin Diabetes Center and associate professor of ophthalmology, Harvard Medical School, said in a statement.1 “Our findings show that potentially, the use of machine learning algorithms may further refine the risk of disease progression and personalize screening intervals for patients, possibly reducing costs and improving vision-related outcomes.”
The number of people with diabetic retinopathy is expected to nearly double by 2050 and affect more than 14 million people in the United States.2 Determining the risk of diabetic retinopathy progression can prove difficult in the clinical setting, as medical knowledge and clinical experience required to estimate said risk can vary between clinicians. The current severity scale for diabetic retinopathy may inform clinicians of the progression risk and provide updated recommendations for follow-up and treatment.
The current analysis evaluated how the usage of AI algorithms may improve the process of estimating the risk of diabetic retinopathy progression. To do so, the investigative team created and validated machine learning models for diabetic retinopathy progression from ultra-widefield retinal images. The investigative team labeled each image for baseline diabetic retinopathy severity and progression based on a clinician review of the images and 3-year longitudinal follow-up using the Early Treatment Diabetic Retinopathy Study (ETDRS) severity scale.
The data showed 8 classes: no DR non-progression (14.62%), mild non-proliferative DR (NPDR) progression (10.16%)/non-progression (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%). The analysis split 9970 unique images into the train, the validation, and the test datasets based on 60-20-20 proportions.
Investigators indicated the class imbalance was addressed during the model building using data augmentation. In addition, they noted the ResNet model trained on the dataset had a classification test accuracy of 81% and an area under the curve (AUC) of 0.967 on the test dataset. The objective of the model was to reduce false negatives, which refers to predicting a class that is less progressive than the true label.
Upon analysis, investigators found the predicted labels for 91% of the images were either correct or showed greater progression than the original labels. As a result, the analysis demonstrates the accuracy and feasibility of using machine learning models for identifying DR progression developed using ultra-widefield images.
The investigative team suggests the use of machine learning algorithms may further refine the risk of disease progression and personalize screening intervals that may reduce costs and improve vision-related outcomes.