New research out of the National Eye Institute (NEI) supports the development of artificial intelligence/machine learning (AI/ML)-based systems that are focused on preventive measures for patients with age-related macular degeneration (AMD).
Tell me more about these systems.
These AI/ML-based systems are designed to not only screen for AMD, but also predict which patients may be likely to progress to late AMD within 2 years as well as evaluate their chances for developing late neovascular (wet) AMD based on their risk for late dry (geographic atrophy [GA]) AMD.
Have any companies released their own systems?
Indeed. New York-based, AI-ML-based system developer iHealthScreen, has launched the iPredict, a retinal image-grading system designed to allow non-eye care practitioners (ECPs) to screen for AMD and predict which patients diagnosed with early AMD may be at risk for a later late-stage diagnosis.
Tell me more about the iPredict.
The iPredict was trained and validated via a prediction model by iHealthScreen founder Alauddin Bhuiyan, PhD, using 93,830 color fundus images from patients within the early and late AMD groups in the NEI-funded Age-Related Eye Disease Study [AREDS]). Further validation was provided via images from the Nutritional AMD Treatment-2 (NAT-2) study.
Results from the iPredict included an 86% prediction accuracy in the 2-year risk for progression to late AMD (of the AREDS data) and an 84% prediction accuracy using the NAT-2 data.
See here for more details.
How does it work?
A non-ECP can use a fully-automated fundus camera to capture bilateral color images of the retina along with recording pertinent patient data (age, gender, smoking, etc.). The data is sent to a centralized server and is then analyzed by the iPredict’s algorithm for any indications and diagnosis of AMD.
A report is created in approximately 1 minute, which determines whether the patient is referable or non-referable for AMD. For referable patients, a prediction score is generated (0 to 100% scale) that assesses the predictive value of a patient for developing late AMD in the next 1 to 2 years.
Any real-world use yet?
That's in progress. The system’s accuracy is currently being tested (NCT04863391) in a non-ECP setting at four primary care clinics in New York, contracted through the NY Eye and Ear Infirmary. Providers are screening adults (ages 50+) during annual check-ups; results will be compared to a retinal specialist’s assessments.
Any other AI/ML-based systems being studied?
Yes … the NEI has developed one in collaboration with the National Center for Biotechnology Information, an extension of the National Library of Medicine.
While similar in development to the iPredict—in that images from the AREDS study were used—researchers also utilized data of patients with intermediate AMD from the AREDS2 study and trained the model to predict the probability for late AMD progression using AREDS/AREDS2 data.
According to researchers, when validated against an independent test dataset of 601 patients, the model’s prognostic accuracy outperformed that of retinal specialists using two clinical standards.
Any other similarities?
Both systems are designed to identify reticular pseudodrusen, which is often a challenge to detect due to its composition and location.
Data thus far points to reticular pseudodrusen being connected to a higher risk for late AMD progression inclusive of geographic atrophy; however, scientists state more research is needed to understand its mechanisms.
What else should I know?
According to Bhuiyan, the iPredict screening model has been prospectively validated and submitted to the FDA for clearance; if cleared, the company hopes to market it to the primary care practice system by the end of 2023.