Technology ID
TAB-4995

Using Artificial Intelligence To Predict The Risk Of Age-Related Macular Degeneration

E-Numbers
E-057-2020-0
Lead Inventor
Chew, Emily (National Eye Institute (NEI))
Co-Inventors
Lu, Zhiyong (National Library of Medicine, NIH)
Keenan, Tiarnan (National Eye Institute (NEI))
Wong, Wai (National Eye Institute (NEI))
Chen, Qingyu (National Library of Medicine, NIH)
Agron, Elvira (National Eye Institute (NEI))
Applications
Diagnostics
Development Stages
Pre-clinical (in vivo)
Development Status
prototype
Lead IC
 

Summary: 
The National Eye Institute seeks research co-development partners and/or licensees for a deep learning algorithm that can predict the probability of progression to late age-related macular degeneration.

Description of Technology: 
In 2024, an estimated 200 million people worldwide suffer from age-related macular degeneration (AMD); projected to affect ~288 million people by 2040. AMD is the leading cause of blindness in all developed countries. Identifying eyes at high risk of progression to late AMD, the stage associated with blindness, is vital. This would allow timely medical treatments, lifestyle interventions, more tailored home monitoring and improved clinical trials for patients.

Reticular pseudodrusen (RPD) is an AMD disease feature recently discovered to confer greatly increased risk of progression to late AMD. However, RPD is often very difficult to detect on clinical examination or on color fundus photography (CFP). Detection usually requires specialized imaging (especially fundus autofluorescence) and highly expert grading typically available at few specialized centers. For these reasons, RPD have not been incorporated into AMD risk classification systems.

We used Artificial Intelligence (AI) to predict the risk of progression to late AMD using over 80,000 images from almost 3300 participants from the Age-Related Eye Disease Studies AREDS and AREDS2. Using independent test data, our deep learning algorithm produced 5% higher prognostic accuracy compared to existing clinical standards. The predictive accuracy of the new approach was 5% higher than that of the two traditional approaches ((i) AREDS Simplified Severity Scale, and (ii) the Casey AMD online calculator). Our approach can make predictions over a wide range of time intervals (1-12 years), and separately for the two subtypes of late AMD (geographic and neovascular AMD). In contrast, the AREDS Simplified Severity Scale can make predictions at one fixed interval only (5 years), and for late AMD only (not separately by subtype). By separating the deep learning extraction of retinal features from the survival analysis, the final predictions are more explainable and biologically plausible, and error analysis is possible. By contrast, end-to-end ‘black-box’ deep learning approaches are less transparent and may be more susceptible to failure
A fully automated device that contains this novel image processing method has also been developed.

Potential Commercial Applications: 


• Diagnostic tool predicting risk of AMD

Competitive Advantages:


• Widely available via device 
• Fully automated analysis of the CFP and no requirement for human grading of the CFP, either by retinal specialists or by reading center experts
• More predictive, accurate approach compared using the same test set of AREDS and AREDS2 participants 
• Predictive over a wide range of time intervals (1-12 years) and separately for the two subtypes of late AMD (geographic and neovascular AMD) 
• Two-step method separates the deep learning extraction of retinal features from the survival analysis
• Two-step method produces final predictions that are more explainable and biologically plausible
• Error analysis 
• Our approach has the advantage of not requiring genetic information to provide a high level of predictive accuracy
 

Licensing Contact:
Alsaffar, Hiba
hiba.alsaffar@nih.gov