A new AI model from the New York Eye and Ear Infirmary of Mount Sinai boasts 99.2% accuracy for diagnosing intermediate or late AMD.
R. Theodore Smith, MD, PhD
Results of a study from the New York Eye and Ear Infirmary of Mount Sinai (NYEE) suggest newly developed artificial intelligence (AI) algorithm could increase the speed and accuracy detecting age-related macular degeneration (AMD).
The study, which used data from more than 100k color fundus photos from more than 4000 patients, indicated the deep-learning model was able to diagnose intermediate or late AMD with 99.2% accuracy and predict 2-year incidence of late AMD with 86.36% accuracy.
“We are excited to have built a deep-learning form of AI that can be trained to match the performance of a human expert to accurately diagnose AMD grade and stage based on scanning retinal photographs, without using other information,” said lead investigator R. Theodore Smith, MD, PhD, professor of ophthalmology at the Icahn School of Medicine at Mount Sinai, in a statement. “This is an important step in identifying those at risk for late-stage AMD and may allow them to get quick referral to an eye specialist for timely, preventive treatment.”
In an effort to build, and subsequently validate, an AI-based model for AMD screening to aid in the detection of late dry and wet AMD progression, investigators designed a 2-step study to create and assess a model for this purpose. The first step included creating a deep-learning prediction model using data from the Age-related Eye Disease Study (AREDS) and testing it using data from the Nutritional AMD Treatment-2 (NAT-2) study.
From AREDS data, investigators were able to create an AI model trained through 116,875 color fundus photos from 4139 individuals. The model was designed to determine the presence and severity of AMD while also stratifying them according to the AREDS 12-level severity scale. The resulting AMD scores were then combined with sociodemographic clinical data along with other clinical data to predict risk for progression to late AMD within 1 or 2 years.
Upon analysis, results indicated the model achieved 99.2% (95% CI, 99.02—99.39) accuracy in distinguishing early from intermediate/late AMD in binary screening of AMD stage (sensitivity of 98.9% [95% CI, 98.64-99.66]; specificity of 99.5% [95% CI, 98.85-99.80]).
When examining incidence of late AMD in 2 years, the model achieved 86.36% (95% CI, 84.22-88.31) accuracy, 92.42% (88.64—95.25) sensitivity, and 84.39% (81.78-86.76) specificity. For 1-year incidence of late AMD, the model achieved 86.19% (84.03-88.15) prediction accuracy with 90.74% (86.64-93.92) sensitivity and 84.74% (82.15-87.09) specificity.
Investigators examined prediction of late dry and wet AMD separately, based on incident type. Results of this analysis indicated the model achieved 66.88% (95% CI, 64.01-69.66) accuracy with 69.16% (59.50-77.73) sensitivity and 66.63% (63.60—69.56) specificity for the 2-year incidence of late dry AMD and 67.15% (64.29-69.93) accuracy with 71.43% (63.19-78.74) sensitivity, and 66.53% (63.44-69.51) specificity for 2-year incidence of late wet AMD.
To further evaluate the model, investigators assessed 2-year late AMD prediction using NAT-2 data. Results of this analysis indicated the model achieved an accuracy of 84% (95% CI, 74.75-91.02) with a sensitivity of 90% (95% CI, 73.47-97.89) and a specificity of 81% (95% CI, 68.59-90.13).
“This may become an important and cost-effective tool for high-risk or low-income groups who may not have direct or frequent access to eye screening, as early detection is critical to preventing AMD,” said Smith in the aforementioned statement.
This study, “Artificial Intelligence to Stratify Severity of Age-Related Macular Degeneration (AMD) and Predict Risk of Progression to Late AMD,” was published in Translational Vision Science and Technology.