AI Screening for Referable AMD Shows 90% Sensitivity
Key Takeaways
- In adults older than 50 without a prior AMD diagnosis who were screened in New York City primary care and ophthalmology clinics, iPredict-AMD identified referable AMD with strong discrimination in this study.
- Most participants with disease beyond early AMD were correctly classified on automated screening.
- Negative automated screens were rarely associated with referable AMD, suggesting strong rule-out performance in this cohort.
Investigators conducted a prospective iPredict-AMD clinical validation study across three primary care clinics and three general ophthalmology clinics in New York City in adults older than 50 years without prior AMD, enrolling 845 subjects and including 696 in the primary endpoint analysis, with 339 from primary care and 357 from ophthalmology clinics. One non-dilated color fundus image per eye was captured with the DRSPlus camera for artificial intelligence (AI) input, while dilated images were independently graded by three masked ophthalmologists as no AMD, early AMD, intermediate AMD, or late AMD, with majority decision used as ground truth. The automated report was generated within a minute and could return referable AMD, non-referable AMD, or insufficient image quality.
Reported subject-level and eye-level AMD screening performance showed that subject-level detection of mteAMD had an area under the curve (AUC) of 0.92. Subject-level sensitivity was 90% and specificity was 83%. Negative predictive value (NPV) was 98%. In this cohort, negative screens were rarely associated with referable AMD.
Referable AMD prevalence was 16% (95% CI 14%-19%) at the subject level and 16% (95% CI 15%-19%) at the eye level. The automated system correctly identified 102 of 113 subjects with referable AMD. The authors report that some participants were later found by graders to have diabetic retinopathy, epiretinal membrane, high myopia, or macular pucker were also flagged as referable.
Generalizability may be limited by the single-city setting and by the 149 enrolled subjects not included in the primary endpoint analysis because of eligibility issues, age ineligibility, undisclosed prior AMD, or incomplete ground-truth imaging and protocol follow-through. The model was not designed or trained to detect diseases other than AMD, so incidental non-AMD flags should not be interpreted as validated performance outside its intended use. The report also described deployment considerations such as access to imaging, internet bandwidth, trained personnel, and possible electronic health record (EHR) integration.
Clinician Questions
What counted as referable AMD in the iPredict-AMD validation?
Referable AMD was defined as more than early age-related macular degeneration (mteAMD). Three ophthalmologists graded dilated fundus images as no AMD, early AMD, intermediate AMD, or late AMD using a 4-level Age-Related Eye Disease Study (AREDS) framework, and the tool was evaluated against identification of disease beyond early AMD.
How was ground truth established for iPredict-AMD in this clinical validation?
Dilated fundus images were captured by a trained imaging operator and independently graded by three masked ophthalmologists. Majority decision among those three graders served as the reference standard for comparison with the automated output.
Does this validation establish detection of diabetic retinopathy or other retinal disease?
No. The model was not designed or trained to detect diseases other than AMD, and any referable flags seen in diabetic retinopathy, epiretinal membrane, high myopia, or macular pucker were incidental observations rather than validated non-AMD performance.
Which patients and clinical settings do these iPredict-AMD results apply to?
The validation involved adults older than 50 years without a prior AMD diagnosis who were screened in three primary care clinics and three general ophthalmology clinics in New York City; patients with known retinal diseases such as diabetic retinopathy or retinal vein occlusion were excluded. The single-city setting and pre-endpoint attrition limit how far those findings can be generalized beyond that workflow.
Recommended Reading
- For more on earlier AMD detection: Improving Early Detection of Geographic Atrophy
- For more on early AMD management: Novel Approaches in Early and Intermediate AMD: A Mechanistic Deep Dive