Artificial intelligence platforms are increasingly capable of identifying diabetic retinopathy (DR) severity with accuracy comparable to human experts, according to data presented at the Clinical Trials at the Summit. The Ocula360 algorithm, developed by Ophthalytics, demonstrated high agreement with human reading centers in classifying retinal images, suggesting a potential shift in how clinicians monitor disease progression and select patients for clinical trials.
How Ocula360 Performs Against Human Grading
The Ocula360 system uses a web-based platform to analyze ultra-widefield retinal images. In a validation study involving 120 images from the ONYX-1 study, the AI achieved a kappa value of 0.89, indicating excellent agreement with human experts.
According to findings presented by SriniVas R. Sadda, MD, of the Doheny Eye Institute, the platform categorized images into three severity classes:
- Class 1: DR severity scale 43 or lower.
- Class 2: DR severity scale 47 or 53.
- Class 3: DR severity scale 60 or higher.
The AI classified 51 images as class 1, 33 as class 2, and 31 as class 3. In comparison, human reading centers classified 50, 32, and 33 images into those same categories, respectively.
Why Automated DR Screening Matters
Current FDA-approved automated tools often face limitations when attempting to screen for specific disease severity levels. While many systems can identify the presence of disease, the ability to provide a granular, quantitative score remains a clinical challenge.
Dr. Sadda noted that the ideal diagnostic future involves the automated detection of all retinal lesions. By using ultra-widefield imaging, AI can potentially facilitate a continuous scoring system that captures peripheral disease—an area often missed by traditional, narrower field-of-view imaging. This transition toward quantitative, automated assessment could improve the consistency of patient monitoring in both routine clinical practice and high-stakes clinical research.
Future Directions for AI in Retinal Health
The integration of AI into retinal imaging is expected to influence how clinical trials identify eligible participants. By incorporating peripheral lesion data into standardized grading, researchers may refine the criteria used to assess DR severity.
As the technology evolves, the focus remains on ensuring these algorithms can handle the complexities of real-world imaging. While Ocula360 is currently trained on over 487,000 images, ongoing research aims to further validate its performance across diverse patient populations and varying clinical settings. The ability to automatically classify disease severity allows for more efficient screening, potentially reducing the burden on specialized reading centers while maintaining high diagnostic accuracy.
Summary of AI Diagnostic Accuracy
| Metric | Ocula360 Performance | Human Reading Center |
|---|---|---|
| Class 1 (Mild) | 51 images | 50 images |
| Class 2 (Moderate) | 33 images | 32 images |
| Class 3 (Severe) | 31 images | 33 images |
Data based on the ONYX-1 study validation results presented at the 2026 Clinical Trials at the Summit.