Article Impact Level: HIGH Data Quality: STRONG Summary of Cardiovascular Diabetology, 24(1), 3. https://doi.org/10.1186/s12933-024-02564-w Dr. Mohammad Ghouse Syed et al.
Points
- A deep-learning AI model analyzed retinal photographs from diabetic patients to predict their 10-year risk of major adverse cardiovascular events like heart attack or stroke with high accuracy.
- The AI model’s predictive performance was found to be comparable to the traditional pooled cohort equation clinical risk score, both achieving an area under the curve of 0.697.
- Combining the AI retinal score with the clinical risk score and a polygenic risk score significantly improved predictive accuracy for cardiovascular events compared to using the clinical score alone.
- Researchers observed that a measurable increase in a person’s AI-predicted retinal risk score over three years was associated with a higher likelihood of future major adverse cardiovascular events.
- This technology suggests that routine diabetic eye screenings could become a valuable one-stop opportunity for non-invasive cardiovascular disease risk assessment in the future for populations at risk.
Summary
A research team developed a deep-learning artificial intelligence model to predict 10-year cardiovascular disease (CVD) risk using routine diabetic retinal screening photographs. The study included 6,127 individuals with type 2 diabetes, with the cohort divided into training, validation, and testing cohorts of 1,241 individuals. The AI model, an EfficientNet-B2 network, was trained to predict the time to a first major adverse CV event (MACE). The primary analysis focused on comparing the AI-derived retinal risk score to the traditional pooled cohort equation (PCE) risk score and a coronary heart disease polygenic risk score (PRS).
The results from the test cohort, which experienced 288 MACE events, demonstrated a strong correlation between the AI-predicted risk and the PCE risk score (r = 0.66) but not the PRS (r = 0.05). Higher retina-predicted risk was significantly associated with an increased 10-year risk of MACE (HR 1.05 per 1% increase; 95% CI 1.04–1.06, p < 0.001), an association that persisted after adjustment for both the PCE and PRS (HR 1.03; 95% CI 1.02–1.04, p < 0.001). The retinal risk score’s performance was equivalent to the PCE, with both achieving an AUC of 0.697 for MACE prediction.
Furthermore, the study found that combining the AI retinal score with the PCE and PRS significantly improved risk prediction compared to the PCE alone, increasing the AUC to 0.728. This combined approach could potentially identify an additional three at-risk individuals per 100 people screened. The researchers also noted that an increase in a patient’s retinal-predicted risk over three years was associated with a subsequently higher likelihood of MACE. These findings suggest AI-powered retinal analysis could enable a “one-stop” CVD risk assessment during routine screenings.
Link to the article: https://cardiab.biomedcentral.com/articles/10.1186/s12933-024-02564-w
References Syed, M. G., Trucco, E., Mookiah, M. R. K., Lang, C. C., McCrimmon, R. J., Palmer, C. N. A., Pearson, E. R., Doney, A. S. F., & Mordi, I. R. (2025). Deep-learning prediction of cardiovascular outcomes from routine retinal images in individuals with type 2 diabetes. Cardiovascular Diabetology, 24(1), 3. https://doi.org/10.1186/s12933-024-02564-w
