The system compared retina images with medical records.
Which makes me think the condition was already noted in the patient before the retina image was taken?
The trick is to identify the condition from the retina before the doctor notices it, not after. Which would require photographing a large number of volunteers, seeing which ones the system assessed as possibly having CV issues and then checking their BP etc and comparing.
Having looked at the paper, I see that predicted vs actual mmHg looks like quite a bad correlation (ie a fuzzy cloud not on the slope) That major adverse cardiac events is AUC of 0.70 which is "fair" not "good" match