Researchers evaluated the feasibility, accuracy, and acceptability of an automated retinal photography system paired with artificial intelligence for cardiovascular disease risk assessment.
In the study, published in npj Digital Medicine, they found that the AI-derived risk scores were comparable in predictive accuracy to traditional World Health Organization cardiovascular disease risk assessments, and both patients and general practitioners reported satisfaction.
The researchers enrolled 361 adult patients aged 45 to 70 years without a history of CVD from 2 general practice clinics in Victoria, Australia. The participants underwent nonmydriatic retinal imaging using an automated fundus camera (Mediworks FC162), which generated AI-based retinal photographic CVD (rpCVD) risk scores in near real time.
The rpCVD algorithm, which was developed using UK Biobank data, utilized retinal image features and demographic inputs to predict 10-year CVD risk. Each participant also underwent a conventional WHO CVD risk assessment, which was used as the benchmark.
The researchers noted that rpCVD risk scores had a 93.9% success rate, which was measured by at least one gradable image. The median session time from initiating the image capture to printing the QR code for the report was 1 minute 47 seconds (interquartile range = 1:30–3:21). Over half (65.9%) of the participants successfully had images that met quality standards on the first attempt.
Using UK Biobank data, the researchers found that the predictive accuracy of 10-year incident CVD for rpCVD was comparable to the WHO CVD risk score. Scores were also comparable for 10-year coronary heart disease or stroke alone.
Risk category agreement between rpCVD and WHO CVD scores was 63.4%. Overestimation occurred in 17.1% of cases and underestimation in 19.5%. Older male participants with diabetes or a history of smoking were more likely to have underestimated risk, whereas younger female participants who didn't smoke were more likely to have overestimated risk.
Among the participants who responded to the satisfaction survey, 95% found the system easy to use, and 92.8% said the output was easy to understand. Most (92.5%) expressed overall satisfaction. Most clinicians (87.5%) also said they were satisfied with the system and would be likely to continue to use the tool if it integrated into their workflows. While 62.5% thought the tool would increase willingness to perform CVD risk assessments, the lab-based WHO model was still the most preferred by 75% of clinician respondents.
The researchers acknowledged that the algorithm was trained using a relatively healthy UK Biobank population and with a different camera system (TOPCON 3D OCT vs Mediworks FC162), which may have limited generalizability. They suggested further algorithm refinement, as well as real-world longitudinal outcome tracking and improved model calibration for high-risk groups to “support its wider clinical deployment as a convenient and noninvasive approach for CVD risk assessment in primary care.”
No competing interests were declared.
