A recent study highlighted a pilot project aimed at using artificial intelligence (AI)-based oculomics—the integration of ophthalmic features to develop biomarkers for systemic diseases—to assess cardiovascular risk factors through fundus imaging.

At the time of their study, deep learning algorithms have been developed to diagnose a range of ocular conditions including diabetic retinopathy, age-related macular degeneration, multiple sclerosis, and retinopathy of prematurity, along with analysis of candidates for complex surgical procedures, the researchers note in their recent article, published in the Asia-Pacific Journal of Ophthalmology.

The study focused on estimating HbA1c levels—a critical biomarker for diabetes and cardiovascular disease—from retinal images. The data-driven model compared multiple deep learning architectures, including VGG19 and ResNet-50. The VGG19 model demonstrated superior accuracy in estimating HbA1c levels from retinal images and its performance showed potential for non-invasive HbA1c estimation. VGG19’s success highlighted the potential for fundus imaging as a non-invasive tool for diabetes management and cardiovascular risk assessment.

Still, the researchers identified age and sex as contributing factors to bias in AI models, which could impact prediction accuracy. This limitation emphasized the importance of transparency and dataset diversity in AI model development.

The authors also discussed the use of Grad-CAM and ensemble models to improve model performance and reliability, specifically regarding better visualization of areas that are critical for the model's decision-making. However, they noted that ensemble models add complexity when determining model reliability and robustness, which could impact patient health and safety. They suggested analyzing reliability in multiple ways, rather than only using the performance of a single test set to consider the true generalizability of the model chosen.

Finally, the researchers discussed the importance of physicians avoiding overreliance on AI technology as they become more comfortable and confident in its abilities.

A full list of author disclosures can be found in the published research.