In a new study published in the Asia-Pacific Journal of Ophthalmology, researchers explored the use of artificial intelligence (AI) in oculomics for assessing cardiovascular risk factors through retinal imaging, specifically focusing on predicting HbA1c levels. The study demonstrates how AI could change healthcare, especially for diabetes management, by providing non-invasive methods to estimate this critical biomarker.
Oculomics, the study of how eye-related data can reflect systemic diseases, is gaining traction in medical research. This pilot study used a dataset of over 6,000 fundus images, taken from patients with varying HbA1c levels—categorized as normal, pre-diabetic or diabetic. The researchers experimented with different AI model architectures to estimate HbA1c levels and explored how the AI could be fine-tuned to improve accuracy.
The study revealed that deep learning models like VGG19 achieved an accuracy of about 60.87%, while an ensemble model improved performance to 62.73%. These results, though promising, demonstrate that further refinements are necessary to enhance the reliability of these tools before they can be widely adopted in clinical settings.
The research highlights several challenges that must be addressed for AI-based oculomics models to be trusted by healthcare professionals. The study observed significant biases in performance based on the age and sex of the participants, raising concerns about the model’s generalizability across different patient demographics. For instance, younger patients with higher HbA1c levels showed distinct differences in model accuracy compared to older patients. Additionally, the performance of the AI in differentiating between male and female fundus images led researchers to identify potential bias in the dataset that may affect predictions.
To address these concerns, the study emphasizes the need for a larger, more diverse dataset and the inclusion of more transparent and interpretable AI models. Techniques such as Grad-CAM (Gradient-weighted Class Activation Mapping) were used to make the AI model’s decision-making process more interpretable, showing which parts of the retinal images were most relevant to the AI’s predictions. This transparency is key to increasing trust in AI-driven medical tools.
Despite these challenges, the potential for oculomics to provide a rapid, non-invasive means of assessing cardiovascular risk factors is particularly promising for underserved populations that may lack access to specialized care. By leveraging retinal imaging, patients in remote or rural settings could receive better diagnostic services without the need for blood tests or invasive procedures.
The next steps in this research will focus on addressing the limitations of the current AI models, ensuring they are more robust and less biased. Improving model transparency and incorporating larger, more representative datasets will be key to achieving reliable, trustworthy AI in ophthalmology.
Disclosures: This article was created by the touchOPHTHALMOLOGY team utilizing AI as an editorial tool (ChatGPT (GPT-4o) [Large language model]. https://chat.openai.com/chat.) The content was developed and edited by human editors. No funding was received in the publication of this article.
