AI-driven oculomics assesses HbA1c for cardiovascular risk

Integrating AI with oculomics shows promise in predicting diabetes, with the study revealing that diverse datasets can help develop more accurate and trustworthy models to assess HbA1c levels, enhancing patient care in clinical settings.

Study: Development of oculomics artificial intelligence for cardiovascular risk factors: A case study in fundus oculomics for HbA1c assessment and clinically relevant considerations for clinicians. Image Credit: Stas Ponomarencko / Shutterstock.com

Oculomics is a relatively new and powerful technology that integrates ophthalmic features to identify biomarkers that can be used to predict systemic diseases. A recent Asia-Pacific Journal of Ophthalmology pilot study investigates the potential advantages of incorporating artificial intelligence (AI) with oculomics.

Using oculomics for HbA1c estimation

The objective of the current study was to apply oculomics to fundus images to assess glycated hemoglobin A1c (HbA1c) levels.

The percentage of HbA1c is typically used to diagnose and monitor the progression of diabetes. Despite its widespread use, HbA1c measurements can be inaccurate in some instances, such as in patients with comorbidities like sickle cell anemia, those who have recently received a blood transfusion, or pregnant patients.

In the current study, a total of 6,118 fundus images were obtained, 1,138 of which were diagnosed as normal. Initially, the researchers compared the performance of a monolithic model to ensemble architecture and its reliability and bias from the effects of age and sex.

The VGG19 model, which is based on convolutional neural networks (CNN), exhibited the best performance across all metrics. Compared to the single model, the ensemble model showed a performance improvement of about 2%.

For the model to be reliable, its size as compared to the dataset size must be carefully considered. In the current study, the VGG19 model outperformed other larger models, which may be due to the dataset size, as complex and more extensive models require larger amounts of data to accurately estimate parameters.

It is also crucial to evaluate reliability along other dimensions than just performance on a single testing set, as safety-critical applications of AI could significantly impact patient health and safety.

The effects of age and gender on model performance

Higher accuracy was observed in the model that was trained on samples obtained from both youth and seniors as compared to those where the training set comprised a single group. While developing AI solutions, the ability of a model to show robust and reliable performance across a diverse population is crucial for minimizing bias.

The performance was best when the training set included both sexes. In fact, a 5% degradation was observed when the model was trained on either males or females alone. An additional model was trained to output gender from fundus images.

The model was associated with an overall accuracy of 87%; however, this exceptional performance could be attributed to potential bias in the training dataset. Grad-CAM, a well-known interpretable AI technique, was used to identify key features of the fundus images that provide important insights for various classification labels. These results complemented those produced in previous studies and emphasized the need for a diverse dataset to deliver reliable and robust performance.

Challenges in developing trustworthy AI in oculomics

The study findings confirm the importance of high-quality and diverse datasets to train models, which will ultimately improve their robust and reliable performance in different conditions. Model outputs should also be transparent to ensure healthcare providers can understand and trust their predictions. Ensuring fairness and addressing bias in model predictions is also extremely important.

The future of AI in oculomics

Adaptability to diverse clinical environments and compliance with regulatory guidelines is fundamental to ensure the trustworthy deployment of AI in oculomics. Maintaining transparency also supports the development of explainable and interpretable AI models in which medical practitioners can easily understand the logic of their predictions.

AI models may experience performance degradation due to out-of-distribution (OOD) inputs or unforeseen clinical scenarios during deployment. However, incorporating continuous learning frameworks could mitigate this issue.

Anomaly detection algorithms may also serve as a safeguard; therefore, models should be periodically updated to introduce novel and diverse data. These efforts have the potential to sustain the accuracy and relevance of AI applications in clinical environments.

In the future, AI systems for oculomics must be developed to enhance human-centered healthcare by simplifying clinicians’ workflows, improving patient health outcomes, and raising the quality of care provided. To this end, ongoing collaboration among various stakeholders is necessary throughout the development and implementation phases to reach AI’s full potential in transforming healthcare delivery.