OR WAIT null SECS
An analysis of the AlzEye study presented at ARVO 2023 suggests a deep-learning model showed promise in predicting three-year incident MACE using OCT.
An in-house pre-training strategy significantly outperformed the baseline supervised strategy, according to an analysis of the AlzEye study presented at the 2023 Association for Research in Vision and Ophthalmology (ARVO) Annual Meeting in New Orleans.
“The potential for using retinal scans to predict some of the biggest causes of death is exciting because eye scans are safe, affordable, and widely available,” said Mark Chia, MBBS, University College London Institute of Ophthalmology.2 “We might one day reach a point where a patient can be identified as ‘high-risk’ during a routine glasses check, which could lead them to having the necessary treatments to prevent a heart attack or stroke.”
Recent literature has indicated the crucial role of the retina outside the realm of vision, providing insight into health disorders, including cardiovascular disease (CVD). As CVD is the leading cause of death globally, Chia and colleagues noted it represented a promising area of research. However, there have been few studies using the 3-dimednsionla imaging capabilities of OCT to further explore the link.
As a result, the investigative team developed a deep learning model to predict the 3-year incidence of MACE using retinal OCT. The team collected and analyzed data from AlzEye, a retrospective cohort study linking retinal imaging of patients aged ≥40 years with systemic disease data from hospital admissions between January 2008 – March 2018.
The study defined MACE as ischemic stroke, myocardial infarction, heart failure, and atrial fibrillation, according to the International Classification of Disease, 10th revision. For the analysis, investigators used the left retinal OCT from a single visit for each patient and split the dataset into train, validation, and test sets in the ratio 55:15:30.
Then, two models were developed using pre-training strategies for comparison: an in-house strategy trained on unlabeled natural and retinal images; a baseline strategy trained on labeled natural images from ImageNet-21k. Investigators performed fine-tuning on the AlzEye training set and evaluated on the internal test set in both models.
Based on the findings, Chia and colleagues suggest the model performed “reasonably well” at predicting MACE within 3 years, adding that expanding this work to other populations may have potential.Within AlzEye, the analysis showed 5,382 patients had a MACE incident within 3 years of undergoing retinal OCT.
The primary model using the investigative team’s in-house strategy achieved an area under the receiver operating characteristic (AUROC) of 0.796 (95% CI, 0.795 - 0.797) and an area under the precision recall (AUPR) of 0.769 (95% CI, 0.768 - 0.771). Moreover, data showed the F1-score and sensitivity was 0.713 (95% CI, 0.711 - 0.714) and 0.700 (95% CI, 0.699 - 0.702), respectively.
Investigators indicated the in-house pre-training strategy significantly outperformed the supervised baseline strategy. Comparatively, the supervised baseline strategy achieved an AUROC of 0.743 (95% CI, 0.741 - 0.744), AUPR of 0.713 (95% CI, 0.710 - 0.715), F-1 score of 0.660 (95% CI, 0.659 - 0.662), and sensitivity of 0.637 (95% CI, 0.635 - 0.639).
Chia and colleagues suggested subsequent research “will explore multimodal prediction models, external validation of our findings, and the potential for clinical translation.”