Neurovascular retinomics for predicting heart diseases

To investigate the use of retinomics as a composite biomarker for incident heart diseases, leveraging the multidimensional relationships between vascular parameters from colour fundus photography (CFP) and neural parameters from optical coherence tomography (OCT).

Feature selection was performed using LASSO regression to minimize overfitting and multicollinearity. Subsequently, the predictive value of retinomics was assessed with the Gradient Boosting Machine model and compared with the widely used WHO-CVD risk score. Associations between retinomics and incident heart diseases were analysed using Cox regression. 39 450 participants, 3310 developed heart disease over 11.4 years. Retinomics with age and sex achieved a concordance index (C-index) of 0.731 for the entire follow-up duration and a C-index of 0.761 for 5-year incident events. For the 5-year follow-up, the high-risk group (the third tertile) had a hazard ratio of 9.30 compared with the low-risk group. Compared with the WHO-CVD risk score (C-index: 0.708), the retinomic model demonstrated significantly improved discrimination (C-index: 0.736; P < 0.001). Association analysis revealed 17 retinomic features associated with the risk of incident heart diseases, even after adjusting for age, sex, and the presence of hypertension, diabetes, and obesity.

Our simple model incorporating neurovascular retinomics with age and sex demonstrated good predictive performance and effective risk stratification for incident heart disease. Notably, the age at which heart disease risk increases coincide with the age at which people get eye exams due to age-related ocular conditions, making this simple model a low-cost, timely, non-invasive, and highly feasible screening tool, especially in resource-limited settings.