Retinal Age Prediction in Lithuanian Population

Abstract

Ageing changes are heterogenous, with substantial variation in health impacts of ageing across populations, individuals and tissues [Grimbly MJ, et al. BMJ Open Ophth 2024;9:e001794. doi:10.1136/bmjophth-2024-001794]. Biological ageing markers have emerged to better represent the ageing process and predict the risk of diseases. Retinal age is an imaging-based biomarker for biological age assessment from retinal fundus photographs. Accelerated ageing, or Retinal age gap (RAG) – the difference between calculated retinal age and chronological age, provides a valuable metric for quantifying the risk of vascular, neurodegenerative, metabolic diseases and even assessing the risk of mortality. However, an undisclosed controversy remains whether generated retinal age models are accurate predictors of biological age. Moreover, further clinical trials exploring their applicability across diverse healthy populations are warrant. We aimed to determine the accuracy of retinal age prediction models and evaluate their ability to reflect age-related parameters from retinal images exploring their estimation to chronological age of healthy Lithuanian population subjects. Color fundus images from 92 patients were analyzed for age prediction, allocating 80% of patients (n = 74) for training and 20% (n = 18) for validation. Images were split into training and validation sets at the patient level to avoid data leakage. Each patient contributed 8 images, except for one patient who had 6 images, resulting in a total of 734 fundus images included in the analysis. Preliminary experiments using grayscale OCTA images resulted in insufficient prediction performance, therefore the analysis focused exclusively on color fundus data. A convolutional neural network based on EfficientNet-B2 was adapted for regression by replacing the classification head with a single linear output unit predicting normalized age values. EfficientNet-B2 was selected based on findings from our previous study, in which several convolutional and transformer-based architectures were compared for classifying fundus images of healthy versus pathological images. The evaluated models in-