Hierarchical Sampling Improves Transformer-Based Oculomics Predictions

Global: Hierarchical Sampling Improves Transformer-Based Oculomics Predictions

A team of scientists has introduced a new hierarchical sampling technique designed to enhance retinal imaging models that predict systemic illnesses such as cardiovascular disease and dementia. The method, evaluated on a five‑year forecast of major adverse cardiovascular events (MACE) in a diverse cohort, aims to preserve patient‑specific information that conventional data augmentations often disrupt.

Background on Oculomics and Transformer Models

Oculomics leverages retinal photographs to infer health conditions beyond the eye, capitalizing on the rich vascular and structural cues captured in these images. Recent advances rely on transformer‑based foundation models, exemplified by RETFound, which demonstrate strong data efficiency for such high‑dimensional medical tasks.

Challenges with Conventional Augmentation Techniques

Standard image‑level mixed‑sample augmentations, including CutMix and MixUp, combine pixels and labels from unrelated examinations. While these techniques improve generalization for natural‑image datasets, they can inadvertently alter clinical attributes tied to individual patients—such as comorbidities or temporal disease progression—thereby compromising the integrity of medical predictions.

Introducing Oculomix: A Hierarchical Sampling Approach

The proposed Oculomix strategy incorporates two clinical priors. First, images captured from the same patient during a single exam share identical attributes; second, images from the same patient across different visits exhibit a soft temporal trend, reflecting the typical increase in morbidity over time. By constraining the mixing space to the patient and exam levels, Oculomix maintains these relationships while still providing the regularization benefits of mixed‑sample augmentation.

Empirical Evaluation and Outcomes

Researchers applied Vision Transformer (ViT) architectures to the Alzeye dataset, a large, ethnically diverse population used for five‑year MACE prediction. Compared with traditional CutMix and MixUp, Oculomix consistently delivered higher discrimination, achieving up to a 3 % increase in area under the receiver operating characteristic curve (AUROC).

Broader Significance

The findings underscore the importance of respecting patient‑level structure when augmenting medical imaging data for deep‑learning models. By aligning augmentation practices with clinical realities, the approach may accelerate the deployment of reliable oculomic tools in preventive healthcare and personalized risk assessment.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.