Beyond classification: a systematic review of advanced predictive methodologies in retinal image analysis

Abstract

The retina provides a unique, non-invasive window into the human microvascular and central nervous systems. Recent advancements in deep learning have catalyzed the emergence of “Oculomics” transitioning automated retinal image analysis from localized ophthalmic diagnostics to holistic systemic health assessment. This systematic review synthesizes the state-of-the-art computational pipelines, training paradigms, and explainability frameworks that enable the extraction of predictive systemic biomarkers from retinal imaging. We trace the evolution of methods from traditional pixel-wise classification to complex topology-aware graph representations. We also curated a comprehensive landscape of fundus imaging data sets. We further document the growing practice of aggregating many public data sets into large corpora for pretraining vision–language foundation models. The results are organized into four distinct views to provide a clearer perspective on the field’s development. By leveraging multivariate biobanks and advanced Foundation Models, these structural representations demonstrate remarkable accuracy in predicting cardiovascular, renal, and neurodegenerative risks, often rivaling invasive clinical metrics. Crucially, we highlight a methodological mandate for eXplainable Artificial Intelligence (XAI) in clinical translation. Techniques such as concept-based interpretability, spatial attention mapping, and human-in-the-loop validation are actively mitigating the algorithmic black box, directly linking predictive risk scores — such as retinal biological age — to clinically verifiable physiological changes. Despite persistent challenges related to data set homogeneity, cross-domain generalization, sensitivity to image quality, “residual black-box” modules, scarce longitudinal cohorts, and lack of regulatory standardization frameworks integrating robust graph-based and multimodal architectures with transparent XAI modules, positions automated retinal biomarkers as a transformative “human-in-the-loop” tool for non-invasive, population-level systemic risk stratification.