From fundus to filtration: AI-driven retinal phenotyping as a framework for non-invasive prediction of kidney pathological categories (the "virtual renal biopsy" concept)- a narrative review.

BACKGROUND: The burden of chronic kidney disease (CKD) continues to grow, affecting over 850 million individuals globally. Current methods to precisely and definitively classify CKD still rely on renal biopsy, a procedure with a 5.1% rate of major complications, and further cost and access complications in low-resource settings. The retina, containing the only accessible microvasculature for direct and non-invasive visualization, shares deep developmental and structural features with, and biologically plausible pathogenic overlap with, the kidney (see Sect. 3 for the distinction between established and hypothesized components of this relationship). Oculomics, an emerging discipline that incorporates AI to analyze retinal images to identify systemic disease, has shown promising results for binary CKD screening, with AUC scores of 0.83-0.93; this is distinct from, and should not be conflated with, the prediction of specific renal pathological categories discussed below. A 2025 study introducing the Kidney Intelligent Diagnosis System (KIDS) was among the first to demonstrate that specific renal pathological categories-IgA nephropathy, idiopathic membranous nephropathy, arterionephrosclerosis, diabetic nephropathy, and a combined idiopathic minimal change disease/focal segmental glomerulosclerosis category-could be predicted from retinal images using five separate binary prediction tasks, reporting AUC values of 0.790-0.932 (internal and, for a subset, external validation; hybrid image-plus-clinical-data models performed at the higher end of this range). This is best regarded as an early proof-of-concept for a non-invasive risk-stratification tool, conceptually described in this review as a 'virtual renal biopsy'-a metaphor for a proposed research framework, not a claim of diagnostic equivalence with tissue histopathology (Meng et al.'s DeepDKD model had earlier used retinal images to distinguish biopsy-defined diabetic from non-diabetic kidney disease, so KIDS is best described as an early multi-category model rather than the first retinal-AI study in this space). AIM: This narrative review aims to consolidate various biological, technological, and clinical components of AI-driven retinal phenotyping and outline the path toward non-invasive prediction of kidney pathological categories. This includes the aspects of continuous progression from binary screening of CKD towards predicting specific histopathologies while considering the current state of explainability of AI, the performance metrics beyond AUC that are required for clinical deployment, and offering a detailed clinical research roadmap. METHODS: Narrative synthesis was utilized to integrate data on CKD and deep learning, retinal vascular research, research on retinal-renal connections, AI explainability, and the KIDS model. A structured search of PubMed/MEDLINE, Embase, Scopus, and Google Scholar was performed for English-language, peer-reviewed original studies, systematic reviews, and clinical guidelines relevant to retinal-renal biology, oculomics, and AI-based renal pathology prediction; representative search terms, eligibility criteria, and the handling of preprints are detailed in Sect. 2. The data were organized by biological plausibility, evolution of modeling, integration of imaging technologies, metrics required for clinical validation, and research in translational barriers. CONCLUSIONS: The combination of the biological plausibility of a retinal signature of renal injury, the advances in oculomics, AI, and early multi-category models such as KIDS lays a preliminary foundation for research toward non-invasive prediction of kidney pathological categories. In order to advance this research agenda, numerous multidisciplinary clinical studies are required along with the integration of different imaging technologies, rigorous reporting of calibration and clinically relevant performance metrics, and the creation of ethical and regulatory frameworks. The 'virtual renal biopsy' concept will not replace histopathology and is not yet supported by evidence sufficient for clinical deployment; if the underlying research agenda is executed rigorously, it may in time provide an adjunctive, non-invasive risk-stratification tool for the hundreds of millions of patients who do not have access to renal histopathological diagnosis.