AI-Based ZEBRA Model May Improve Detection of Fabry Nephropathy on Kidney Biopsy

An artificial intelligence (AI)–assisted model may support earlier detection of Fabry disease (FD), particularly in patients with Fabry nephropathy (FN), according to findings from a new study.1

The findings are derived from the ZEbra Bodies Recognition by Artificial Intelligence (ZEBRA) pipeline, which investigators describe as a potential new histological measure of podocyte involvement in FN.1

“AI applications have been explored for tasks such as glomerular classification, segmentation, and quantification of histological features, aiming to augment traditional pathology workflow,” wrote investigator Vincenzo L’Imperio, MD, from the University of Milano-Bicocca. “In the context of FN, the diagnostic challenge is represented by the insidious clinical presentation that can be totally unspecific.”1

FD is a rare lysosomal storage disorder caused by intracellular globotriaosylceramide (Gb3) accumulation, which can lead to progressive organ damage. Renal involvement, FN, is a major contributor to morbidity and mortality and may present with proteinuria, declining kidney function, and, in advanced stages, end-stage renal disease.

Despite these risks, FN is often difficult to diagnose, particularly in its early stages. Clinical presentation may be nonspecific or clinically silent, especially in patients with late-onset or atypical variants and in female patients, in whom X-chromosome inactivation can attenuate disease expression. As a result, kidney involvement may go unrecognized until irreversible damage has occurred.

Renal biopsy remains a cornerstone of early FN detection, allowing identification of Gb3-related podocyte changes on light and electron microscopy before overt systemic manifestations develop. However, diagnosis relies heavily on nephropathologist expertise, as hallmark features such as podocyte cytoplasmic vacuolization can be subtle, focal, or unevenly distributed, particularly in atypical or female cases. Against this backdrop, AI-assisted digital pathology has emerged as a potential adjunct to renal biopsy interpretation.1,2

To assess whether AI could assist in identifying early FN-related changes, L’Imperio and colleagues conducted a multicenter analysis of formalin-fixed, paraffin-embedded renal biopsies collected over a 15-year period from patients with genetically confirmed FD. Investigators trained multiple AI-based classification and segmentation models to detect foamy podocytes at the glomerular level.1

Building on this framework, the team developed a custom inference pipeline, ZEBRA, designed to streamline glomerular annotation and quantify podocyte involvement in a reproducible manner. The ZEBRA algorithm integrates a glomerular-level classification component with a segmentation module that delineates affected podocytes at the microscopic level.

Using the ZEBRA model, investigators calculated the proportion of glomerular area occupied by foamy podocytes (fpA) relative to total glomerular area (tgA), generating the ZEBRA score (ZS = fpA/tgA, expressed as a percentage).1

The study included 77 participants, divided into FN cases (n = 37) and controls (n = 40). All patients underwent kidney biopsy based on standard nephrological indications, such as persistent proteinuria (>0.5–1 g/day), hematuria, unexplained decline in kidney function, or other laboratory or clinical findings, regardless of whether FD was suspected at the time of biopsy.1

In total, investigators analyzed 1,075 glomeruli from FN cases and 689 from controls. Compared with controls, FN cases had higher estimated glomerular filtration rate (mean 98 vs 69 ml/min/1.73 m²; P <.001), lower proteinuria (median 0.2 vs 1.28 g/day; P <.001), and milder interstitial fibrosis and tubular atrophy (P = .01), underscoring the early-stage nature of kidney involvement in many FN cases.1

Upon receiver operating characteristic analysis, investigators observed strong diagnostic performance for the ZEBRA score, with an area under the curve of 0.93. A preliminary cutoff of 0.19 (mean ZS per case) yielded a sensitivity of 0.91 and a specificity of 0.81. The ZEBRA score also correlated with the manual podocyte vacuolization score in both hematoxylin and eosin–stained (rs = 0.66) and periodic acid–Schiff–stained slides (rs = 0.71), although case-level discrepancies were noted.1

“This study introduces and validates a novel digital pathology pipeline for FN, culminating in the ZEBRA score, a new histological measure of podocyte involvement,” investigators concluded. “The AI system is designed as a high-sensitivity screening tool, providing comprehensive image analysis that can alert human operators to potential oversights and help prevent missed diagnoses.”1

Investigators emphasized further studies are needed to evaluate the prognostic significance of the ZEBRA score and to determine its potential role in informing clinical decision-making in FD.1

References

1. Cazzaniga G, Carbone M, Barretta R, et al. Zebra bodies recognition by artificial intelligence (ZEBRA): a computational tool for Fabry nephropathy. Scientific Reports. Published online January 12, 2026. doi:https://doi.org/10.1038/s41598-026-35466-w

2. Del Pino M, Andrés A, Bernabéu AÁ, et al. Fabry Nephropathy: An Evidence-Based Narrative Review. Kidney and Blood Pressure Research. 2018;43(2):406-421. doi:https://doi.org/10.1159/000488121