AI classified common glomerular diseases accurately on retrospective data

Design

Systematic review and meta-analysis of diagnostic accuracy

Population

15 studies, 39,536 validation units of renal pathology images

Primary outcome

Sensitivity, precision, accuracy and AUC for glomerular disease subtypes

Effect

AUC 0.98 membranous, 0.96 IgA, 1.00 minimal change disease

A meta-analysis in Renal Failure (6 October 2026) pooled 15 studies of pathology-image AI models for glomerular disease, with 39,536 validation units (not necessarily separate patients) and pathologist diagnosis as reference.

For the subtypes with enough data, performance was high: membranous nephropathy (sensitivity 0.96, AUC 0.98), IgA nephropathy (sensitivity 0.92, AUC 0.96) and minimal change disease (sensitivity 0.92, AUC 1.00). Some studies reported AI accuracy above senior pathologists, but such comparisons were few. Most studies were retrospective and heterogeneity was substantial.

These tools look promising for screening and quantification, but they were trained and tested against the very diagnosis they replicate, mostly on single-source images. They are not yet a substitute for integrated light, immunofluorescence and electron microscopy reporting.

• Treat AI glomerular classifiers as research tools pending prospective validation
• Check whether any vendor model was validated externally on your staining and scanners
• Keep immunofluorescence and electron microscopy in the diagnostic workup
• Consider AI for quantification tasks under pathologist oversight

Why it matters

Renal pathology expertise is scarce, and AI could support it - once proven prospectively.

Don't overread it

Sparse head-to-head data - this does not show AI outperforms experienced renal pathologists.

The statistics, in plain English

AUC is the probability a model ranks a true case above a non-case; 0.96-1.00 is very high. But validation units were not unique patients, studies were retrospective, and heterogeneity was large, so real-world performance is likely lower.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.