- Design
- Systematic review and meta-analysis of 15 diagnostic studies
- Population
- 39,536 validation samples from renal pathology images
- Primary outcome
- Classification accuracy for glomerular disease subtypes
- Effect
- Membranous: sensitivity 0.96, AUC 0.98; IgA: 0.92, 0.96; MCD: 0.92, 1.00
A meta-analysis pooled 15 studies of artificial intelligence models that classify glomerular disease from renal pathology images, comprising 39,536 validation samples (not necessarily separate patients).
For the three diseases with enough data, performance was high: membranous nephropathy (sensitivity 0.96, AUC 0.98), IgA nephropathy (0.92, 0.96) and minimal change disease (0.92, 1.00). Some studies reported higher accuracy than senior pathologists, but such comparisons were few. Most studies were retrospective, and datasets, imaging methods and validation approaches varied widely.
These tools may help in settings with few renal pathologists, which describes much of India, as a screening or second-reader aid. They are not ready to replace a pathologist's integrated reading with immunofluorescence, electron microscopy and the clinical picture.
- AI performed well for membranous nephropathy, IgA nephropathy and minimal change disease.
- Head-to-head comparisons with pathologists were sparse.
- Treat AI output as a second reader, not a final diagnosis.
- Integration with clinical and immunofluorescence data is still needed.
Why it matters
Access to expert renal pathology is limited in many regions, and validated AI could shorten the wait for a diagnosis.
Don't overread it
Retrospective, heterogeneous studies; superiority over pathologists is not established.
The statistics, in plain English
An AUC of 0.98 means the model almost always ranks a true case above a non-case in the test data. Performance in a new hospital, with different staining and scanners, is usually lower, which is why external validation matters.
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.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for nephrology, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free