- Design
- Systematic review and meta-analysis of diagnostic studies
- Population
- 15 studies, 39,536 validation samples of renal biopsy images
- Primary outcome
- Classification performance for glomerular disease subtypes
- Effect
- Membranous nephropathy: sensitivity 0.96, AUC 0.98; IgA nephropathy: sensitivity 0.92, AUC 0.96
This meta-analysis pooled 15 studies of artificial intelligence models classifying glomerular disease from renal biopsy images, with more than 39,000 validation samples, which were not necessarily unique patients.
For the three subtypes with the most 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). In the few studies that compared AI with pathologists, AI looked more accurate, but this evidence was sparse. Most studies were retrospective and heterogeneous.
AI is not ready to sign out renal biopsies. Its near-term role is supporting pathologists, especially where renal pathology expertise is scarce, as it is in much of India.
- Treat AI renal pathology as a support tool under pathologist oversight, not a replacement.
- Expect best performance for common patterns such as membranous and IgA nephropathy.
- Ask whether any AI tool has been externally validated on local biopsies before use.
- Continue to integrate biopsy findings with serology, urine and clinical data.
Why it matters
Renal pathology expertise is scarce, and AI may extend it, but the evidence is early.
Don't overread it
Evidence that AI outperforms experienced pathologists is sparse.
The statistics, in plain English
An AUC of 0.96 means very good discrimination between disease types in test data. Because most data were retrospective and from the developers' own datasets, real-world accuracy is likely to be 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.
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