AI classifies common glomerular diseases well, but not yet against pathologists

Design

Systematic review and diagnostic accuracy meta-analysis

Population

15 studies, 39,536 renal pathology validation units

Primary outcome

Classification accuracy for glomerular disease subtypes

Effect

Membranous sensitivity 0.96, AUC 0.98; IgA sensitivity 0.92, AUC 0.96

This systematic review and meta-analysis included 15 studies with 39,536 validation units, which were not necessarily separate patients, testing AI models that classify glomerular disease from renal biopsy images against the pathology diagnosis.

For membranous nephropathy, AI reached sensitivity 0.96 and AUC 0.98; for IgA nephropathy sensitivity 0.92 and AUC 0.96; and for minimal change disease sensitivity 0.92. Some studies reported AI accuracy above that of senior pathologists, but the comparator data were sparse. Most studies were retrospective, and heterogeneity across datasets, imaging methods and validation was substantial.

Renal pathology depends on light, immunofluorescence and electron microscopy together, and these models mostly use one modality. The results support AI as a screening or quantitation aid under pathologist oversight, not as a replacement. Laboratories with limited nephropathology expertise, common outside large Indian centres, may be the first to benefit once prospective validation exists.

• Treat renal biopsy AI tools as decision support; current evidence does not show they beat experienced pathologists.
• Ask whether any tool has been validated on external data similar to your own staining and scanning.
• Keep immunofluorescence and electron microscopy central to the final diagnosis.
• Expect weaker performance for less common glomerular diseases, which had fewer validation datasets.

Why it matters

Vendors may cite these accuracy figures, and the head-to-head evidence behind them is thin.

Don't overread it

Validation units were not unique patients and most studies were retrospective, which tends to inflate accuracy.

The statistics, in plain English

Sensitivity is the share of true cases the model caught; AUC summarises how well it separates cases from non-cases, with 1.0 perfect. High heterogeneity means results varied widely between studies, so pooled figures may not apply to any one laboratory.

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.