Design: Systematic review and meta-analysis of prediction models
Population: 66 machine-learning models for diabetic kidney disease
Primary outcome: Discrimination (AUC), sensitivity and specificity
Effect: Pooled AUC 0.926 (95% CI 0.898–0.947); I² above 85%

This meta-analysis pooled 66 machine-learning models designed to identify diabetic kidney disease. The pooled AUC was 0.93, with sensitivity 0.83 and specificity 0.90; neural networks performed best on paper.

Heterogeneity was very high, and most models lacked external validation and calibration, and were poorly reported. Model type and study design both affected the results.

Simple, validated tools such as eGFR and urine albumin-to-creatinine ratio remain the basis for detecting diabetic kidney disease. These models are not yet ready to replace or supplement them in clinic.

Screen people with diabetes annually with eGFR and urine albumin-to-creatinine ratio.
Do not rely on unvalidated machine-learning risk scores for diagnosis.
Ask for external validation and calibration data before adopting any prediction tool.
Use established risk equations where you need to estimate progression.

Why it matters

High accuracy figures in development studies often shrink in real use.

The statistics, in plain English

An AUC of 0.93 suggests good discrimination, but without external validation it may reflect overfitting to the development data. Calibration, whether predicted risks match real ones, was rarely reported.