Nephrology · 2 h ago
Review finds limited validation of machine-learning models for CKD prognosis
A systematic review of 52 studies assessed machine-learning models predicting chronic kidney disease progression, kidney failure, or kidney replacement therapy from routinely collected health data. Although many models showed good discrimination, limited external validation, inconsistent calibration, and high risk of bias constrain their clinical applicability.
- Review included 52 studies of CKD prediction models.
- 37 studies reported discrimination estimates of at least 0.80.
- External validation and calibration were each reported in 19 studies.
- 41 studies were assessed as having high risk of bias.
Researchers searched MEDLINE, Embase, and CINAHL for studies of machine-learning models predicting CKD progression, kidney failure, or kidney replacement therapy using routinely collected healthcare data. The systematic review included 52 studies and assessed risk of bias with PROBAST. Outcomes, prediction horizons, cohorts, and validation methods varied substantially.
Of the 52 studies, 37 reported an AUROC or concordance estimate of at least 0.80. Where comparisons were available, performance gains over conventional statistical models and established risk scores, including the Kidney Failure Risk Equation, were often modest. External validation and calibration assessment were each reported in 19 studies; 41 studies were judged at high risk of bias. Kidney function, proteinuria-related measures, and age were the most influential predictors.
The findings indicate that strong reported discrimination alone does not establish readiness for clinical use. Differences in handling proteinuria and missing data, along with inconsistent model-development practices, further limit confidence that results will transfer to other settings. The review concludes that rigorous external validation, transparent methods, and evidence of real-world clinical utility are needed before implementation. The authors also note substantial heterogeneity across the included evidence.
In practice
For clinicians considering machine-learning tools to estimate CKD progression or kidney failure risk, this review does not support adopting a model based on discrimination measures alone. While 37 of 52 studies reported AUROC or concordance estimates of at least 0.80, external validation and calibration assessment were each reported in only 19 studies, and 41 studies had high risk of bias. The evidence is heterogeneous, and comparative gains over established scores such as KFRE were often modest; clinical utility remains to be demonstrated.
AI-generated clinical interpretation · Not editor-reviewed
Is this summary clinically accurate?
Help fellow clinicians: your rating sends inaccurate summaries straight to our editors.
Sign in to rate this summary →Source
Clinical Kidney Journal: Performance of machine learning models for CKD progression and kidney failure: a systematic review ↗This is an automated AI-condensed summary that has not yet been reviewed by an editor. Always consult the full item at the original source.