Comparative performance of risk prediction models for kidney disease: an external validation using 0.5 million UK Biobank participants

To compare the external validation performance of existing kidney disease risk prediction models for the general population, individuals with type 2 diabetes, and across various subgroups. We identified and compared 16 risk prediction models for chronic kidney disease (CKD) or kidney failure from 3 recent systematic reviews (7 models for the whole population, 9 models specific for type 2 diabetes). We analysed 497,896 adults (age 38–73) in the UK Biobank data; of which 4·7% (n = 23,298) had type 2 diabetes. Models were evaluated by discrimination and calibration performance with subgroup analyses by age, sex, ethnicity and pre-existing hypertension. During a total follow-up of 5·95 million person-years (median: 12·2 years; IQR: 1·4), predictive models for people without diabetes exhibited fair-to-excellent discrimination performance (c-indices: 0·695-0·806) but severely overpredicted risk. The O’Seaghdha model demonstrated the best overall performance for discrimination (c-index: 0·806 [0·806-0·807]) and calibration (slope: 0·69, intercept: -0·011; Brier score: 0·03 [0·02,0·04]). Models including medications for diabetes showed superior performance. Discriminative performance was poorer for people with diabetes or hypertension. Severe miscalibration occurred for many models. Most models demonstrated fair to excellent discrimination for CKD and good to excellent discrimination for kidney failure. Calibration performance was frequently suboptimal; most models substantially overpredicted CKD risk while underpredicting kidney failure risk, indicating that recalibration is warranted prior to clinical application. Model performance in individuals with diabetes or hypertension was poorer. Future CKD risk prediction model development should incorporate diabetes medication use to enhance discriminative capability.