Fig. 3: Clinical and protein risk models for time to death within 5 years.

Data are presented as AUC +/- standard error for risk models predicting time to death occurring within 5-years in participants who have undergone hemodialysis ≥ 1 year(prevalent time period). Top panel: Variable selection and model fitting were performed in CRIC 80% training set, and models were validated in PACE(whole cohort). CRIC AUCs are shown for the 20% testing set (N = 87). CRIC calibration N = 349, PACE N = 230. Bottom panel: Variable selection and model fitting were performed in pooled cohort 80% training set. AUCs are shown for the pooled cohort 20% testing set (N = 133). Calibration was performed in the training sets: N = 530 for clinical models, N = 533 for protein models. The Comorbidity model is based on the index developed by Liu et al.¹⁷ and includes diabetes, myocardial infarction, stroke, peripheral arterial disease, chronic obstructive pulmonary disease, arrhythmia, malignancy, gastrointestinal bleeding and heart failure. The Clinical model is based on the model developed by Wagner et al.¹⁸ and includes age, sex, self-reported race, diabetes, etiology of kidney failure, CVD, diabetes x CVD, smoking, and serum creatinine, hemoglobin, albumin, and calcium. The Cohort Clinical was developed by screening comorbidity and other clinical variables in the pooled cohort, and includes age, log(serum albumin), hemoglobin, log(time since dialysis initiation), triglycerides, BMI, log(PTH), phosphate. The 3-protein model includes SVEP1, R-spondin 4 and tetranectin. Model coefficients are listed in Supplementary Data 15–17. One-sided p-values were obtained with paired bootstrapping t-tests. P-values < 0.001 indicate that, in the 1000 bootstrap replicates, the model AUC exceeded the comparator. Source data are provided as a Source Data file.