JMIR Publications

ABSTRACT

Background:

Chronic kidney disease (CKD) affects approximately 14% of U.S. adults, accrues over USD 89,000 per year per dialysis patient, and is strongly stratified by neighborhood-level socioeconomic deprivation. Existing CKD prediction models commonly rely on private electronic-health-record cohorts or synthetic data, which limits independent replication. We constructed and validated a fully open, two-stage CKD surveillance and screening pipeline using only publicly available data sources, with all code, random seeds, and pipeline scripts archived for replication.

Objective:

To (1) estimate small-area CKD prevalence from the Area Deprivation Index (ADI) calibrated against CDC PLACES, and (2) build a patient-level CKD classifier from National Health and Nutrition Examination Survey (NHANES) data using only non-kidney clinical variables. Both models are evaluated with calibration assessment, geographic generalization, decision-curve analysis, and survey-weighted bootstrap uncertainty estimation suitable for population-health deployment.

Methods:

Stage 1 (ecological model) trains an XGBoost regressor on 60,445 U.S. census tracts where the 2020 University of Wisconsin Neighborhood Atlas ADI block-group file (n=242,335 block groups, aggregated to tracts by population-weighted mean) intersected the CDC PLACES 2022 tract-level CKD prevalence release (n=72,337 tracts; BRFSS 2020 underlying data). The outcome was the observed PLACES CKD crude prevalence; features were tract-mean ADI national rank, ADI state rank, and ADI quintile. Performance was assessed with random 5-fold cross-validation, leave-one-Census-region-out cross-validation, calibration deciles, capacity sensitivity sweep, and three feature-set ablations (raw ADI components; adding ACS poverty/education; adding spatial lag). A boundary-misalignment sensitivity analysis used the Census 2010-to-2020 tract crosswalk to recover ~12,400 previously-excluded tracts. Stage 2 (patient-level classifier) trains an XGBoost classifier on NHANES 2017–March 2020 pre-pandemic and 2021–August 2023 cycles (n=15,150 adults aged ≥18 with serum creatinine or urine albumin available; survey-weighted using NHANES MEC weights divided by 2 for combined-cycle weighting per CDC guidance). The outcome was a KDIGO-2024 CKD label (eGFR<60 mL/min/1.73 m² by CKD-EPI 2021 race-free OR urine albumin-to-creatinine ratio ≥30 mg/g). Features included demographics, body mass index, oscillometric blood pressure (mean of up to three readings), and self-reported hypertension, diabetes, heart failure, and stroke; no kidney biomarkers were used. We computed survey-weighted AUROC, AUPRC, Brier score, calibration slope, decision-curve net benefit, and subgroup AUROC by age, sex, and race/ethnicity, with 1,000-replicate bootstrap 95% confidence intervals throughout. The classifier was benchmarked against a regularized Logistic Regression on the same features; the Tangri KFRE was excluded from comparison because it requires the kidney biomarkers our classifier deliberately omits.

Results:

Stage 1: Cross-validated R² = 0.450 ± 0.006, mean absolute error = 0.45 percentage points (Figure 1, Table 2). Leave-one-Census-region-out R² ranged from 0.239 (West) to 0.415 (South); calibration slopes ranged 0.71–1.14 (Figure 3). Predicted CKD prevalence was monotonically associated with ADI quintile (Figure 5; Spearman rank correlation overall = 0.71). Three feature-set ablations are reported in Supplementary Table S2: raw ADI components increase R² to 0.468; adding ACS poverty/education to 0.472; adding spatial lag to 0.488 (the latter is reported as an upper bound rather than a recommended design). Capacity sweep showed R² varying by < 0.005 across five hyperparameter settings (Figure 4). The boundary-misalignment sensitivity analysis (Supplementary Table S1) confirmed that the original analytic decision is robust: recovering ~12,400 excluded tracts via the 2010-to-2020 Census crosswalk changes Stage 1 R² by < 0.01. Stage 2: survey-weighted AUROC = 0.773 (95% bootstrap CI 0.756–0.789; per-fold range 0.760–0.800; Figure 8), AUPRC = 0.404 (95% CI 0.371–0.438; baseline 0.139; Figure 9), Brier score = 0.102 (95% CI 0.098–0.106). Calibration slope was 0.955 with intercept +0.015 across decile-binned predictions (Figure 10). Sensitivity at 90% specificity was 45.4% (threshold 0.288); the operating-point trade-off across alternative specificity targets is reported in Supplementary Table S4 (sensitivity 70.4% at 70% specificity, threshold 0.139). Decision-curve net benefit exceeded both treat-all and treat-none strategies across thresholds 0.05–0.40 (Figure 12). Subgroup AUROC ranged 0.637–0.819 with 95% bootstrap CIs reported in Figure 11. The Logistic Regression baseline on identical features achieved AUROC 0.756 (95% CI 0.738–0.773); the XGBoost classifier outperformed it by 0.017 AUROC points (DeLong's z = 2.91, p = 0.004). Missing-data proportions for each predictor are tabulated in Supplementary Table S3; multiple-imputation sensitivity analysis changed survey-weighted AUROC by < 0.001.

Conclusions:

A fully open, reproducible two-stage pipeline for CKD surveillance and individual screening can be constructed from public datasets alone. The ecological model identifies high-burden neighborhoods with calibrated R² ≈ 0.45 against CDC ground truth and a Spearman rank correlation of 0.71, supporting tract-level prioritization for resource allocation. The patient-level classifier achieves AUROC ≈ 0.77 without using kidney biomarkers, supporting its use as a screening triage tool that does not require kidney-biomarker laboratory infrastructure. All code, data sources, and random seeds are documented to enable replication and external validation. Prospective deployment evaluation in partnership with a state health department is in progress (Supplementary Material S5).