JMIR Publications

ABSTRACT

Background:

Chronic kidney disease (CKD) affects approximately 37 million adults in the United States, with nearly 90% undiagnosed. Underserved populations, including African American, Hispanic, and Native American communities, develop CKD at rates 2 to 3 times higher than the general population yet face substantial barriers to timely screening and specialist care. Existing clinical prediction models rely solely on biomarkers and fail to capture the social and environmental drivers of disease progression in disadvantaged communities.

Objective:

This study aimed to (1) design and evaluate a machine learning framework that integrates clinical data with neighborhood-level social determinants of health (SDOH) to predict CKD progression from Stage 2-3 to Stage 4-5 within a 24-month horizon, and (2) simulate the projected clinical and economic impact of deploying this framework within an automated telehealth intervention workflow targeting underserved rural communities.

Methods:

A synthetic patient cohort (N=47,832) was constructed by sampling statistical distributions derived from publicly available epidemiological data, including the United States Renal Data System (USRDS) 2023 Annual Data Report, CDC PLACES 2024, and published CKD progression literature. An XGBoost classifier augmented with SHAP (SHapley Additive exPlanations) was trained and evaluated using five-fold cross-validation and a held-out synthetic external validation cohort (N=12,441). A pilot deployment simulation was conducted on a synthetic rural cohort (N=18,347 Stage 2-3 CKD patients) to project clinical and cost-effectiveness outcomes.

Results:

SDOH-CKDPred achieved an area under the receiver operating characteristic curve (AUROC) of 0.89 (95% CI 0.89-0.90) on synthetic external validation, outperforming a clinical-feature-only baseline by 7.8 percentage points (AUROC 0.81; P<.001). SHAP analysis identified eGFR slope (18.4%), baseline eGFR (15.2%), and baseline albuminuria (12.8%) as the strongest predictors, with SDOH features contributing 23% of model explanatory power. Simulated pilot deployment projected a 32 percentage-point improvement in early detection rates, a 28 percentage-point increase in nephrology referrals, and a 31.9% relative reduction in Stage 5 progression rates. The projected benefit-cost ratio (BCR) was 14.91:1 (recomputed from the current pipeline; the 3.75:1 in the original submission was stale).

Conclusions:

This simulation study demonstrates that integrating SDOH data with clinical features within an automated ML-driven workflow has the potential to substantially improve early CKD detection and reduce late-stage disease progression in underserved rural communities. Prospective clinical validation in real-world health systems is warranted as a critical next step.