JMIR Publications

ABSTRACT

Background:

In the United States (U. S. ), 37 million adults have Chronic Kidney Disease (CKD) and nearly 90% of them are unaware of their condition. Additionally, 120 billion USD are spent by the U. S. Medicare system each year for over 120 billion USD sites substantial medical expenses associated with it. CKD detection is primarily reliant on the symptomatic complaints of patients which creates a reactive situation for existing health care system. This problem is magnified in underserved areas because key communities such as African Americans, Hispanics, and Native Americans develop CKD at rates 2-3 times higher compared to that in the broader community.

Objective:

This study aims to develop an AI system that considers medical data in combination with socio-demographic and Social Determinants of Health (SDOH) factors to accurately predict the risk of developing Stage 2-4 CKD for residents in underserved populations. This allows proactive interventions to be undertaken before symptoms manifest.

Methods:

Mining of Electronic Health Records (EHR), integration of Social Determinants of Health (SDOH) data, environmental data and insurance claims sources are used to construct the model. An XGBoost model is employed for prediction and a relevant set of SDOH features is incorporated for model explanation via the SHAP technique. The model forecasts the onset of late-stage CKD (i. e. , Stage 5 or Stage 4) for targeted cohorts presently diagnosed with Stage 2 or 3 of CKD within a 24 months' timeframe. Residents in SDOH affected ZIP code areas with high value for immediate physician medical intervention are connected to telehealth virtual assistance and/or mobile blood draw labs services in home settings.

Results:

Model developed in this project demonstrated excellent performance on the test data set (N = 208,465)[. . . ]

Results:

SDOH-CKDPred predicted progression to Stage-5 CKD with sensitivity of 87% (AUROC = 0.87, 95% CI 0.85-0.89) for 24 months into the future. Performance of SDOH-CKDPred exceeded the clinical feature-only baseline by 8.3% (95% CI in sensitivity 1.12-10.21%). Our three-site rural pilot study showed 32% increase in_detection of early stage CKD, 28% increase in_patient referrals to nephrologists, and 31.9% reduction in_rate of high-risk CKD patients progressing to Stage-5.

Conclusions:

SDOH-CKDPred satisfied all criteria for health equity with its transparent model-agnostic interpretability technique for patient-specific explanations that helped physicians confirm clinical reasoning. The BCR was 3.75:1, signifying economic impact to make the solution sustainable. Our PILA-based telehealth architecture demonstrates that real-time solutions like SDOH-CKDPred will reduce, and not exacerbate, health disparities in chronic diseases.