Abstract: FR-PO0003
NHEM: A Dual-Model Machine Learning Pipeline for Early CKD Progression Detection in Medicaid Populations with Type 2 Diabetes
Session Information
- Artificial Intelligence: Clinical Prediction, Decision Support, and Digital Kidney Care
October 23, 2026 | Location: Exhibit Hall A, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Author
- Onyeachom, Uchenna, Health For Mankind, Boston, Massachusetts, United States
Background
CKD progression in type 2 diabetes (T2DM) is frequently detected at advanced stages in Medicaid populations, where care fragmentation and incomplete laboratory data limit eGFR threshold–based monitoring. Validated models such as the KFRE require UACR, which is tested in only 21–53% of at-risk patients nationally. Earlier identification of high-risk trajectories in data-sparse populations is a critical unmet need.
Methods
NHEM™ (Novel Holistic Early-warning Model) is a patent-pending dual-model ML pipeline integrating two orthogonal domains: biologic kidney trajectory risk (XGBoost; 60% weight), encoding eGFR slope, acceleration, and baseline deviation via tsfresh temporal features; and care-continuity risk (LightGBM; 40% weight), encoding lab completion, visit frequency, and medication adherence. Missing data is treated as a directional positive risk signal rather than imputed or excluded. SHAP explainability is computed per output. The model delivers scores (0–100) via CDS Hooks in Epic FHIR R4. CKD progression was defined as: sustained ≥40% eGFR decline, eGFR slope worse than −5 mL/min/1.73 m/year, CKD stage 4–5 transition, kidney failure, dialysis, transplant waitlisting, or nephrology referral for progressive CKD.
Results
NHEM identified progression risk 6–12 months before standard threshold detection in retrospective modeling. Four risk tiers (Critical ≥80, High ≥60, Moderate 35–59, Low 35) map to tiered care pathways: SGLT2i review, RAAS optimization, UACR testing prompts, finerenone assessment, CHW deployment, nephrology eConsult, and direct referral. The absence of albuminuria is a recognized limitation; a hybrid UACR integration roadmap leveraging albuminuria when available while maintaining universal computability is under development.
Conclusion
NHEM is an EHR-integrated AI platform detecting high-risk DKD trajectories before conventional identification, with a missing-data-as-signal architecture addressing structural limitations of existing models in underserved populations. A prospective stepped-wedge cluster randomized trial at UMass Chan Medical School will evaluate chronic eGFR slope, downstream impact metrics (UACR testing, SGLT2i/RAAS optimization, referral timeliness), and equity across demographic and socioeconomic subgroups.
Funding
- Commercial Support – Health For Mankind