A Novel Chronic Kidney Disease Phenotyping Algorithm Using Combined Electronic Health Record and Claims Data

Area of Science:

• Nephrology
• Health Informatics
• Biostatistics

Background:

• Chronic kidney disease (CKD) is frequently under-coded in healthcare claims data.
• Accurate CKD phenotyping is crucial for research and clinical applications.
• Electronic health records (EHRs) contain valuable laboratory data for CKD assessment.

Purpose of the Study:

• To develop and validate prediction models using claims data to identify CKD phenotypes.
• To overcome limitations of under-coding in claims data for CKD diagnosis.
• To enable robust CKD phenotyping without direct access to laboratory results.

Main Methods:

• Linked EHR data with Medicare claims for training and validation cohorts.
• Included individuals aged ≥65 with serum creatinine in EHRs (2007-2017).
• Utilized LASSO regression to predict estimated glomerular filtration rate (eGFR) categories (<60, <45, <30 mL/min/1.73m²).

Main Results:

• The validation cohort included 56,744 patients.
• The model achieved high Area Under the Curve (AUC) values: 0.81 for eGFR <60, 0.88 for eGFR <45, and 0.92 for eGFR <30 mL/min/1.73m².
• Positive predictive values ranged from 0.38 to 0.80 for the respective eGFR categories.

Conclusions:

• A claims-based prediction model for CKD stages defined by eGFR was successfully developed.
• This model serves as a valuable proxy for clinical endpoints and confounders in research.
• Facilitates enhanced subgroup effect assessment in studies lacking direct laboratory data.