Abstract: FR-OR019

Electronic Health Records-Based Computable Phenotype for CKD Diagnosis and Staging

Session Information

Category: Chronic Kidney Disease (Non-Dialysis)

  • 302 CKD: Estimating Equations, Incidence, Prevalence, Special Populations

Authors

  • Shang, Ning, Columbia University, New York, New York, United States
  • Kiryluk, Krzysztof, Columbia University, New York, New York, United States
  • Drawz, Paul E., University of Minnesota, Minneapolis, Minnesota, United States
  • Carroll, Robert J, Vanderbilt University Medical Center, Nashville, Tennessee, United States
  • Arruda-Olson, Adelaide M., Mayo Clinic, Rochester, Minnesota, United States
  • Mohan, Sumit, Columbia University, New York, New York, United States
  • Ionita-Laza, Iuliana, Columbia University, New York, New York, United States
  • Gharavi, Ali G., Columbia University, New York, New York, United States
  • Weng, Chunhua, Columbia University, New York, New York, United States
  • Hripcsak, George, Columbia University, New York, New York, United States
Background

Chronic Kidney Disease (CKD) diagnosis can be made by blood, urine, or imaging tests. This study is to design and evaluate a portable electronic phenotype for automated detection and staging of CKD based on electronic health records (EHR).

Methods

With urine tests data of 68,617 patients from three major health care systems (Columbia, Minnesota, and Vanderbilt), we used machine learning to design a universal albuminuria (A-stage) classifier that accommodates five common methods for quantification of urinary protein excretion, including two urine dipstick scales in combination with urine specific gravity. The performance of the classifier for each type of urine test was assessed by a 10-fold cross-validation procedure against matched UACR data. By integrating our A-stage classification with kidney-related diagnostic codes and serum Cr-based eGFR, a rule-based method, each individual is automatically assigned as CKD case or control. Each CKD case is additionally placed on a “staging grid” of albuminuria (A-stage) by eGFR (G-stage).

Results

The CKD algorithm has been designed, implemented and tested at Columbia University, achieving 100% and 83% PPVs for cases and controls respectively by manual chart review. We then performed external validation at Vanderbilt University and Mayo Clinic; where our electronic phenotype had PPVs of 85% and 98% for cases and controls, respectively. Based on expert review of 265 charts across all three sites, the PPVs to diagnose CKD Stage 1, 2, 3a, 3b, 4, and 5 were 75%, 98%, 95%, 93%, 89%, and 79%, respectively.

Conclusion

Using machine learning and rule-based methods, we developed and validated a portable CKD diagnosis and staging algorithm in a large multi-center effort. The electronic phenotype follows the latest guidelines and can be applied to both pediatric and adult patients. This algorithm can be implemented within any EHR to enable real-time automated detection and staging of CKD, enabling the implementation of stage-specific clinical decision support tools.

Funding

  • Other NIH Support