(45) Development and Validation of a Chronic Kidney Disease Phenotype Using Combined Electronic Health Record and Claims Data

Publication Number: 376

Background: Chronic kidney disease (CKD) is a major public health burden. There is growing interest in using real-world evidence to inform clinical care in this patient population who are underrepresented or excluded from clinical trials. However, accurately predicting and defining a CKD phenotype in real world data remains understudied.

Objectives: To build prediction models for identifying a CKD phenotype using electronic health record (EHR) and claims data.

Methods: We used the Research Partners Data Repository at the Mass General Brigham hospital system linked with Medicare claims data (EHR-Medicare: 2007-2017). Patients with at least 90 days of continuous Medicare enrollment were included if they had a valid serum creatinine laboratory result and excluded if they had end-stage kidney disease or were on dialysis. The primary outcomes were continuous estimated glomerular filtration rate (eGFR), and eGFR in categories ( < 60, < 45 and < 30 ml/min/1.73m2). Baseline characteristics were assessed during the 90 days before and after cohort entry. Patients were partitioned into training and validation sets. Potential variables included demographic characteristics, comorbidities, medication use, and healthcare utilization. Variables were selected using least absolute shrinkage and selection operator (LASSO) regression, after which linear regression models were developed to predict continuous eGFR, and logistic regression models were used to predict eGFR in categories. Model performance characteristics were assessed by calculating the area under the receiver operating curves (AUC) for the predicted probability of eGFR in the training and validation sets.

Results: Our study included 174,220 patients in the EHR-Medicare cohort. The training set included 117,476 patients (mean age 74.8 years, female 58.2%, mean eGFR 70.9 ml/min/1.73m2) and the validation set included 56,744 patients (mean age 73.8 years, female 59.6%, mean eGFR 72.5 ml/min/1.73m2). We included 124 variables to predict eGFR. The model R2 for predicting continuous eGFR was 0.35. The AUC for predicting eGFR < 60 ml/min/1.73m2 was 0.81 (training) and 0.82 (validation), eGFR < 45 ml/min/1.73m2 was 0.87 (training) and 0.91 (validation), and eGFR < 30 ml/min/1.73m2 was 0.93 (training) and 0.97 (validation).

Conclusions: EHR and claims data together can be used to predict with sufficient accuracy a CKD phenotype. Our findings suggest that databases with EHR-claims data could be used for future real-world research studies in patients with CKD stages 3-5.