Assessing the Impact of Computable Type 2 Diabetes Phenotypes on Predicting Healthcare Utilization Using Electronic Health Records and Administrative Claims

Abstract

Background/Objectives: Type 2 Diabetes (T2D) computable phenotypes, which leverage electronic health records (EHRs) and administrative claims data, provide the basis for T2D population health research. Our study investigates how T2D phenotypes affect downstream healthcare utilization prediction, specifically inpatient (IP) and emergency room (ER) admissions. Methods: This study included 15,338 adult patients from a large academic medical center with both EHR and claims data from 2017 to 2019. We compared widely adopted and locally developed T2D phenotypes. EHR predictors and claims-based outcomes were used for utilization prediction. Models were developed using a 70/30 training-and-test split over 100 iterations. Mean area under the curve (AUC), odds ratios (ORs), positive predictive values, and negative predictive values were reported. Results: Models had comparable performance in concurrent predictions. Impact of phenotypic variation impact was more apparent in prospective predictions. The CMS Chronic Conditions Data Warehouse (CCW) phenotype was more discriminatory in predicting concurrent IP and ER admissions (AUCs of 0.80 and 0.74) than prospective IP and ER visits (0.70 and 0.73) in 2019. Conclusions: Our study demonstrated how phenotypic variations and data sources impact healthcare utilization prediction in T2D patients. Furthermore, we highlight the significance of phenotype selection for targeted T2D population health initiatives and management strategies.