Jennifer Lyons1, Vibha Desai2, Yaping Xu3, Greg Ridgeway4, William Finkle5, Paul Solari3, Sean Sullivan6, Stephan Lanes2. 1. HealthCore, Wilmington, DE, USA. Electronic address: jglyons@bu.edu. 2. HealthCore, Wilmington, DE, USA. 3. Genentech, South San Francisco, CA, USA. 4. Consolidated Research, Inc., Los Angeles, CA, USA; University of Pennsylvania, Philadelphia, PA, USA. 5. Consolidated Research, Inc., Los Angeles, CA, USA. 6. University of Washington, Seattle, WA, USA.
Abstract
BACKGROUND: The accuracy with which hemophilia A can be identified in claims databases is unknown. OBJECTIVE: Develop and validate an algorithm using predictive modeling supported by machine learning to identify patients with hemophilia A in an administrative claims database. METHODS: We first created a screening algorithm using medical and pharmacy claims to identify potential hemophilia A patients in the US HealthCore Integrated Research Database between January 1, 2006 and April 30, 2015. Medical records for a random sample of patients were reviewed to confirm case status. In this validation sample, we used lasso logistic regression with cross-validation to select covariates in claims data and develop a predictive model to estimate the probability of being a confirmed hemophilia A case. RESULTS: The screening algorithm identified 2,252 patients and we reviewed medical records for 400 of these patients. The screening algorithm had a positive predictive value (PPV) of 65%. The predictive model identified 18 predictors of being a hemophilia A case or noncase. The strongest predictors of case status included male sex, factor VIII therapy, office visits for hemophilia A, and hospitalizations for hemophilia A. The strongest predictors of noncase status included hospitalizations for reasons other than hemophilia A and factor VIIa therapy. A probability threshold of ≥0.6 resulted in a PPV of 94.7% (95% CI: 92.0-97.5) and sensitivity of 94.4% (95% CI: 91.5-97.2). CONCLUSIONS: We developed and validated an algorithm to identify hemophilia A cases in an administrative claims database with high sensitivity and high PPV.
BACKGROUND: The accuracy with which hemophilia A can be identified in claims databases is unknown. OBJECTIVE: Develop and validate an algorithm using predictive modeling supported by machine learning to identify patients with hemophilia A in an administrative claims database. METHODS: We first created a screening algorithm using medical and pharmacy claims to identify potential hemophilia Apatients in the US HealthCore Integrated Research Database between January 1, 2006 and April 30, 2015. Medical records for a random sample of patients were reviewed to confirm case status. In this validation sample, we used lasso logistic regression with cross-validation to select covariates in claims data and develop a predictive model to estimate the probability of being a confirmed hemophilia A case. RESULTS: The screening algorithm identified 2,252 patients and we reviewed medical records for 400 of these patients. The screening algorithm had a positive predictive value (PPV) of 65%. The predictive model identified 18 predictors of being a hemophilia A case or noncase. The strongest predictors of case status included male sex, factor VIII therapy, office visits for hemophilia A, and hospitalizations for hemophilia A. The strongest predictors of noncase status included hospitalizations for reasons other than hemophilia A and factor VIIa therapy. A probability threshold of ≥0.6 resulted in a PPV of 94.7% (95% CI: 92.0-97.5) and sensitivity of 94.4% (95% CI: 91.5-97.2). CONCLUSIONS: We developed and validated an algorithm to identify hemophilia A cases in an administrative claims database with high sensitivity and high PPV.