Estimating the Risk of Myocardial Infarction in Persons with Hemophilia A Using a Machine-Learning Approach with US Claims Data

Introduction

Epidemiological studies in the literature provide conflicting conclusions regarding the risk of myocardial infarction (MI) in people diagnosed with congenital hemophilia A (HA). As a result, the question of protection from MI conferred by HA remains debated. Using insurance claims data, we conducted a study to explore this potential relationship.

Methods

Initially, a traditional pharmaco-epidemiological approach was conducted using the Truven MarketScan Commercial Database and/or Medicare Supplemental Database. A cohort of persons with HA (PwHA) was identified based on the following criteria: having had a confirmed diagnosis of congenital HA between Jan 1, 2000 and Sept 30, 2017; at least three claims for HA within 365 consecutive days; and continuous enrollment with insurance coverage for the 6 months after first diagnosis of HA. In addition, all individuals were required to be aged ≥18 years, male, and with no evidence of a diagnosis of von Willebrand disease (VWD), hemophilia B, acquired HA, or MI prior to their first HA diagnosis. Based on the results of the first analysis, a second, novel approach was undertaken, using a 2-step method incorporating machine learning and drug utilization for cohort identification. For this approach, the inclusion criteria for the study were further refined to include persons with at least one medical or pharmacy claim for factor VIII (FVIII) therapy, activated prothrombin complex concentrate, or activated factor VIIa therapy; or at least one medical or pharmacy claim for FVIII/VWD therapy; or at least one medical or pharmacy claim for desmopressin and at least one medically-attended visit with a diagnosis of HA in the same claim line; or at least one medically-attended visit with a HA diagnosis. The earliest date for fulfilling any of these inclusion criteria was deemed the individual's index date. Participants also had to have 183 days of continuous insurance enrollment prior to study entry in order to participate. Secondly, a HA classification algorithm, first developed and validated by Lyons et al (Value in Health 2018), was adapted and applied to the above refined cohort. A cohort of individuals with no evidence of HA in the study period was then randomly selected from the MarketScan database and frequency matched to the HA cohort by age, sex, insurance type, region, enrollment length, diabetes status, and hypertensive status at a 1:3 ratio, yielding a control cohort of 18,817 individuals. A Poisson regression model was then fitted to estimate the adjusted incidence rate ratio (IRR). The model was adjusted for all baseline covariates as well as HIV and hepatitis C status, with age as a time-varying covariate.

Results

Based on the chosen criteria, an initial cohort of 2148 PwHA was identified. The crude incidence rate of MI in this cohort was estimated to be 1.75 (95% confidence interval [CI] 1.43-2.11) per 100 person-years. Relative to the matched cohort of individuals with no evidence of HA (N=10,661), this yielded an IRR of 1.68. Such a high relative risk was investigated further by examining concomitant medications. This revealed evidence of misclassification bias with 16% of participants having been prescribed anticoagulants and a low frequency of hemophilia drug utilization. These results prompted the revision of methods underlying cohort identification to the machine learning/drug utilization approach. The revised cohort yielded a 98.5% specificity and 77.8% sensitivity for selection of PwHA, identifying 3154 individuals with a ≥90% probability of being true PwHA. Ten were excluded as they had a previous history of MI, leaving a final cohort of 3144 PwHA. The crude incidence rate of MI in this HA cohort was calculated to be 0.25 (95% CI 0.15-0.34) per 100 person-years and 0.22 (95% CI 0.18-0.27) in the non-HA population, yielding an unadjusted IRR of 1.13. The adjusted IRR was estimated to be 1.31 (95% CI 0.85-2.00, p=0.22), indicating no statistically significant difference in the risk of MI in the HA population versus a matched non-HA control.

Conclusions

No evidence of a different risk of MI in PwHA relative to non-HA counterparts was observed in this analysis. To our knowledge, this represents the largest real-world data study investigating MI in the HA population. While every effort was taken to mitigate for the effects of confounders and bias, the results should be interpreted in the context of the limitations of a secondary data use study.