Skip to main content

Model Development Summary

Senior Asian Couple Adults with Medical Staff at Home
Top

The Anticoagulant-Related Bleeding Risk Prediction Model was developed using 2016–2018 Medicare fee-for-service claims data. To be included in the study cohort, a Medicare beneficiary had to be over 18 years of age, have at least 13 months of continuous enrollment in Parts A/B/D and no Part C, and have at least one Part D claim for an anticoagulant of interest to the study. Each beneficiary could contribute multiple observation windows based on anticoagulant exposures; each window comprised a 180-day look-back period and a 90-day outcome measurement period. New observation windows were created for the beneficiary by sliding the index date (i.e., the start of bleeding outcome measurement) 30 days forward repeatedly until reaching the end of the data time period. Windows were truncated at the date of death, Medicare disenrollment, or receiving hospice care, and no additional windows were created past these events. Once the beneficiary stopped the use of anticoagulants, defined as the end date of last anticoagulant prescription plus 10 days, the beneficiary did not contribute new windows unless the person restarted therapy later during the data time frame.

Bleeding outcomes were identified from ED, observation stay, and inpatient claims using the present-on-admission indicator and all 25 diagnosis positions following a previously validated diagnosis code set (Shehab et al., 2019. https://doi.org/10.1002/pds.4783). Risk variables were selected based on a literature review of bleeding risk factors, prior risk prediction schemes, and the synthesis of expert opinions from three clinical experts. Variables were specified in ICD-10-CM diagnosis codes, a limited set of ICD-10-PCS procedure codes, National Drug Codes, and demographic and healthcare utilization information available in Medicare claims.

A multiple logistic regression model was constructed, using the 30-day sliding window approach, from 53 million observation windows (N = 3.3 million beneficiaries, divided 60:20:20 into training, validation, and testing cohorts). The model was fitted on the testing dataset: it was well calibrated and showed moderate discrimination with a c-statistic of 0.678. When further narrowing to a smaller set of risk variables including only demographic and clinical risk factors (Groups 1, 2, and 3 of the risk variables), the model continued to show good calibration and moderate discrimination with a c-statistic of 0.665.