OBJECTIVES : Major bleeding is an infrequent side effect during anticoagulants therapy. ABC (age, biomarkers and clinical history) was reported to be the best model among existing bleeding risk algorithms. We aim to identify a better performing bleeding risk prediction model to enrich patients for clinical studies.

METHODS : Optum Pan-Therapeutic Electronic Health Records from 2006 to 2017 were used in the analysis. Patients with major bleeding events associated with hospitalization while on Factor Xa treatment regimen were defined as case group. Similarly, controls without major bleeding during Factor Xa treatment regimen were drawn from the same period. The last Factor Xa treatment regimen start date was used as index date. Predictive modeling was then performed to predict subsequent major bleeding event using predictors including demographic variables, indicators of comorbidity, medications, indicators for lab test requisitions, healthcare service utilization extracted at 3 months, 12 months and lifetime (3 years) before the index date. An ensembled prediction of extreme gradient boosting (XGB) models was applied.

RESULTS : 9,376 subjects in case group and 175,213 subjects in control group met the inclusion criteria. Univariate analysis for lab tests +/- 30 days prior to index date were screened to identify 48 (out of 99) lab measurements that differentiated cases from controls (p < 0.05/99). 31 (out of 48) were replicated in the replication samples (p < 0.05/48). Joint analysis identified 53 analytes (p < 0.05/99). With only 10 predictors, an area under the receiver operating characteristic curve (AUC ROC) of 84%, balanced accuracy of 76%, sensitivity of 75%, and specificity of 76% have been achieved in independent test samples.

CONCLUSIONS : Risk models can be derived based on EHR records to identify subjects at-risk to major bleeding for patient enrichment in clinical studies and or real-world prescription practice.