top of page

Anticoagulation Study

Real-world evidence (RWE) pharmacoepidemiologic comparative safety study comparing Direct Oral Anticoagulants (DOACs, e.g., Apixaban/Rivaroxaban) versus Warfarin in patients with non-valvular Atrial Fibrillation to evaluate incidence of major bleeding events (gastrointestinal and intracranial hemorrhage) and support formulary and clinical pathway decisions.

Business question: Does initiation of DOACs confer a significantly lower risk of major bleeding complications compared to Warfarin in non-valvular Atrial Fibrillation patients, and does baseline balance hold after large-scale propensity score adjustment?

Technology stack: R (Posit Cloud), OHDSI HADES ecosystem (CohortMethod, FeatureExtraction, DatabaseConnector, Cyclops), DuckDB / OMOP Common Data Model (CDM v5.3/v5.4).

covariate_balance.png
preference_score_distribution.png
kaplan_meier.png
Key Insights

​

  • Strong Clinical Equipoise: The preference score distribution shows significant overlap between cohorts without substantial clustering near 0 or 1, confirming comparable baseline indications and validating active-comparator selection.

  • Effective Confounding Control: Large-scale propensity score matching balanced thousands of extracted baseline covariates. Pre-match standardized mean differences (SMDs) reaching up to ~ 0.35 were consistently brought within acceptable clinical thresholds (SMD < 0.1) post-match.

  • Safety Profile Differentiation: The Kaplan-Meier survival curves display early and sustained divergence over the follow-up period. Patients in the target arm demonstrated higher event-free survival compared to the active comparator group.

  • Survival Disparity: The comparator group experienced a steeper and earlier decline in event-free probability, reaching approximately 94.5% survival by study end, compared to over 96.0% in the target treatment arm.

1. Data Source

​

  • Standardized Real-World Data Model: Data is mapped to the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM) v5.3 architecture, enabling reproducible pharmacoepidemiologic pipelines.

  • Longitudinal Patient Trajectories: Sourced from an established OMOP CDM synthetic benchmark population (Eunomia testbed), containing detailed longitudinal clinical histories.

  • Core Clinical Domains Analyzed: Includes Person (demographics), Drug Exposure (active drug eras, washouts, and duration of use), Condition Occurrence (baseline comorbidities and adverse outcome events), and Cohort (target, comparator, and outcome phenotype definitions).

  • Automated Feature Extraction: Utilized OHDSI HADES feature extraction protocols to assemble thousands of patient-level covariates covering medical conditions, prior medication exposures, and demographic risk factors.

2. Data Exploration
 

  • Data was extracted and processed directly from an OMOP CDM v5.3 database in R using the OHDSI HADES analytics suite (DatabaseConnector, FeatureExtraction, and CohortMethod).

  • Standardized OMOP concept sets were queried across drug_era, condition_occurrence, and cohort tables to define an active-comparator, new-user study population of 1,797 target patients and 830 comparator patients.

  • Through FeatureExtraction, 34,774 longitudinal baseline covariates (patient demographics, general conditions, and prior medication exposures) were constructed, with study drug exposures excluded to prevent collinearity.

  • Exploratory cohort workflows enforced a 180-day pre-index washout period to isolate incident users and established an on-treatment time-at-risk window (day 1 post-initiation to day 30 post-discontinuation) for downstream survival analysis.

  • A high-dimensional propensity score model was fitted via regularized logistic regression (L1 Laplace prior using Cyclops), followed by 1:1 nearest-neighbor caliper matching and stratified Cox proportional hazards regression.

3. Data Visualization
 

  • Diagnostic and outcome visualizations were generated directly through OHDSI CohortMethod plotting engines and exported as high-resolution figures for reporting.

  • Visualizations generated for the study include:

  • Attrition Diagram: The complete OHDSI flowchart illustrating each filter step (original cohorts, 180-day washout, index restriction, and at-risk requirements) with exact attrition drop numbers.

  • Patient Population: A two-panel vertical bar chart displaying 1,797 vs 830 patients before matching, side-by-side with the balanced 829 vs 829 matched cohorts.

  • Preference Score Density Plot: Displays patient distributions across the propensity score continuum to verify clinical equipoise and positive overlap between treatment arms.

  • Covariate Balance Scatter Plot: Compares pre- versus post-matching Standardized Mean Differences (SMD) against the diagonal identity line, demonstrating the reduction of baseline imbalances below the 0.1 threshold.

  • Kaplan-Meier Survival Curves: Illustrates cumulative event-free survival probabilities over time with shaded 95% confidence intervals and an integrated number-at-risk table (829 matched patients per arm).

Other projects

bottom of page