네트워크 분석을 활용한 교란변수 탐색 고도화-심방세동 환자에서 약물 효과 비교 사례-

Identification of Confounders Using Network Analysis:A Case Study in Patients with Atrial Fibrillation

This study aimed to improve confounding control in an observational study using health insurance claims data by applying diagnosis-based network analysis for covariate selection. We investigated whether proton pump inhibitor (PPI) co-therapy reduces the risk of upper gastrointestinal bleeding (UGIB) in patients with atrial fibrillation (AF) treated with direct oral anticoagulants (DOACs). A retrospective cohort study was conducted using Korean National Health Insurance claims data from July 2015 to June 2020. Adult patients newly prescribed DOACs for non-valvular AF were identified and classified according to PPI use. Two propensity score matching (PSM) models were constructed: Model 1 included conventional covariates from previous literature; Model 2 incorporated additional variables identified through diagnosis-based network analysis, which evaluated the strength and structure of co-occurring disease patterns. Among 94,482 eligible patients, 21.3% received PPI co-therapy. Network analysis identified 25 highly connected diagnoses, including gastroesophageal reflux disease (GERD), not previously used as covariates. When using Model 1, GERD showed substantial imbalance between groups (standardized mean difference [SMD]=0.76), which was resolved in Model 2 (SMD=0.01). Matching using Model 2 achieved improved covariate balance, suggesting better confounding adjustment. These findings demonstrate that network-based covariate selection can enhance the validity of real-world data analyses and support more reliable comparative effectiveness research.