(156) High-dimensional propensity score implementation utilising electronic health records in Hong Kong

Background: Residual confounding remains a challenge for causal inference in observational studies using electronic health records. Traditionally, researchers mitigate confounding bias through pre-specified selection of covariates based on clinical and epidemiological knowledge. The high-dimensional propensity score (HDPS) empirically selects data-driven covariates from databases to further control for confounding but it has not been applied in routine clinical data in hospital setting.

Objectives: The study aims to demonstrate the implementation of HDPS methods in routine clinical data in hospital setting from Hong Kong using the association between antihypertensives and dementia as a case-study, and discuss HDPS as an approach to identify database-specific covariates that are usually neglected for confounder adjustment.

Methods: HDPS was applied to a cohort study comparing initiation of different classes of antihypertensives and the risk of dementia using routine clinical data captured from hospitals in Hong Kong. Patients were followed up from antihypertensive initiation until the earliest of dementia diagnosis, death, or study end date. We identified three data dimensions – diagnosis, prescriptions, and laboratory tests. Recurrence assessment was conducted with a one-year window prior to study entry. HDPS covariates were prioritized using the Bross formula and the top 250 covariates were used in the main analysis. Inverse probability of treatment weighting was used with Cox regression to estimate hazard ratios (HRs) comparing different antihypertensive drug classes with angiotensin-converting enzyme inhibitors (ACEi). Diagnostics were reviewed to evaluate the HDPS covariates included.

Results: We included 434,506 antihypertensive users and observed no difference in the hazard of dementia by comparing different antihypertensive drug classes with ACEi using researcher-specified propensity score. After implementing HDPS, we observed a lower risk of dementia associated with beta-blockers compared with ACEi (HR: 0.89, 95% Confidence Interval: 0.82-0.97). Reviewing HDPS covariates, we identified that protein-calorie malnutrition and fracture of base of skull were potential strong confounders not included in the pre-specified propensity score.

Conclusions: Inclusion of HDPS covariates led to a lower hazard of dementia amongst beta-blocker users compared with ACEi users; this was not detected using pre-specified covariates only. The potentially influential covariates identified were hard-to-measure markers of frailty, highlighting the potential for HDPS to identify database-specific covariates that may be difficult to discern based on prior knowledge alone. An R package was also developed for implementing HDPS in any database.