OBJECTIVES: Machine learning techniques have been recommended as promising alternatives to logistic regression for the estimation of propensity scores. This study compared the performance of the machine learning approach Inverse Probability Weighting- Gradient Boosted Modeling (IPTW-GBM) with traditional Propensity Scores (PS) method in predicting risk of serious events across individual Cholinesterase inhibitors (ChEIs) use in older adults with dementia.

METHODS: This was a retrospective cohort study of older adults (aged≥65 years) with dementia diagnosis (ICD-9/10-CM codes) using 2013-2015 Medicare claims data involving Parts A, B, and D. The study identified new users of ChEIs (donepezil, galantamine, or rivastigmine) with six months of washout period. Patients were followed for up to 180 days for the occurrence of serious events (emergency department visits, inpatient hospitalizations, or death). The PS and IPTW-GBM were calculated using 82 covariates. Covariate balance was assessed using standardized mean differences (SMD) < 0.1, which was calculated before and after implementing PS methods. Cox proportional hazards regression models were implemented using PS as covariates and the IPTW-GBM to assess the risk of serious events across individual ChEIs.

RESULTS: The study identified 524,975 older adults with dementia who were incident users of ChEIs. The SMD of < 0.1 were observed for 100% of covariates in both methods. Cox model with both methods found that hazard ratios were similar for the risk of serious events (PS as covariate: rivastigmine vs. donepezil HR 1.12, 95% CI 1.03-1.22, galantamine vs. donepezil HR 1.51, 95% CI 1.24-1.84; IPTW- GBM: rivastigmine vs. donepezil HR 1.14, 95% CI 1.09-1.18, galantamine vs. donepezil HR1.52, 95% CI 1.26-1.75).

CONCLUSIONS: The GBM performed similarly to traditional PS models in terms of covariate balance and the risk of the outcome. Multiple methodological approaches can help to evaluate the robustness of the study findings in pharmacoepidemiology research.