Evaluation of Machine Learning-Based Propensity Score Estimation: A Benchmarking Observational Analysis Against a Randomized Trial.

Machine learning (ML) approaches for propensity score estimation are increasingly used with the expectation of improving covariate balance and reducing bias, but their validity in selecting appropriate confounders remains controversial. In this study, we estimated the effectiveness of sacubitril/valsartan versus angiotensin-converting enzyme inhibitor and angiotensin receptor blocker on all-cause mortality among heart failure patients with implantable cardioverter defibrillators in the U.S. Department of Veterans Affairs from 2016 to 2020. We compared results from traditional logistic regression- and ML-based propensity score methods and benchmarked them against the PARADIGM-HF randomized trial. The estimate from logistic regression with confounder selection (HR = 0.93, 95% CI 0.61 - 1.42; 27-month RR = 0.87, 95% CI 0.59 - 1.21) most closely aligned with the trial result (HR = 0.81; 95% CI 0.61 - 1.06). In contrast, generalized boosting models did not outperform traditional logistic regression, and may amplify bias when combined with a data-driven confounder selection (HR = 0.63, 95% CI 0.31 - 1.30; RR = 0.61, 95% CI 0.33 - 1.04). Our findings suggest that ML-based propensity scores may introduce overadjustment bias and underscore the importance of subject-matter knowledge in causal inference with high-dimensional real-world data.