Machine Learning Versus Logistic Regression for Propensity Score Estimation: A Benchmark Trial Emulation Against the...

Area of Science:

• Cardiovascular Medicine
• Health Informatics
• Biostatistics

Background:

• Machine learning (ML) is increasingly used for propensity score estimation to improve covariate balance and reduce bias.
• The validity of ML in selecting appropriate confounders for causal inference remains controversial.
• Heart failure management often involves comparing drug effectiveness using real-world data.

Purpose of the Study:

• To estimate the effectiveness of sacubitril/valsartan versus traditional therapies on all-cause mortality in heart failure patients.
• To compare propensity score estimation using traditional logistic regression versus ML approaches.
• To benchmark real-world evidence (RWE) findings against the PARADIGM-HF randomized controlled trial.

Main Methods:

• Retrospective cohort study of U.S. Department of Veterans Affairs heart failure patients (2016-2020) with implantable cardioverter defibrillators.
• Propensity scores were estimated using logistic regression with *a priori* confounder selection and ML-based methods (generalized boosting models).
• Effectiveness was measured by all-cause mortality, hazard ratios (HR), and risk ratios (RR), compared to the PARADIGM-HF trial.

Main Results:

• Logistic regression with *a priori* confounder selection yielded results (HR=0.93, 95% CI 0.61-1.42) closely aligning with the PARADIGM-HF trial (HR=0.81, 95% CI 0.61-1.06).
• ML-based propensity scores, particularly with data-driven confounder selection, did not outperform logistic regression and potentially amplified bias (HR=0.63, 95% CI 0.31-1.30).
• ML methods may introduce overadjustment bias in complex, high-dimensional datasets.

Conclusions:

• Subject-matter expertise is crucial for confounder selection in causal inference with real-world data.
• Traditional logistic regression with careful confounder selection may be more reliable than ML approaches for estimating treatment effectiveness in this context.
• Over-reliance on data-driven methods without clinical validation can lead to biased estimates in cardiovascular research.