Evaluating bias in target trial emulation for heart failure across statistical and deep learning methods

Abstract

Abstract Target trial emulation (TTE) is increasingly used for causal inference from observational data, but remains vulnerable to confounding by indication, and whether advanced adjustment methods mitigate this bias is unclear. Using Clinical Practice Research Datalink Aurum, we emulate target trials of beta-blockers (positive control) and digoxin (negative control) versus usual care on two-year all-cause mortality in patients with heart failure with reduced ejection fraction. We apply four adjustment strategies: propensity score matching, inverse probability of treatment weighting, targeted maximum likelihood estimation, and a Transformer-based deep learning approach. No method reproduces the randomised controlled trial (RCT) benchmarks: all suggest neutral or harmful effects for beta-blockers and elevated mortality for digoxin. In semi-synthetic simulation, all methods recover the true effects when confounders are observed, yet fail in real-world data. TTE, even with advanced adjustment, may not yield trial-equivalent estimates when confounding is strong; randomised evidence remains essential for clinical and policy decisions.