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

Area of Science:

• Epidemiology
• Biostatistics
• Health Informatics

Background:

• Target trial emulation (TTE) is a method for causal inference from observational data.
• Confounding by indication remains a significant challenge in TTE.
• The effectiveness of advanced adjustment methods in mitigating this bias is not well understood.

Purpose of the Study:

• To evaluate the performance of TTE with various adjustment strategies in heart failure patients.
• To compare TTE estimates against randomized controlled trial (RCT) benchmarks.
• To assess the impact of confounding on TTE accuracy.

Main Methods:

• Emulation of target trials for beta-blockers and digoxin in heart failure with reduced ejection fraction using Clinical Practice Research Datalink Aurum.
• Application of four adjustment strategies: propensity score matching, inverse probability of treatment weighting, targeted maximum likelihood estimation, and a Transformer-based deep learning approach.
• Validation using semi-synthetic simulations.

Main Results:

• No TTE method reproduced RCT benchmarks, suggesting neutral/harmful effects for beta-blockers and increased mortality for digoxin.
• Advanced adjustment methods failed to fully mitigate confounding in real-world data.
• In simulations, methods recovered true effects with observed confounders but failed with unobserved ones.

Conclusions:

• TTE, even with advanced adjustments, may not provide trial-equivalent estimates in the presence of strong confounding.
• Observational data and TTE are insufficient for definitive clinical and policy decisions.
• Randomized evidence remains the gold standard for establishing treatment efficacy and safety.