Step-by-step causal analysis of EHRs to ground decision-making

Area of Science:

• Medical Informatics
• Causal Inference
• Machine Learning

Background:

• Electronic health records (EHRs) offer vast observational data for treatment effect estimation.
• Randomized controlled trials (RCTs) face limitations in scope and applicability.
• Causal inference methods are crucial for reliable analysis of EHR data.

Purpose of the Study:

• To propose and illustrate a framework for causal inference using EHR data.
• To estimate the effect of albumin on mortality in sepsis patients.
• To compare sensitivity analyses with RCT results as a gold standard.

Main Methods:

• Utilized the target trial and PICOT framework for study design.
• Selected confounding variables based on expert knowledge.
• Employed a doubly robust estimator (AIPW) with random forests for analysis.

Main Results:

• Identified and mitigated immortal time bias related to treatment initiation.
• Demonstrated that adding confounders improves recovery of RCT results.
• The AIPW estimator with random forests proved most reliable.

Conclusions:

• A structured, step-by-step framework is essential for valid causal inference from EHRs.
• Albumin treatment showed improved efficacy in specific subgroups (older males, septic shock).
• Causal inference enhances machine learning for personalized clinical decision-making.