Translational & Implementation Science
Causal Decision Modules in the EHR: Bedside Care in Acute Decompensated Heart Failure Ishan Suthar* Ishan Suthar Suthar University of South Carolina
Epidemiological analyses of large volumes of electronic health records (EHRs) have largely focused on retrospective estimation and prediction. Most EHR based Clinical Decision Support (CDS) still remain predictive rather than causal and rarely inform interventional knowledge at the moment of care, limiting their utility. To answer a clinician’s question, “what’s next?”, we propose a deployable, EHR-native causal inference framework that generates interpretable, action-oriented recommendations for inpatient acute decompensated heart failure (ADHF) decisions.
We propound a theory-driven causal inference framework that embeds target trial emulation directly within ADHF inpatient workflows to align causal estimation with real clinical decision points. ADHF care is framed as a sequence of target trials at key decisions (initial decongestion, escalation, therapy, discharge readiness), each with prespecified eligibility, treatment options, and decision-aligned estimands. Eligibility is reassessed at each decision point, enabling rolling enrollment and avoiding artificial baselines. Intervention strategies reflect bedside actions: IV loop diuretic intensity/continuation vs temporary holding of guideline-directed therapy, addressing treatment–confounder feedback.
Clinical notes are leveraged to measure time-varying severity as confounders while accounting for documentation bias and measurement error. Causal effects are identified using doubly robust g-methods with flexible nuisance models using ML, under assumptions of consistency, sequential exchangeability, and positivity. To emulate real-world timing, we use cloning–censoring–weighting to address immortal time and selection biases.
Each target trial is packaged as a reusable EHR-based causal decision module and evaluated prospectively in silent mode before clinician-facing release to assess safety, equity, and decision concordance. This provides a feasible path to transparent, causal-based bedside decision support for ADHF.
