Expert-Guided g-computation with Large Language Models for Estimating Causal Effects on Timings: Applications to Hospital Quality Improvement

Hospital quality improvement (QI) programs frequently need to prioritize candidate interventions to reduce patient length of stay (LOS). However, traditional quantitative methods struggle to evaluate hypothetical interventions that lack historical data, while qualitative expert reviews are often prone to cognitive biases. This paper addresses the challenge of estimating the causal effect of these interventions on patient timing metrics.

The authors propose "egg-computation," a hybrid framework that bridges qualitative Gantt charts (used by clinicians to map patient journeys) with formal causal DAGs. The approach identifies causal effects by:

1. Expert-Guided DAGs: Representing patient trajectories as DAGs where nodes are event times and edges are dependencies.
2. Hybrid Inference: Using data-driven models for events where historical data is reliable, while eliciting expert judgment for components that are unidentifiable from data.
3. LLM Integration: Using an LLM-assisted pipeline to extract these causal structures and reasoning from unstructured clinical notes, allowing the method to scale without manual expert intervention for every patient.

In simulation studies, egg-computation outperformed conventional causal inference methods, particularly when patient causal structures and intervention mechanisms were diverse. In a real-world application involving eleven candidate QI interventions at an urban safety-net hospital, the LLM-generated graphs and time-saving estimates showed high concordance with human experts. The study demonstrates that LLMs can effectively serve as a proxy for human experts in causal reasoning tasks, providing actionable insights into how interventions influence hospital flow.

This work provides a scalable, rigorous way to evaluate hospital policy changes before they are implemented. By formalizing the link between clinical process maps (Gantt charts) and causal inference, it allows hospitals to move beyond intuition-based decision-making toward evidence-based optimization of patient care pathways.