Primary Submission Category: Applications in Health and Biology
Ensemble Causal Structure Learning for Actionable Insights into Repeated Healthcare Utilization
Authors: Shishir Adhikari, Guido Muscioni, Mark Shapiro, Plamen Petrov, Elena Zheleva,
Presenting Author: Shishir Adhikari*
Understanding the factors that trigger or prevent repeated undesirable health outcomes, such as emergency department (ED) visits and hospital readmissions, is critical for improving quality of care and reducing costs. When randomized controlled trials are infeasible, causal structure learning (CSL) provides an alternative for generating causal hypotheses from observational data, but its reliability is limited by strong assumptions and model uncertainty. We hypothesize that an ensemble of CSL algorithms improves robustness by identifying causes that persist across different assumptions. We propose an end-to-end framework that integrates an ensemble of CSL algorithms with causal effect estimation to identify and rank causal risk and preventive factors, and to quantify heterogeneous effects across subpopulations. The framework outputs candidate causes and effect modifiers with confidence scores based on agreement across methods. Experiments on synthetic and semi-synthetic data show that a majority-voting ensemble improves recall of causal factors while maintaining precision. Application to real-world healthcare data yields clinically plausible hypotheses aligned with existing knowledge and identifies subpopulations with differential susceptibility. This approach enables data-driven identification of actionable interventions, supporting targeted strategies to improve patient outcomes while reducing avoidable healthcare utilization and associated costs.