Heterogeneous Causal Discovery of Repeated Undesirable Health Outcomes

Shishir Adhikari, Guido Muscioni, Mark Shapiro, Plamen Petrov, Elena Zheleva

Abstract

Understanding factors triggering or preventing undesirable health outcomes across patient subpopulations is essential for designing targeted interventions. While randomized controlled trials and expert-led patient interviews are standard methods for identifying these factors, they can be time-consuming and infeasible. Causal discovery offers an alternative to conventional approaches by generating cause-and-effect hypotheses from observational data. However, it often relies on strong or untestable assumptions, which can limit its practical application. This work aims to make causal discovery more practical by considering multiple assumptions and identifying heterogeneous effects. We formulate the problem of discovering causes and effect modifiers of an outcome, where effect modifiers are contexts (e.g., age groups) with heterogeneous causal effects. Then, we present a novel, end-to-end framework that incorporates an ensemble of causal discovery algorithms and estimation of heterogeneous effects to discover causes and effect modifiers that trigger or inhibit the outcome. We demonstrate that the ensemble approach improves robustness by enhancing recall of causal factors while maintaining precision. Our study examines the causes of repeat emergency room visits for diabetic patients and hospital readmissions for ICU patients. Our framework generates causal hypotheses consistent with existing literature and can help practitioners identify potential interventions and patient subpopulations to focus on.

Introduction

In healthcare, identifying why a patient experiences repeated adverse events—such as frequent emergency room (ER) visits or hospital readmissions—is vital for developing effective interventions. While Randomized Controlled Trials (RCTs) remain the gold standard for establishing causality, they are often prohibitively expensive or ethically complex in observational settings. Causal Structure Learning (CSL) provides a data-driven alternative, yet it frequently suffers from sensitivity to model assumptions and a failure to account for treatment effect heterogeneity across different patient subpopulations.

Methodology

The proposed framework addresses these limitations through a two-stage approach: a robust ensemble for causal discovery and a detailed analysis of effect heterogeneity.

Causal Structure Learning Ensemble

To mitigate the risk of relying on a single algorithm's assumptions, the framework employs an ensemble of CSL algorithms, including Hill Climbing (HC), Fast Greedy Equivalence Search (FGES), and Max-Min Hill Climbing (MMHC). A "cause support" metric is introduced to quantify the consensus among these algorithms, prioritizing causal relationships identified by multiple models.

Temporal Constraints and Data Representation

The framework leverages the inherent temporal order of medical events. Since causes must precede effects, time-stamped data allows for the orientation of edges in the causal graph. The process involves mapping dense event sequences into sparse representations suitable for causal analysis.

Mapping temporal event sequences to sparse representations for causal discovery.

Heterogeneity Analysis

Beyond identifying average causal effects, the framework seeks effect modifiers—variables such as age, comorbidities, or specific medical histories that alter the impact of a cause on the outcome. By utilizing backdoor adjustment and linear models with interaction terms, the system estimates Conditional Average Treatment Effects (CATE). This allows for the identification of specific contexts where an intervention might be more or less effective.

Causal graph illustrating different types of effect modifiers (EMs) and their relationships with treatment, mediator, and outcome.

Validation with Large Language Models

To bridge the gap between data-driven hypotheses and clinical knowledge, the authors utilized Large Language Models (LLMs) to cross-reference discovered causal relationships with existing medical literature. This provides a scalable validation mechanism in the absence of extensive expert review.

Experimental Evaluation

The framework was validated using both real-world and synthetic datasets:

  • Diabetic ER Visits: Analysis of 461,754 claims from approximately 12,000 diabetic patients.
  • MIMIC-IV: Investigation of 30-day hospital readmissions using over 335,000 clinical records.
  • Synthetic Data: Evaluation against known causal structures (e.g., Erdos-Renyi graphs) to benchmark precision and recall.

Key Results

The ensemble approach significantly improved recall on synthetic data while maintaining high precision, demonstrating that algorithmic consensus is a reliable indicator of true causality. In real-world applications, the framework identified established risk factors, such as kidney failure and medication non-compliance.

Furthermore, the analysis revealed actionable insights regarding preventive care. For instance, "General Medical Exams" (Z000) and "Encounters for Immunization" (Z23) were found to have significant negative causal effects on repeat ER visits. Crucially, the effectiveness of these interventions varied based on patient medical history, such as a history of convulsions, illustrating the power of identifying effect modifiers for targeted care management.

Conclusion

This work demonstrates that combining ensemble-based causal discovery with heterogeneity analysis provides a robust tool for healthcare research. By identifying not just the causes of adverse outcomes but also the specific patient contexts that modify these effects, the framework offers a path toward more personalized and effective clinical interventions. Future research will focus on relaxing the assumption of causal sufficiency to better account for unmeasured confounding variables.

Summary updated

Abstract

Understanding factors triggering or preventing undesirable health outcomes across patient subpopulations is essential for designing targeted interventions. While randomized controlled trials and expert-led patient interviews are standard methods for identifying these factors, they can be time-consuming and infeasible. Causal discovery offers an alternative to conventional approaches by generating cause-and-effect hypotheses from observational data. However, it often relies on strong or untestable assumptions, which can limit its practical application. This work aims to make causal discovery more practical by considering multiple assumptions and identifying heterogeneous effects. We formulate the problem of discovering causes and effect modifiers of an outcome, where effect modifiers are contexts (e.g., age groups) with heterogeneous causal effects. Then, we present a novel, end-to-end framework that incorporates an ensemble of causal discovery algorithms and estimation of heterogeneous effects to discover causes and effect modifiers that trigger or inhibit the outcome. We demonstrate that the ensemble approach improves robustness by enhancing recall of causal factors while maintaining precision. Our study examines the causes of repeat emergency room visits for diabetic patients and hospital readmissions for ICU patients. Our framework generates causal hypotheses consistent with existing literature and can help practitioners identify potential interventions and patient subpopulations to focus on.
BibTeX
@misc{adhikari2025heterogeneous,
  author = {Shishir Adhikari and Guido Muscioni and Mark Shapiro and Plamen Petrov and Elena Zheleva},
  title = {{Heterogeneous Causal Discovery of Repeated Undesirable Health Outcomes}},
  year = {2025},
  eprint = {2503.11477},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.11477},
}
Figure