OBJECT-CENTRIC CAUSAL DISCOVERY IN MIMIC-IV: A PARADIGM SHIFT FOR IDENTIFYING RISK FACTORS FOR HOSPITAL-ACQUIRED PRESSURE INJURIES IN REAL-WORLD CLINICAL DATA
Author(s)
Jenny Alderden, PhD1, Michael Haft, PhD2, Andy Wilson, PhD3;
1Boise State University, Boise, ID, USA, 2Xplain Data, Munich, Germany, 3University of Utah, Salt Lake City, UT, USA
1Boise State University, Boise, ID, USA, 2Xplain Data, Munich, Germany, 3University of Utah, Salt Lake City, UT, USA
OBJECTIVES: Traditional causal discovery algorithms face critical limitations in healthcare: they assume "flat" datasets, struggle with EHR relational complexity, and yield equivalence classes rather than fully-oriented graphs. We present an object-centric approach that encodes clinical constraints (temporal precedence, relational structure, informative missingness) while enabling data-driven discovery within that structure. We applied this method to identify causal factors for hospital-acquired pressure injuries (HAPrI) in MIMIC-IV (v3.1).
METHODS: Using MIMIC-IV adult ICU admissions, we represented patient journeys as temporally ordered "objects" (admissions, procedures, medications, labs) and applied a causal discovery algorithm operating directly on this relational structure without flattening. Domain knowledge constrained the search space, enforcing temporal precedence to minimize target leakage and encoding missingness mechanisms, while the algorithm screened millions of candidate dependencies to construct a sparse causal graph. The algorithm explored dependencies without pre-specified candidates, though we prioritized albumin's relationship to HAPrI for focused evaluation given its contested causal status.
RESULTS: The algorithm identified hypoalbuminemia as a significant upstream driver of HAPrI, supporting albumin's potential protective benefit, a relationship frequently obscured by confounding in conventional analyses. It also rediscovered established risk factors (immobility, moisture, hemodynamic instability, mechanical ventilation) without manual feature engineering, validating clinical plausibility. Object-centric search condensed millions of correlations into interpretable causal structure, surfacing under-recognized factors including fluid management patterns and albumin timing.
CONCLUSIONS: Object-centric causal discovery offers a middle path between purely data-driven search and expert-specified DAGs: clinical knowledge disciplines the algorithm, while the algorithm surfaces relationships experts might overlook. Future work will engage clinicians to refine the graph and employ modern causal estimators (TMLE, longitudinal g-methods) to quantify effects of albumin and other interventions on HAPrI risk.
METHODS: Using MIMIC-IV adult ICU admissions, we represented patient journeys as temporally ordered "objects" (admissions, procedures, medications, labs) and applied a causal discovery algorithm operating directly on this relational structure without flattening. Domain knowledge constrained the search space, enforcing temporal precedence to minimize target leakage and encoding missingness mechanisms, while the algorithm screened millions of candidate dependencies to construct a sparse causal graph. The algorithm explored dependencies without pre-specified candidates, though we prioritized albumin's relationship to HAPrI for focused evaluation given its contested causal status.
RESULTS: The algorithm identified hypoalbuminemia as a significant upstream driver of HAPrI, supporting albumin's potential protective benefit, a relationship frequently obscured by confounding in conventional analyses. It also rediscovered established risk factors (immobility, moisture, hemodynamic instability, mechanical ventilation) without manual feature engineering, validating clinical plausibility. Object-centric search condensed millions of correlations into interpretable causal structure, surfacing under-recognized factors including fluid management patterns and albumin timing.
CONCLUSIONS: Object-centric causal discovery offers a middle path between purely data-driven search and expert-specified DAGs: clinical knowledge disciplines the algorithm, while the algorithm surfaces relationships experts might overlook. Future work will engage clinicians to refine the graph and employ modern causal estimators (TMLE, longitudinal g-methods) to quantify effects of albumin and other interventions on HAPrI risk.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
MSR72
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
SDC: Sensory System Disorders (Ear, Eye, Dental, Skin)