Causality-informed models for multidrug-resistant organism prediction: Enabling effective ICU antibiotic stewardship

Antimicrobial resistance, particularly from multidrug-resistant organisms (MDROs), remains a global health crisis that escalates clinical and economic burdens. To facilitate effective antibiotic stewardship, identifying causality-informed predictors while accounting for the temporal dependencies of clinical data is essential for developing more robust and actionable predictive models.

This study aims to (1) identify and quantify causality-informed predictors of MDRO, and (2) evaluate the performance of 5 predictive models for MDRO-positive cultures using 2 real-world datasets.

We identified causality-informed variables within the Markov boundary of each MDRO to characterize local dependencies. Temporal measurements of these selected variables from the first 24 hours of ICU admission were then combined with static features to predict MDRO culture positivity during ICU stays.

Across 5 MDRO types, 32 causality-informed features were identified. Validated on MIMIC-IV and MIMIC-III datasets, the Bi-GRU model outperformed competitors, achieving AUROCs of 0.8261-0.8280 and AUPRCs of 0.5227-0.5375 (improving upon XGBoost by up to 1.03% and 25.09%, respectively), while maintaining AUROC >0.75 across all MDRO subtypes. Notably, models trained on the causality-informed reduced feature set performed comparably to or better than those using the full feature set.

The results highlight the capacity of causality-informed features to reduce spurious correlations. This work offers mechanistic insights to guide ICU antibiotic stewardship.