A causal machine learning framework for ICU discharge decision support using electronic health records

Discharging patients from the intensive care unit (ICU) requires balancing the risks of premature transfer against the costs and capacity constraints of prolonged ICU stay. While early discharge may increase the risk of deterioration and readmission, delayed discharge strains critical care resources. We aimed to develop and evaluate a computationally efficient causal machine learning framework to estimate ICU discharge decisions associated with lower 30-day readmission risk using observational data. We analyzed two large public electronic health record databases, MIMIC-IV (45,281 ICU stays) and eICU (38,967 ICU stays), for model development and external validation. The framework combined inverse propensity weighting with doubly robust estimation, implemented using gradient boosting decision trees and logistic regression. The system was designed for computational efficiency, requiring less than 4 GB of RAM and under 2 h of training time on standard CPU hardware. The models demonstrated good discriminatory performance, with AUROC values of 0.847 in MIMIC-IV and 0.821 in eICU. In cohorts with baseline 30-day ICU readmission rates of 8.7% (MIMIC-IV) and 7.3% (eICU), counterfactual evaluation of the learned discharge policy projected relative reductions of 18.3% in the MIMIC-IV cohort and 15.7% in the eICU cohort, corresponding to absolute risk reductions of approximately 1.6 and 1.1% points, respectively. Temporal patterns of vital signs, laboratory measurements, and fluid balance were among the features most influential in policy assignment. This study demonstrates that policy learning methods can be applied to routinely collected ICU data to estimate counterfactual discharge strategies in a computationally efficient manner. While the findings are derived from retrospective observational data and rely on modeling assumptions, they suggest potential value for informing ICU discharge decision support. Prospective evaluation will be required to assess clinical impact and safety.