Graph-Based Polypharmacy Risk Networks for Adverse Event Prediction in ICU Patients: A MIMIC-IV Analysis

<title>Abstract</title> <p>—Polypharmacy– the concurrent use of multiple medications– is common in intensive care and is a recognized driver of adverse events, yet most machine-learning studies on MIMIC-IV target mortality prediction rather than modeling drug-interaction risk directly. We address this gap by constructing a population-level drug co-administration graph from 9,974 adult ICU stays and comparing three patient-level feature representations– node2vec graph embeddings, hand-crafted graph-theoretic statistics, and a one-hot tabular baseline– for predicting three polypharmacy-relevant outcomes: acute kidney injury (AKI, KDIGO-based), a delirium proxy, and a bleeding proxy. Across four classifiers and stratified 5-fold cross-validation, we find a mixed but interpretable pattern: the tabular baseline wins when given unrestricted access to drug identity (AKI AUROC0.763, bleeding AUROC 0.864), while graph embeddings outperform a leakage-corrected tabular baseline for delirium (AUROC 0.876 vs. 0.858, paired bootstrap ∆=+0.018, 95% CI [0.011, 0.025], p &lt; 0.0001), replicating on an independent, zero-overlap second subsample (∆=+0.015). We further report a label-leakage correction identified during development (delirium label circularity with tabular drug features, corrected AUROC 0.944 → 0.858) and perturbation-based, reliability-filtered drugrisk rankings. This comparison is evidence for when graph-based representations add value over tabular baselines in EHR-based polypharmacy modeling.</p>