Unsupervised machine learning analysis to identify patterns of ICU medication use for fluid overload prediction

Area of Science:

• Critical Care Medicine
• Pharmacology
• Data Science

Background:

• Intravenous (IV) medications are a significant contributor to fluid overload (FO) in intensive care units (ICUs).
• The precise relationship between IV medication administration patterns and the occurrence of FO remains incompletely understood.
• Understanding these associations is crucial for optimizing patient management and preventing adverse outcomes.

Approach:

• This retrospective cohort study analyzed fluid balance data and medication administration records for adult ICU patients admitted for at least 72 hours.
• Principal Component Analysis (PCA) and Restricted Boltzmann Machine (RBM) were employed to cluster IV medications based on administration timing and volume.
• Statistical methods, including the Wilcoxon rank sum test, were used to compare medication clusters between patients with and without FO.

Key Points:

• Fluid overload (FO) occurred in 13.7% of the study cohort.
• Ten distinct IV medication clusters were identified, with Cluster 7 showing the strongest association with FO (p<0.0001).
• Cluster 7 medications, including continuous infusions, antibiotics, and sedatives/analgesics, significantly improved FO prediction when added to existing models (AUROC 0.65).

Conclusions:

• Machine learning effectively identified a specific cluster of IV medications strongly linked to FO development in ICU patients.
• Integrating this medication cluster into predictive models enhances the accuracy of FO risk assessment.
• This approach holds potential for real-time clinical applications to facilitate early detection of FO and improve patient outcomes.