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

Area of Science:

• Intensive care medicine
• Data science in healthcare
• Pharmacology

Background:

• Fluid overload (FO) is a common and serious complication in intensive care units (ICUs).
• Intravenous medications are a major contributor to FO, but their complex administration patterns make them difficult to predict.
• Unsupervised machine learning can identify medication administration patterns associated with FO.

Purpose of the Study:

• To apply unsupervised machine learning to uncover medication administration patterns correlating with FO.
• To evaluate the predictive value of identified medication patterns for FO development.

Main Methods:

• Retrospective cohort study of 927 adult ICU patients with stays ≥72 hours.
• FO defined as fluid balance ≥7% of admission body weight.
• Medication administration data analyzed in 3-hour intervals using principal component analysis (PCA) and Restricted Boltzmann Machine (RBM) to identify medication clusters.

Main Results:

• Fluid overload (FO) occurred in 13.7% of patients.
• Ten unique medication clusters were identified from 47,803 intravenous medication administrations.
• A specific medication cluster (Cluster 7), including continuous infusions, antibiotics, and sedatives/analgesics, was significantly associated with FO (mean administrations: 25.6 in FO vs. 10.9 without FO).
• Adding Cluster 7 medications to existing prediction models improved FO prediction accuracy (AUROC from 0.719 to 0.741).

Conclusions:

• Unsupervised machine learning identified a novel medication cluster strongly associated with FO in the ICU.
• This medication cluster significantly improved FO prediction compared to traditional models.
• Integrating this approach into clinical practice may enhance early FO detection and timely interventions.