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

Abstract

Background: Fluid overload (FO) in the intensive care unit (ICU) is common, serious, and may be preventable. Intravenous medications (including administered volume) are a primary cause for FO but are challenging to evaluate as a FO predictor given the high frequency and time-dependency of their use and other factors affecting FO. We sought to employ unsupervised machine learning methods to uncover medication administration patterns correlating with FO.

Methods: This retrospective cohort study included 927 adults admitted to an ICU for ≥72 h. FO was defined as a positive fluid balance ≥7% of admission body weight. After reviewing medication administration record data in 3-h periods, medication exposure was categorized into clusters using principal component analysis (PCA) and Restricted Boltzmann Machine (RBM). Medication regimens of patients with and without FO were compared within clusters to assess their temporal association with FO.

Results: FO occurred in 127 (13.7%) of 927 included patients. Patients received a median (interquartile range) of 31(13-65) discrete intravenous medication administrations over the 72-h period. Across all 47,803 intravenous medication administrations, 10 unique medication clusters, containing 121 to 130 medications per cluster, were identified. The mean number of Cluster 7 medications administered was significantly greater in the FO cohort compared with patients without FO (25.6 vs.10.9, p < 0.0001). A total of 51 (40.2%) of 127 unique Cluster 7 medications were administered in more than five different 3-h periods during the 72-h study window. The most common Cluster 7 medications included continuous infusions, antibiotics, and sedatives/analgesics. Addition of Cluster 7 medications to an FO prediction model including the Acute Physiologic and Chronic Health Evaluation (APACHE) II score and receipt of diuretics improved model predictiveness from an Area Under the Receiver Operation Characteristic (AUROC) curve of 0.719 to 0.741 (p = 0.027).

Conclusions: Using machine learning approaches, a unique medication cluster was strongly associated with FO. Incorporation of this cluster improved the ability to predict FO compared to traditional prediction models. Integration of this approach into real-time clinical applications may improve early detection of FO to facilitate timely intervention.