Machine Learning Improves Fluid Overload Prediction in ICU

Originally published by our sister publication Specialty Pharmacy Continuum

By Gina Shaw

A machine learning (ML)-identified medication cluster more accurately predicts fluid overload (FO) in ICU patients than traditional prediction models, according to a new study from the MRC-ICU Investigator Team, a multicenter group of researchers focused on data-driven, optimal pharmacotherapeutic care for critical care patients (Pharmacotherapy 2025;45[2]:76-86). 

This result also suggests that artificial intelligence could be used to identify other medication-related predictive risk factors in critical care that may be less straightforward than FO, the investigators noted.

“There has been a lot of interest in using AI for prediction models in the ICU, but a lot of the models that have been developed include factors that are fairly easy to get from the medical record, such as lab values or diagnosis codes,” said study co-investigator Kelli Henry, PharmD, a critical care clinical pharmacist in the medical ICU (MICU) at Wellstar MCG Health, in Augusta, Ga. “Few of the prediction models have included medications, and those that have incorporate only very basic facts like the name of the medication and the dose—not necessarily factors such as when they received the medication, whether it is an appropriate dose for the patient based on renal function, and so on. So, [these models] have been fairly simplistic.”

The MRC-ICU team decided to assess whether inclusion of more detailed medication data could improve the accuracy of commonly used methods such as the Sequential Organ Failure Assessment and APACHE II scores in predicting mortality and other outcomes. “Those scores do not include medication data. If we can use AI to incorporate that data, maybe we can make our predictions stronger,” Dr. Henry said.

The investigators chose FO as an outcome in part to test the strength of the ML model on well-understood risk factors. “We know that the medications on this list make a lot of sense: medications that are fluids, medications that cause renal injury, and continuous infusions would all be associated with FO,” Dr. Henry said. “We wanted to see if an AI model could pick up on that.”

The retrospective cohort study included 927 adults admitted to an ICU for 72 hours or more, with FO occurring in 127 (13.7%) of patients. After reviewing medication administration record data in three-hour periods, medication exposure was categorized into clusters. Across all 47,803 IV medication administrations, 10 unique medication clusters, containing 121 to 130 medications per cluster, were identified, with Cluster 7 standing out. The mean number of Cluster 7 medications administered was significantly greater in the FO cohort than in patients without FO (25.6 vs.10.9; P<0.0001).

“The medications in that cluster matched with a lot of the things we expected to see: continuous infusions, vasopressors, fluids themselves and some nephrotoxic medications,” Dr. Henry said.

“Then we took Cluster 7 and added it to our traditional prediction models, including APACHE II [score] and number of diuretics, and did a regression analysis. When we put all of those together, we were able to more accurately predict which patients had FO.” Model predictiveness increased from an area under the receiver operating characteristic curve of 0.719 to 0.741 (P=0.027), she and her colleagues reported.

The investigators plan to validate their findings using a larger data set. “We also want something that can be integrated into the EHR [electronic health record] in real time to use that cluster to identify high-risk patients and send an alert to the provider, so that they can clinically evaluate them and make a management decision based on those alerts.”

The team also hopes to use these findings to help develop other AI-based tools for medication outcome prediction. “For this study, we picked an outcome where we knew what we were looking for, as a proof of concept,” Dr. Henry said. “This makes us more confident that we can use a similar approach in the future for different outcomes that aren’t as clearly related to medications in our surface assessment.” 

For example, she suggested, “Maybe there are medication-related factors in certain outcomes, such as nephrotoxicity, that we don’t yet know to look for, but AI can detect. A tool like this could help with our workload and direct clinical attention to patients who need it most or whose conditions are changing.”

Dr. Henry reported no relevant financial disclosures.