Introduction and Background
Fluid bolus therapy (FBT) is a medical treatment used to quickly administer a large volume of intravenous (IV) fluids to a patient (infusion rate > 248 ml/hr). The purpose of FBT is to rapidly increase the circulating blood volume and improve the patient’s hemodynamic status, particularly when there is a significant deficit of fluids in the body or an acute drop in blood pressure (mean atrial pressure ≤ 65 mmHg). One challenge currently faced in intensive care units (ICUs) by clinicians is achieving optimal fluid balance, which plays an important role in ensuring the body’s organs have enough blood and nutrients, maintain the body’s pH level, regulate blood pressure and body temperature, and more. Clinicians wish to avoid hypovolemia, insufficient blood volume, as well as hypervolemia, an excess of fluid in the body, as both can result in severe outcomes, including death. Traditional approaches rely on clinical expertise and empirical rules, which can be subjective and vary among healthcare providers.
Methods
In a prior study, an approach involving time-aggregated regression models and attention-based recurrent neural networks (RNNs) was proposed for predicting responses to a medical intervention in a specific patient population. We extend this exploration by investigating different algorithms related to patient fluid balance and further exploring the application of other types of neural networks for time-series modeling. While our analytical methodology is different, the primary objective is to improve clinical interoperability, providing healthcare professionals with a better understanding of predictive factors to save patients lives.
Data for our study were obtained from a widely-used medical database, accessible through PhysioNet. The dataset includes de-identified health data for a significant number of patients admitted to a medical center over a specified timeframe. The dataset encompasses various types of information, such as patient demographics, vital sign measurements, laboratory results, procedures, medications, medical notes, imaging reports, and mortality data. Our study cohort was created using criteria similar to a previous study for analysis.
Following cohort selection, we develop a prototype neural network model with a time-series, utilizing a set of features derived from the medical database.
Any outcomes from this research project are confidential until the study is published. Please contact Dr. Chel Hee Lee for questions.