Artificial intelligence to reduce practice variation in the ICU

Introduction

Practice variation refers to how identical patients with exactly the same comorbidities and clinical presentation are managed differently by clinicians. It is increasingly recognized as a major contributor to inefficiencies in healthcare delivery as well as medical errors. The ICU is a place where significant practice variation exists. In this project, the concept of extracting information from a large database in order to reduce practice variation and facilitate care customization is explored. Calculation of maintenance fluid requirement was selected to investigate the feasibility of this approach.

Methods

The Multi-parameter Intelligent Monitoring for Intensive Care (MIMIC II) database consists of >18,000 ICU patients admitted to the Beth Israel Deaconess Medical Center since 2003. Patients who were on vasopressors >6 hours during the first 24 hours of admission were identified. Demographic and physiologic variables that might affect fluid requirement or reflect the intravascular volume were extracted. We represented the variables by learning a Bayesian network from the underlying data.

Results

The network generated has a threshold Bayes factor of 7 representing the posterior probability of the model given the observed data. This minimum Bayes factor translates into P < 0.05. The variables from day 1 that correlated with the fluid intake on day 2 are the number of vasopressors, mean blood pressure, mean heart rate, mean creatinine and day 1 fluid intake. The probability that a patient will require a certain amount of fluid on day 2 can be predicted based on the values of the five variables. In the presence of a larger database, analysis may be limited to patients with good clinical outcomes; that is, resolution of lactic acidosis, improvement in organ dysfunction, survival to discharge.

Conclusion

By better predicting maintenance fluid requirements based on the previous day's physiologic variables, one might be able to prevent hypotensive episodes requiring fluid boluses during the course of the following day. The application of machine learning to a large high-resolution database may also facilitate a more evidence-based customization of care by limiting the analysis to patients of certain demographics and/or those with specific clinical presentations.