Research

An artificial intelligence tool to predict fluid requirement in the intensive care unit: a proof-of-concept study

Leo Anthony Celi¹ , L Christian Hinske² , Gil Alterovitz³ and Peter Szolovits⁴

¹  Laboratory of Computer Science, Massachusetts General Hospital, 50 Staniford Street, 7^th floor, Boston, MA 02114, USA

²  Decision Systems Group, 900 Commonwealth Avenue, 3rd Floor, Boston, MA 02215, USA

³  Children's Hospital Informatics Program, Enders Building 6th Floor, room 624.1, 320 Longwood Avenue, Boston, MA 02115, USA

⁴  The Stata Center, Building 32, 32 Vassar Street, Cambridge, MA 02139, USA

author email corresponding author email

Critical Care 2008, 12:R151doi:10.1186/cc7140

Published:	1 December 2008

See related commentary by Lane and Boyd, http://ccforum.com/content/13/1/111

Abstract

Introduction

The goal of personalised medicine in the intensive care unit (ICU) is to predict which diagnostic tests, monitoring interventions and treatments translate to improved outcomes given the variation between patients. Unfortunately, processes such as gene transcription and drug metabolism are dynamic in the critically ill; that is, information obtained during static non-diseased conditions may have limited applicability. We propose an alternative way of personalising medicine in the ICU on a real-time basis using information derived from the application of artificial intelligence on a high-resolution database. Calculation of maintenance fluid requirement at the height of systemic inflammatory response was selected to investigate the feasibility of this approach.

Methods

The Multi-parameter Intelligent Monitoring for Intensive Care II (MIMIC II) is a database of patients admitted to the Beth Israel Deaconess Medical Center ICU in Boston. Patients who were on vasopressors for more than six hours during the first 24 hours of admission were identified from the database. Demographic and physiological variables that might affect fluid requirement or reflect the intravascular volume during the first 24 hours in the ICU were extracted from the database. The outcome to be predicted is the total amount of fluid given during the second 24 hours in the ICU, including all the fluid boluses administered.

Results

We represented the variables by learning a Bayesian network from the underlying data. Using 10-fold cross-validation repeated 100 times, the accuracy of the model in predicting the outcome is 77.8%. The network generated has a threshold Bayes factor of seven representing the posterior probability of the model given the observed data. This Bayes factor translates into p < 0.05 assuming a Gaussian distribution of the variables.

Conclusions

Based on the model, the probability that a patient would require a certain range of fluid on day two can be predicted. In the presence of a larger database, analysis may be limited to patients with identical clinical presentation, demographic factors, co-morbidities, current physiological data and those who did not develop complications as a result of fluid administration. By better predicting maintenance fluid requirements based on the previous day's physiological variables, one might be able to prevent hypotensive episodes requiring fluid boluses during the course of the following day.