Risk Prediction of Critical Vital Signs for ICU Patients Using Recurrent Neural Network

Abstract

Monitoring vital signs for Intensive Care Unit (ICU) patients is an absolute necessity to help assess the general physical health. In this research, we use machine learning to make a classification forecast that uses continuous ICU vital signs measurements to predict whether the vital signs of the next hour would reach the critical value or not. With the early warning, nurses and doctors can prevent emergency situations that may cause organ dysfunction or even death before it is too late. In this study, the data includes vital sign measurements, laboratory test results, procedures, medications collected from over 40,000 patients. After data preprocessing, bias data balancing, feature extraction, and feature selection, we have a clean dataset with informative and discriminating features. Then, various machine learning algorithms including Random Forest, XGBoost, Artificial Neural Networks (ANN), and LSTM were developed to predict critical vital signs of ICU patients 1-hour beforehand. We particularly developed predictive models to predict Heart Rate, Blood Oxygen Level (SpO2), Mean Arterial Pressure (MAP), Respiratory Rate (RR), Systolic Blood Pressure (SBP). The results demonstrated the accuracy of the developed methods.