Predicting the Need For Vasopressors in the Intensive Care Unit Using an Attention Based Deep Learning Model

Abstract

Background: 

Previous models on prediction of shock mostly focused on septic shock and often required laboratory results in their models. The purpose of this study was to use deep learning approaches to predict vasopressor requirement for critically ill patients within 24 h of intensive care unit (ICU) admission using only vital signs.

Methods: 

We used data from the Medical Information Mart for Intensive Care III database and the eICU Collaborative Research Database to develop a vasopressor prediction model. We performed systematic data preprocessing using matching of cohorts, oversampling, and imputation to control for bias, class imbalance, and missing data. Bidirectional long short-term memory (Bi-LSTM), a multivariate time series model, was used to predict the need for vasopressor therapy using serial physiological data collected 21 h prior to prediction time.

Results: 

Using data from 10,941 critically ill patients from 209 ICUs, our model achieved an initial area under the curve of 0.96 (95% CI 0.96–0.96) to predict the need for vasopressor therapy in 2 h within the first day of ICU admission. After matching to control class imbalance, the Bi-LSTM model had area under the curve of 0.83 (95% CI 0.82–0.83). Heart rate, respiratory rate, and mean arterial pressure contributed most to the model.

Conclusions: 

We used Bi-LSTM to develop a model to predict the need for vasopressor for critically ill patients for the first 24 h of ICU admission. With attention mechanism, respiratory rate, mean arterial pressure, and heart rate were identified as key sequential determinants of vasopressor requirements.