The Value of Short-term Physiological History and Contextual Data in Predicting Hypotension in the ICU Settings

Abstract

Hypotension frequently occurs in intensive care units (ICUs) and is correlated to worsening patient outcomes. In this study, we propose a machine learning (ML) algorithm that predicts hypotensive events in ICUs by extracting the information from patients' contextual data and physiological signals. The algorithm uses patients’ history including demographics, pre-ICU medication, and pre-existing comorbidities, and only five minutes of prior physiological history to predict hypotension up to 30 min in advance. We show that adding demographic information to the physiological data does not improve the algorithm's predictive performance of 84% sensitivity, 89% positive predictive value (PPV), and 98% specificity. Furthermore, the results show that including features extracted from patients’ pre-ICU medications and comorbidities lowers the learning algorithm’ prediction performance and leads to 2% degradation in its F1-score. The feature importance analysis showed that the ratio of MAP to HR (MAP2HR) and the average of RR intervals on the ECG (RRI), both extracted from physiological signals, have the highest weights in the prediction of hypotension.