Early Detection of Hypotension in ICU Settings Using Supervised Classification Algorithms
DescriptionAbnormal trends in blood pressure (BP) are common in critically ill patients and may lead to various upcoming physiological events such as hypotension. Early prediction of hypotensive events in the intensive care units (ICUs) or operating rooms (ORs) allows clinicians to pre-emptively treat the patient and avoid possible organ damage due to the low blood perfusion. However, the real-time prediction of hypotension is a challenge. This is because of the dynamic changes in patients’ physiological status under the drug administration, which also limits the amount of useful physiological data available for the algorithm. To address this challenge, we proposed to use short-term physiological history to predict hypotensive events. We trained a machine learning algorithm to predict hypotension ICUs based on only 5 minutes of patient’s physiological history. To mimic the real-time monitoring application, we labeled the majority of available data points from patients’ physiological record to train and test the algorithm. Further, we investigated the performance of various supervised Machine learning classiﬁcation algorithms along with the proposed real-time labeling technique. It is shown that logistic regression and SVM (support vector machine) yields a better combination of speciﬁcity, sensitivity and PPV (positive predictive value). Logistic regression is able to predict 85% of events within 30 minutes of their onset with 96% speciﬁcity, 81% PPV, while SVM results in 96% speciﬁcity, 83% sensitivity and 82% PPV. Throughout this study, we used publicly available minute-by-minute numerical signals from the MIMIC III database.