Deep Learning Based Classification of True/False Arrhythmia Alarms in the Intensive Care Unit

Abstract

Once a cardiac alarm is triggered in the intensive care unit (ICU), accurately classifying whether the alarm is true of false is of critical importance. Incorrect classification may lead to patient's death if the alarm is true or to disruption in patient care if false. There has been a body of research, as signified by the 2015 PhysioNet/CinC Challenge; due accomplishments have been made in the relevant computational technology, and yet the highest accuracy known thus far is in the mid-80% range (85%). Our work achieved much higher accuracy and, additionally, very early classification almost at the onset of an arrhythmia alarm, by utilizing state of the art deep learning methods. The machine learning model used is a Residual Network (ResNet) and a Bi-directional Long Short Term Memory (BiLSTM) connected in tandem. Using the Phy-sioNet dataset of 750 recorded ECG segments published with the Challenge, our method performed the classification with 96% accuracy in 0.52 seconds from the onset of an alarm on average over all test ECG segments.