Detecting Artifactual Events in Vital Signs Monitoring Data

In this study we collect and annotate data from 27 intensive care unit (ICU) patients from the Southern General Hospital in Glasgow. Two models are compared for the detection, removal and cleaning of artifact in the vital signs data, namely the Factorial Switching Linear Dynamical System (FSLDS) and the Discriminative Switching Linear Dynamical System (DSLDS). We also consider a combination of the two, called the α-mixture (as described in sec. 7.3). Three types of artifactual events are considered: blood sample, damped trace (in the arterial line), and suction events. The area under ROC curve (AUC) scores for the detection of these events are: blood sample 0.95, damped trace: 0.79, suction 0.64 (α-mixture), with similar results for the FSLDS and DSLDS. The system is able run in realtime, and we discuss issues that had to be addressed to achieve this.