Cumulative Time Series Representation for Code Blue prediction in the Intensive Care Unit

Rebeca Salas-Boni; Yong Bai; Xiao Hu

Cumulative Time Series Representation for Code Blue prediction in the Intensive Care Unit

AMIA Jt Summits Transl Sci Proc. 2015 Mar 25:2015:162-7. eCollection 2015.

Authors

Rebeca Salas-Boni¹, Yong Bai², Xiao Hu³

Affiliations

¹ Department of Physiological Nursing, University of California, San Francisco, CA.
² Department of Bioengineering, University of California, Los Angeles, CA.
³ Department of Physiological Nursing, University of California, San Francisco, CA ; Department of Neurosurgery, University of California, San Francisco, CA ; Institute for Computational Health Sciences, University of California, San Francisco, CA ; Affiliate, UCB/UCSF Graduate Group in Bioengineering, University of California, San Francisco, CA.

PMID: 26306261
PMCID: PMC4525278

Abstract

Patient monitors in hospitals generate a high number of false alarms that compromise patients care and burden clinicians. In our previous work, an attempt to alleviate this problem by finding combinations of monitor alarms and laboratory test that were predictive of code blue events, called SuperAlarms. Our current work consists of developing a novel time series representation that accounts for both cumulative effects and temporality was developed, and it is applied to code blue prediction in the intensive care unit (ICU). The health status of patients is represented both by a term frequency approach, TF, often used in natural language processing; and by our novel cumulative approach. We call this representation "weighted accumulated occurrence representation", or WAOR. These two representations are fed into a L1 regularized logistic regression classifier, and are used to predict code blue events. Our performance was assessed online in an independent set. We report the sensitivity of our algorithm at different time windows prior to the code blue event, as well as the work-up to detect ratio and the proportion of false code blue detections divided by the number of false monitor alarms. We obtained a better performance with our cumulative representation, retaining a sensitivity close to our previous work while improving the other metrics.