Predict In-Hospital Code Blue Events using Monitor Alarms through Deep Learning Approach.

Abstract

Bedside monitors in hospital intensive care units (ICUs) are known to produce excessive false alarms that could desensitize caregivers, resulting in delayed or even missed clinical interventions to life-threatening events. Our previous studies proposed a framework aggregating information in monitor alarm data by mining frequent alarm combinations (i.e., SuperAlarm) that are predictive to clinical endpoints, such as code blue events, in an effort to address this critical issue. In the present pilot study, we hypothesize that sequential deep learning models, specifically long-short term memory (LSTM), could capture time-depend features in continuous alarm sequences preceding code blue events and these features may be predictive of these endpoints. LSTM models are trained from continuous alarm sequences in various window lengths preceding code blue events, and the preliminary results showed the best performance reached an AUC of 0.8549. With the selection of optimal cutoff threshold, the 2-hour window model achieved 85.75% sensitivity and 72.61% specificity, respectively.