Author: Andrea Villaroman
Multiple modalities of patient data are now more available to hospitals, most notably in the form of electronic health records (EHR) and physiological monitoring data. In acute care settings, these can provide a wealth of information but can also compromise patient safety via alarm fatigue and information overload to clinicians. In our previous work, we have shown that combinations of patient monitor alarms can predict in-hospital code blue events two hours ahead of time with a sensitivity of 80%. These combinations, called SuperAlarm patterns, are generated by first encoding alarms using a class-attribute contingency coefficient (CACC) algorithm and then by mining sequences of alarms that are more frequent in code blue patients over matched control patients at the University of California, Los Angeles (UCLA). To expand upon this work, we have also demonstrated that adding a new data modality, laboratory test results, enhanced the predictive power of using SuperAlarm patterns in predicting code blue events at two different institutions. Data between UCLA and the University of California, San Francisco (UCSF) were encoded into a single schema and SuperAlarm patterns were subsequently mined using maximal frequent itemset (MAFIA) algorithm. Using the new SuperAlarm patterns on a simulated online analysis, code blue events were predicted two hours ahead of time with a sensitivity of 90%. Moreover, we demonstrated there was a significant advantage in using SuperAlarm’s sequential patterns over using raw alarms, potentially due to the filtering of false or irrelevant alarms when the data is pre-processed using the SuperAlarm algorithm.  The incremental development of the SuperAlarm algorithm has been demonstrated in our previous work. However, we have limited our study to monitor alarms, which reflect abnormal physiological signals, and lab test results. Building on the multi-institute database developed for our previous studies, we plan to incorporate several other key data modalities, including advanced electrocardiogram (ECG) metrics, vital signs, medication administrations, procedure orders, and nurse and physician notes. Furthermore, we limited our markers of clinical deterioration to code blue events as the most extreme and reliably documented hospital event. In the context of clinical deterioration, the algorithm specificity can be improved by marking new key endpoints of interest, including severe sepsis, unplanned intubation, and neonatal code blue events. We have already begun identifying these endpoints through the use of procedure codes or other surrogate data and are involving expert clinicians in a comprehensive chart review of event candidates. Currently, the SuperAlarm II database contains all patients who were in any one of 5 UCSF adult intensive care units from March 2013 through December 2015 (269 code blue patients). UCLA patients in this database were admitted from March 2010 through June 2012 (176 code blue patients and 1,766 control patients), with plans to expand the UCLA database to December 2015. For each patient, comprehensive EHR data were extracted from the hospital data warehouse, including demographics, all lab test results, nurse-charted vital signs, and diagnoses. Physiological data were extracted from a central repository archiving patient monitor data, including alarms and physiological waveforms. With a larger dataset, more clinical endpoints, and many possibilities for algorithm refinement using additional machine learning techniques, the SuperAlarm algorithm can improve hospital monitoring by extruding critical early warning patterns for clinicians that would signal clinical deterioration.
Co Author/Co-Investigator Names/Professional Title: Andrea Villaroman, MTM, UCSF Richard Fidler, PhD, Assistant Professor, UCSF Xiao Hu, PhD, Associate Professor, UCSF