A diverse array of continuous, multi-parameter and alarm-equipped physiologic monitoring devices have been deployed in modern intensive care units (ICUs) and other critical care settings to detect changes in a patient's status. Alarm signals activated by the monitors are intended to alert caregivers to either abnormalities in a patient's normal state or device malfunctions in order to prevent adverse events, and hence improve quality of care and patient safety. The majority of patients who eventually experience adverse events such as in-hospital cardiac arrest (IHCA) frequently exhibit signs of clinical deterioration that are evidence in symptoms and abnormalities in the physiological vital signs and laboratory test results preceding the events. Unfortunately, the signs of deterioration are often unrecognized and missed by caregivers due to the widespread and well-documented alarm fatigue problem, which is chiefly attributable to the excessive number of false and nuisance alarms generated by the physiologic monitors and subsequently leads to the failure-to-rescue (FTR) problem.
The overarching goal of the present dissertation is to predict patient deterioration, particularly code blue events and offer a potential solution for alarm fatigue problem by leveraging monitor alarms available from physiologic monitors and laboratory test results available in the electronic heath record (EHR) system. Several studies are performed in this dissertation to achieve this objective.
First, caregivers in the ICUs undergoing alarm fatigue are typically overloaded by the overwhelming number of heterogeneous raw data that have not been utilized effectively to uncover the underlying knowledge of patient deterioration. To overcome the issue of data overload, we have developed a data fusion framework to identify multivariate combinations of monitor alarms and laboratory test results that co-occur high frequently in a time window preceding code blue events but rarely among control patients. We proposed two approaches to integrate laboratory test results with monitor alarms. We exploited the maximal frequent itemset algorithm to mine the multivariate combinations in a time window preceding code blue events. The resultant combinations are further filtered out if they also occur sufficiently often among all control patients. The remainder of combinations are termed "SuperAlarm patterns".
Moreover, deploying SuperAlarm patterns to monitor patients and detection of the emerging ones can alert caregivers to the changes in the patient's status. The emerging SuperAlarm patterns are termed "SuperAlarm triggers". The consecutive SuperAlarm triggers over time generate "SuperAlarm sequences". Whereas, the SuperAlarm triggers may have redundancy that may also lead to alarm fatigue. Thereby, we developed a sequence classifier to recognize temporal patterns in SuperAlarm sequences. The sequence classifier essentially functions as a filter of SuperAlarm triggers. In addition, we tested the hypothesis that SuperAlarm sequences may contain more predictive temporal patterns than monitor alarms sequences. We proposed novel method to sample subsequences, and utilized the term frequency inverse document frequency (TFIDF) to represent the subsequences. We used the information gain (IG) to selection the most relevant SuperAlatm patterns to the code blue events, and the weighted support vector machine (SVM) to perform classification. The results have demonstrated that sequence classifier based on SuperAlarm sequences outperformed that based on monitor alarm sequences.
Furthermore, a large-scale, comprehensive patient dataset is required for the development and evaluation of advanced SuperAlarm models and algorithms. To fulfill this need, we have created a SuperAlarm study database by consolidating and aggregating a large volume of physiologic and clinical temporal data. The SuperAlarm study database included patient demographics, admission-discharge-transfer (ADT) information, monitor alarms, laboratory test results, physiologic waveforms and vital signs that were collected from a cohort of a large amount of identified adult coded patients and control patients admitted to UCLA and UCSF medical centers. We designed naming codebooks to map and unify alarms and laboratory tests extracted from the two institutions. We also developed a software application to extract physiological waveforms and vital signs and save them into binary files for further analysis.
Finally, we have proposed a novel representation method to convert SuperAlarm sequences into fixed-dimensional vectors, called Time Weighted Supervised Sequence Representation (TWSSR). Unlike TFIDF representation method, the TWSSR is not only a supervised weighting scheme that takes into account the distribution of sequences between coded patients and control patient, it also considers the impact of time on the weight of a SuperAlarm trigger that occurs in a SuperAlarm sequence. We used the monitor alarms and laboratory test results in the established SuperAlarm study database to mine SuperAlarm patterns and further generate SuperAlarm sequences. The support vector machine based recursive feature elimination (SVM-RFE) algorithm was applied to perform classification in conjunction with feature selection. The results suggested that the performance of the sequence classifier based on the TWSSR representation method was higher than that based on TFIDF method.
In summary, we have proposed the SuperAlarm framework to integrate heterogeneous EHR and patient monitoring data to develop predictive models. This framework recognizes patterns not only across different data modalities but also across the temporal dimension through sequence classification. The SuperAlarm framework by means of data fusion could provide a potential paradigm that transforms the critical care patient monitoring into a more integrated precise system for recognizing adverse events and ensuring prompt interventions and treatments, and subsequently improve patient monitoring.