Machine learning for alarm management in hospital ward environments

The diversity and number of different alarms in contemporary hospital settings is substantial. A majority of these alarms are either false or technical events, contributing to alarm fatigue. The desensitization of clinicians to alarms compromises patient safety and hinders the efficiency of patient care. This thesis presents a machine learning approach to reduce alarm burden by recommendation of personalized alarm limits based on retrospective annotation of alarm events. One-hour-long windows of continuously monitored vital sign trends in ward patients were used as features for each alarm event, with the alarm onset at 45 minutes. Thus, each window spans from 45 minutes before, to 15 minutes after an event. An autoencoder with long short-term memory cells was used to reduce the dimensionality of the feature vector and the events were classified by a downstream random forest classifier. The encoder--classifier correctly classifies all events that were considered relevant by clinicians, and 60\% of false alarms are correctly identified. Thus, the reduction of alarm burden is possible by leveraging the model in making informed configuration changes. The classifier is utilized in an alarm threshold recommendation system that optimizes the sensitivity and specificity of the alarms for each patient individually. The system analyzes patient-specific alarm patterns and vital sign trends to recommend delay and threshold settings that minimize false alarms, while ensuring timely detection of clinically significant events. The positive results in leveraging AI in alarm management of continuously monitored ward patients shows potential in reducing false alarms by utilizing contextual information of multiple vital signs.