Background: Too many unnecessary alarms in the intensive care unit are one of the main reasons for alarm fatigue: Medical staff is overburdened and fails to respond appropriately. This endangers both patients and staff. Currently, there are no algorithms that can determine which alarms are clinically relevant and which are not.
Objective: This paper presents a computer-aided method to automatically determine whether and which interventions followed an alarm. Our algorithm annotates a large data set of oxygen saturation alarms. Previous studies only presented analyses on smaller data sets of manually annotated alarms. Future research can use our large data set of labelled alarms to train machine learning models, for example for alarm prioritisation.
Methods: We propose an alarm annotation algorithm that can efficiently label oxygen saturation alarms from respiratory alarm management by actionability. This algorithm is based on an alarm annotation guideline and works on data from 1961 patients from the hospital information system recorded 06/2019–06/2021. The algorithm analyses a pre-defined time frame after an alarm to determine whether an intervention followed or not. The resulting data set can be used to train machine learning models that predict alarm actionability.
Results: Our open-source algorithm is the first to create a large data set of around 2.5 million relevance-annotated alarms in mere hours. A task that would take years using manual annotation. Our algorithm denotes about 9% of the alarms as actionable. This is in line with previous research. The data set also shows which respiratory management interventions medical staff used to counteract the cause of an alarm.
Conclusion: The data set can be a starting point to reduce the number of unnecessary oxygen saturation alarms. For example, it can serve as a training data set for machine learning models that assess future alarms. The algorithm might be re-used to annotate other alarm data sets as well.
