Sequential Rule Analysis of ICU Patient Vital Signals and Alarms

Abstract

In Intensive Care Units (ICUs), excessive medical alarms can cause alarm fatigue and desensitization, compromising patient safety. Alarm management is typically based on manual threshold adjustments, while advanced algorithmic solutions remain underused due to the complexity of patient conditions, dynamic environments, and missing contextual data. Our goal is to investigate the diagnostic utility of combining multiple signals and alarms to enhance relevance and minimize false alarms. A major challenge is integrating heterogeneous data sources, as vital signs are continuously sampled while alarms are event-driven. To bridge this gap, we encode the ICU data into a discretized symbolic representation, reducing dimensionality and improving pattern discovery. We propose a methodology for extracting sequential rules from multivariate datasets, structuring data into a sequence database using a sliding window transformation to capture temporal dependencies. To improve robustness, we introduce a rule ensemble approach, integrating patterns discovered across multiple representations. We applied our method to ICU data from 604 patients, incorporating continuous vital signs and alarm logs. Our findings reveal interpretable sequential rules, analyzed with clinical experts, including patterns highly relevant to intubation events. Our results highlight the potential of data-driven approaches to refine alarm management and improve patient monitoring in critical care settings.