Utilizing Intensive Care Alarms for Machine Learning

Abstract

Alarms help to detect medical conditions in intensive care units and improve patient safety. However, up to 99% of alarms are non-actionable, i.e. alarm that did not trigger a medical intervention in a defined time frame. Reducing their amount through machine learning (ML) is hypothesized to be a promising approach to improve patient monitoring and alarm management. This retrospective study presents the technical and medical pre-processing steps to annotate alarms into actionable and non-actionable, creating a basis for ML applications.

How to Annotate Patient Monitoring Alarms in Intensive Care Medicine for Machine Learning

“…We identi ed 20 studies containing 23 associated alarm annotation reports in our systematic review [3,[15][16][17][23][24][25][26][27][28][29][30][31][32][33][34][35][36][37][38][39][40][41] (details in the PRISMA ow chart, Supplementary Fig. 1 [42]), and summarized the ndings (Supplementary Table 1).…”

“…Although the annotation methods-and outcomes of interest-differed across all studies, most of the reports speci ed de nitions or protocols prior to the annotation process to classify the alarms; 10 reports focused on the classi cation of alarms into true and false [16, 17, 30-33, 39, 40], audible [38] or not, or based on the alarm type [36]. The remaining 13 reports evaluated if the alarm required or was followed by a medical action such as a therapeutic or diagnostic intervention while using different de nitions and terms: "true positive, clinically relevant" [15], "clinically relevant" [3,23,24,34], "relevant"[28], "relevant or true" [25], "true" [26,27], "actionable" [29,35,41], and "consistent." [37] The researchers mostly used data from monitoring systems including waveforms, measurements, alarms, and settings.…”

“…Video recordings aided the annotation in six reports [3, 23-25, 28, 34], and three studies used documentation from medical records, discharge summaries, consultation notes, etc. [28,38,41].…”