Real alerts and artifact in continuous non-invasive vital sign monitoring: Mono-vs. Multi-process

Abstract

Introduction:
Currently 65% of single parameter VS alarms are due to artifact and result in alarm fatigue. If our hypothesis that artifact and true instability discriminate as mono- vs multi-process VS movement, such information could inform monitoring systems with improved discrimination.

Methods: We prospectively recruited admissions for 8 weeks in a 24 bed trauma unit. Noninvasive VS monitoring data recorded at a frequency of 1/20Hz consisted of heart rate (HR), respiratory rate (RR; bioimpedance), noninvasive (oscillometric) systolic (SBP) and diastolic (DBP) blood pressure, and peripheral oximetry (SpO2). VS events were detected as any VS violation of stability thresholds (HR< 40 or >140, RR< 8 or >36, SBP < 80 or >200, DBP>110, SpO2< 85%). Events were determined as multi-process if at least two VS exhibited either movement beyond stability thresholds or a changing trend, quantified by computation of a normalized score based on the tail probability of mean value in the event period with respect to the distribution of VS values from the prior 10 minutes. Events were mono-process if no other VS exhibited threshold violation or changing trend. Simultaneously, time plots of all VS parameters during events were visually adjudicated by two reviewers (MRP, MH) as real alerts or artifact based on experiential judgment. Results: 308 admissions and >29,000 patient-hours of monitoring data were studied, yielding 812 VS events, of which 26% were judged artifact and 74% real alerts. 632 (78%) of the VS events were multi-process. The probability of the event being a real alert was 76% and 63% for multi- and mono-processes, respectively. The separation was most obvious with RR, where the probability of RR being a real alert was 67% and 40% for multi-and mono-process respectively. However, discrimination was poor for HR events (97% multi- and 95% mono-process).

Conclusions: Determining the mono-or multi-process composition of VS events may be of moderate assistance in discriminating real alerts from artifact, with differences occurring amongst specific VS. RR alerts are more likely to be true when coupled to other VS changes, whereas HR changes tended to be true even if isolated. Accordingly, using this approach, additional VS features can be readily interrogated for their ability to inform artifact detection.