Artifact patterns in continuous noninvasive monitoring of patients

Abstract

INTRODUCTION
Instability can be missed in monitored patients. One contributing factor is alarm fatigue when false alarms originating from monitoring artifact desensitizes clinicians to real alarms. Based on our clinical observation, we hypothesized that artifact is not random, but for each monitored parameter manifests in a small subset of patterns.

OBJECTIVES
We aimed to describe monitor artifact patterns, extract their numeric features, and develop featurized rules for each pattern. Such information could be used to decrease false alarms by automated artifact filtering, and assist clinicians to target action to eliminate artifact.

METHODS
We prospectively recruited admissions for 8 weeks in a 24 bed trauma unit. Noninvasive vital sign (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 deviations (events) beyond stability thresholds (HR< 40 or >140, RR< 8 or >36, SBP < 80 or >200, DBP>110, SpO2< 85%) were visually adjudicated by two reviewers (MRP, MH) as real alerts or artifact. The reviewers developed a limited set of expert rules describing perceived patterns in the artifacts. For each parameter type of artifact (HR, RR, SpO2, etc) we extracted numeric features in the VS artifact signal data as guided by the expert rules (e.g. data density of signal; gap to next time stamp, signal slope). We visualized artifact events for each parameter type in a 1D or 2D space spanned by the extracted features and derived decision boundaries for numeric featurized rules that discriminated between artifact patterns and evaluated which patterns were most useful in achieving discrimination, and whether any cases did not fit any pattern. RESULTS
308 admissions and >29,000 patient-hours of monitoring data were studied, yielding 812 events of VS beyond stability thresholds. Of these, 214 events (26%) were judged to be artifact. Artifacts were most common in RR (65%), followed by SpO2 (17%), DBP (11%), SBP (5%) and HR (2%). All RR artifacts were captured by 2 featurized patterns, with the majority (91%) associated with a pattern of abnormal RR in the absence of HR signals, and the remaining 9% due to oscillatory/sparse signal patterns without change in other VS. All SpO2 artifacts were captured by 3 patterns, with the majority (57%) associated with a pattern of oscillatory step increases and decreases, 24% with a pattern of abrupt step increase at conclusion, and 19% due to sparse signal, also without changes in other VS.

CONCLUSIONS
VS artifact events are common in monitored data, and most artifacts follow predictable patterns for specific VS. Building monitoring systems and sensors to detect artifact based on featurized numeric rules for common artifact patterns could minimize false alarms and improve monitoring utility and patient safety.