Automatic identification of artifacts in monitoring critically ill patients

Abstract

Introduction
We present an analytic system that differentiates true alerts from artifacts in multivariate vitals measured in critically ill patients. It selects easy to interpret projections of data that allow clinicians to validate the alerts. It allows designing reliable decision rules that can be used to identify and dismiss false alerts in real time. It also reduces data review and annotation efforts by experts.

Objectives
To support adjudication of alerts from monitoring critically ill patients, and to reduce false alert rates while maintaining sensitivity of detection.

Methods
Our noninvasive vital sign waveform data includes ECG-derived heart rate (HR) and respiratory rate (RR), oscillometric systolic (SBP) and diastolic (DBP) blood pressure, and peripheral arterial oxygen saturation by finger plethysmography (SpO2). Each exceedence of local stability criteria (HR<40 or >140, RR<8 or >36, systolic BP <80 or >200, diastolic BP>110, SpO2<85%) is assigned a type consistent with the first vital signal that exceeds control limits. Raw data is processed to extract features independently from each vital during the alert and a short window (4 minutes) preceding its onset. The features include common statistics such as mean, standard deviation, minimum, and maximum, and features inspired by domain expertize: data duty cycle (% of non-missing measurements during the alert period), the minimum and maximum of the first order differences and the slope of a linear fit to data inside the alert time window. We model alert-artifact classification separately for RR, BP and SPO2 alerts using RECIP algorithm1. It relies on point estimators for conditional entropy and recovers a desirably small set of projections which accurately classify test alerts. New alerts can be adjudicated using one of the projections from the retrieved set. Results
Models for RR, BP and SPO2 alerts show leave one out cross validation accuracy, precision, and recall of 97.8%/97.9%/99.1%, 88.6%/89.6%/95.8% and 91.2%/91.8%/99.6% respectively. The learned models are generally consistent with clinical intuition. They also provide novel insight, e.g. some highly explanatory projections of data that isolate artifacts due to one particular vital sign often use only features derived from other signs. This suggests existence of informative correlations between vitals that could be leveraged to inform false alert mitigation in existing monitoring systems. The presented method can achieve that goal and also showcase instances of mislabeling of alerts as artifacts and vice versa. It enables new and effective decision rules to filter out artifacts using simple multivariate metrics such as ratios of features extracted from different vitals.

Conclusions
Identification of artifacts in real time high frequency vital sign data can be handled automatically and in a human understandable fashion. It reduces false alert rates without deteriorating the ability to detect true instability. Adjudication of alerts can be made intuitive and highly explanatory, and new multivariate decision rules can be developed for reliable real time filtration of artifacts in waveform vital sign measurements.