Technology, Computing, and Simulation
Section Editor: Jeffrey M. Feldman
An Evaluation of a Novel Software Tool for Detecting
Changes in Physiological Monitoring
J. Mark Ansermino, MBBCh,
MSc(Inf), FRCPC*
Jeremy P. Daniels, BASc, EIT*
Randy T. Hewgill, MD, FRCPC*
Joanne Lim, MASc*
Ping Yang, MASc†
Chris J. Brouse, MASc†
Guy A. Dumont, PhD, PEng†
John B. Bowering, MD, FRCPC*
BACKGROUND: We have developed a software tool (iAssist) to assist clinicians as they
monitor the physiological data that guide their actions during anesthesia. The
system tracks the statistical properties of multiple dynamic physiological processes
and identifies new trend patterns. We report our initial evaluation of this tool (in
pseudo real-time) and compare the detection of trend changes to a post hoc visual
review of the full trend. We suggest a combination of criteria by which to evaluate
the performance of monitoring devices that aim to enhance trend detection.
METHODS: Nineteen children and 28 adults consented to be included in the study,
encompassing more than 68 h of anesthesia. In each surgical case, an anesthesiolo-
gist reported all perceived clinical changes in monitoring in real-time. A trained
observer simultaneously documented the verbally reported changes and every
anesthesiologist action. The same cases were subsequently evaluated offline (in
pseudo real-time) by a novel software tool (iAssist). Heart rate, end-tidal carbon
dioxide, exhaled minute ventilation, and respiratory rate were modeled using a
dynamic linear growth model whose noise distribution was estimated by an
adaptive Kalman filter based on a recursive expectation-maximization method.
Changes were detected by adaptive local Cumulative Sum testing. Changes in the
mean arterial noninvasive blood pressures and oxygen saturation were detected
using adaptive Cumulative Sum testing on a filtered residual from an exponen-
tially weighted moving averaging filter. In post hoc analysis, each change detected
by iAssist was graded independently by two clinicians using a graphical display of
the whole case. Missed changes were recorded.
RESULTS: The iAssist software tool detected 869 true positive changes (at an average
of 12.76/h) with a sensitivity of 0.91 and positive predictive value of 0.87. The post
hoc review identified 91 missed changes (at an average of 1.34/h), resulting in an
overall ratio of true positive rates to false-negative rates of 9.55. The clinicians in
real-time reported 209 changes in trend (at an average of 3.07/h).
CONCLUSION: The algorithms perform favorably compared with a visual inspection
of the complete trend. Further research is needed to identify when and how to
draw the clinician’s attention to these changes.
(Anesth Analg 2009;108:873–80)
Technology designed to assist anesthesiologists in
improving and maintaining vigilance during moni-
toring tasks represents a potential mechanism by
which anesthetic care can be improved. A recent
review concluded that computerized monitoring
has significant potential to improve clinical moni-
toring performance.1
Observers are typically required to monitor dis-
plays over extended periods and to execute overt
detection responses to the appearance of low-
probability critical signals. The signals are usually
clearly perceivable when observers are alerted to
them, but they can be missed in the operating
environment.
Context-Sensitive Monitoring and Trend Detection
The clinical diagnosis of an abnormal physiological
state is rarely dependent on the current observation of
a single physiological variable. For example, a systolic
blood pressure reading of 100 mm Hg cannot be
interpreted in isolation but rather in light of previous
recordings and, most importantly, the trend of record-
ings over time must be considered. In practice, the
expert clinician is able to identify patterns and
changes in single (or multiple) variables over time.
The anesthesiologist judges the current observation in
comparison with previous observations, previous ex-
perience, knowledge about the patient, and rules of
thumb (e.g., that a 20% decrease in arterial blood
From the *Department of Anesthesiology, Pharmacology and
Therapeutics, and †Electrical and Computer Engineering, Univer-
sity of British Columbia, Vancouver, Canada.
Accepted for publication October 30, 2008.
Supported by a Collaborative Health Research Project grant
from the Canadian Institutes of Health Research and the Natural
Sciences and Engineering Research Council of Canada.
Address correspondence and reprint requests to J. Mark Anser-
mino, MBBCh, MSc(Inf), FRCPC, Department of Anesthesia, BC
Children’s Hospital, 4480 Oak St., Vancouver, BC V6H 3V4, Canada.
Address e-mail to anserminos@yahoo.ca.
Copyright © 2009 International Anesthesia Research Society
DOI: 10.1213/ane.0b013e318193ff87
Vol. 108, No. 3, March 2009 873

pressure is significant). The observation is then inter-
preted, with sensitivity to the context in which it has
been obtained, to provide an early warning of poten-
tial deterioration in clinical condition or to reach a
specific diagnosis.
Humans are limited in their ability to accurately
and continuously analyze large amounts of data. The
challenge then is to develop a computer application
that will accumulate all the information on a variable
over time and identify when the trend in observations
has changed. The threshold for providing an alert to
the clinician should not be fixed as in current applica-
tions; rather, it should be dynamically dependent on
all the information available before that observation.
The use of a fixed percentage change as a threshold
for identifying an abnormal observation, as is current
clinical practice, performs poorly, because a percent-
age change is dependent on the level of the compari-
son baseline (the clinical importance of 20% of 180 mm
Hg is very different from 20% of 80 mm Hg). Alter-
natively, the baseline can be compared with the aver-
age value over an interval such as the previous 10 min
(an interval that is arbitrarily chosen). For these meth-
ods, each observation is equally valued in calculating
the comparison observation, whereas in fact the more
recent information is usually more relevant. Using a
measure of statistical distribution, such as standard
deviation, as the comparison observation has similar
limitations.
The mathematical process of detecting trend
changes has been extensively studied in many areas,
such as price-trend prediction and weather forecast-
ing.2 Predictions can be made for the next observation
or observations further in the future. In essence, when
the prediction accurately describes the observation
based on information from all the previous observa-
tions, the trend is unchanged. Alternatively, when the
prediction differs from the observations, it is likely
that the most current observations do not follow the
previous pattern and the trend is changing. The accu-
mulation of differences between the predictions and
observations is used to identify instances where a
significant change in trend has occurred. A similar
mathematical process can be used to identify trend
changes in physiological monitoring.
Physiological observations fluctuate due to the
natural rhythms of the human body and due to
external influences such as drug administration or
surgical stimulation. The changes in trend that are
different from the natural fluctuations over time are
constrained by the physiological processes that pro-
duce the observations. For example, the heart rate
(HR) cannot increase or decrease by 50 bpm in 1 s. The
pattern of change in different physiological variables
differs (e.g., changes in HR have different patterns of
change than those seen in peripheral oxygen satura-
tion [Spo2]). The patterns of change can be used to
distinguish a true change from an artifact. These
patterns of change can be captured using physiologi-
cally based models.
Algorithms implemented on computers provide the
opportunity to detect subtle changes in time-series
data3 that can exceed humans’ discrimination abilities,
especially when the trends are gradual.4 Automated
detection of changes in trend values, as detected by
adaptive algorithms over time, have the potential to
enhance the anesthesiologists’ performance during
periods of sustained monitoring. In this article, we
report on the performance of algorithms designed to
detect subtle trends in six physiological variables: HR,
mean noninvasive arterial blood pressure (NIBPmean),
Spo2, end-tidal carbon dioxide (ETco2), exhaled minute
ventilation (MVexp), and respiratory rate (RR) by com-
paring the performance of the algorithms to a post hoc
review of the visual trend.
METHODS
Algorithm Design and Construction
An abbreviated explanation of the methods used
for detecting a change in trend follows. More detailed
technical descriptions are provided elsewhere.5–7 The
variables studied were chosen to reflect those com-
monly collected and documented during routine
anesthesia.
The Cumulative Sum (CUSUM), a method widely
adopted in process control and monitoring, is a tech-
nique in which the differences between successive
observed values and some target value (a predicted
value in this case) are accumulated. The standard
CUSUM test compares the CUSUM values with a
threshold. A change point is reached when the accu-
mulated value breaks the threshold.
The methods used in this study to detect changes in
HR, ETco2, MVexp, and RR predict future observations
and update the CUSUM threshold in real-time. The
observations are treated as a series of linear segments,
the true levels of which are contaminated by noise
such as measurement error, artifact, and physiological
variations (e.g., respiration). When the current obser-
vation is an extension of the previous pattern, the
increment from the previous observation is similar to
the increment at the previous observation. However,
when the observations start to change into a different
pattern, either the mean level or the slope will be
different. The relationship is described by a matrix-
formatted model called the Dynamic Linear Model.8
Based on this model, an adaptive Kalman filter is used
to estimate the mean and the incremental rate from all
previous observations. The standard Kalman filter
provides the optimal prediction by minimizing the
squared estimation error. Furthermore, it can recur-
sively update the estimates when a new observation
becomes available. In the method used in this study,
the standard Kalman filter is improved by adapting
the filtering process to the signal characteristics. The
adaptive Kalman filter extracts the level of the noise
874 Software Detection of Physiological Changes ANESTHESIA & ANALGESIA