Clinical evaluation of algorithms for context-sensitive physiological
monitoring in children
M. Dosani1, J. Lim1, P. Yang2, C. Brouse2, J. Daniels1, G. Dumont2 and J. M. Ansermino1*
1Department of Anesthesiology, Pharmacology and Therapeutics and 2Department of Electrical and
Computer Engineering, The University of British Columbia, Vancouver, Canada
*Corresponding author: Department of Anaesthesia, BC Children’s Hospital, 1L7-4480 Oak Street, Vancouver, BC,
Canada V6H 3V4. E-mail: anserminos@yahoo.ca
Background. Subtle changes in monitored physiological signals might be used to guide clinical
actions and give early warning of potential adverse events. Automated early warning systems
could enhance the clinician’s interpretation of data by instantaneously processing new infor-
mation and presenting it within the context of previous observations. In this study, we tested
algorithms for tracking the behaviour of dynamic physiological systems and automatically
detecting key events over time.
Methods. Algorithms were activated in real-time during anaesthesia to run context-sensitive
monitoring of six variables (end-tidal PCO2, heart rate, exhaled minute ventilation, non-invasive
arterial pressure, respiratory rate, and oxygen saturation), alongside standard physiological
monitors. The clinical evaluation included real-time feedback on each change point (change
in the physiological trend) detected by the algorithms and the completion of a usability
questionnaire.
Results. Fifteen anaesthetists completed the evaluation during paediatric surgical cases.
A total of 38 cases were evaluated, with a mean duration of 103 (102) min. The mean number
of change points per case was 22.8 (23.4). Sixty-one per cent of all rated change points were
considered clinically significant, and ,7% were due to artifacts.
Conclusions. The algorithms were able to detect a range of clinically significant physiological
changes during paediatric anaesthesia, and were considered useful by participating anaesthe-
tists. These findings indicate that automated detection of context-sensitive changes is possible
and could be used by early warning systems during physiological monitoring. Further investi-
gations are required to assess how this information can best be communicated to the
anaesthetist.
Br J Anaesth 2009
Keywords: computers; equipment, computers; equipment, expert adviser system; model,
mathematical; monitoring, intraoperative
Accepted for publication: February 22, 2009
Current physiological monitoring technologies, com-
bined with clinical information systems, collect large
amounts of data at the bedside. Typically, if the value of
a monitored variable strays outside a preset range, an
alarm will be triggered. Unfortunately, false and
unnecessary alarms are so frequent that, in practice, they
can represent a nuisance rather than improvement to
monitoring.1 Major advances in sensor technology have
resulted in increased information available to the anaes-
thetist. Despite this, the means of extracting important
information to advise clinicians of significant events is
still underdeveloped.
Early warning systems generate scores from a composite
of current observations, which can be used to trigger rapid
intervention for patients at risk. Such systems have been
implemented within the past decade, but recent investi-
gations have shown poor diagnostic accuracy, partially
attributable to the selection of physiological variables and
cut-off values, mistakes in the manual calculation of
scores, and intra- and inter-rater reliability errors.2
# The Author [2009]. Published by Oxford University Press on behalf of The Board of Directors of the British Journal of Anaesthesia. All rights reserved.
For Permissions, please email: journals.permissions@oxfordjournal.org
British Journal of Anaesthesia Page 1 of 6
doi:10.1093/bja/aep045
BJA Advance Access published March 26, 2009

Although rigorous investigations have shown improvement
in the selection and validation of appropriate physiological
variables and cut-off values,3 the distinction between
normal and abnormal values is complex due to the pro-
found variability in physiological measurements between
and within patients.
The likelihood that a patient is at risk, based on the
information available at that point in time, is rarely absol-
ute. It is more common for this likelihood to follow a
probability distribution, with the likelihood of correctly
identifying a patient at risk increasing as more information
becomes available. The use of previously recorded infor-
mation, rather than only the current observation, could sig-
nificantly improve the usefulness of early warning systems
both in the operating theatre and in acute care environ-
ments. Automated early warning systems can enhance the
clinician’s ability to interpret data by instantaneously pro-
cessing new information and presenting it to the clinician
within the context of previous observations.
In this study, we tested algorithms for tracking the beha-
viour of dynamic physiological systems and automatically
detecting key events in the processes over time. After
extensive offline testing4 and refinement, we now report
the results of evaluation in real-time during anaesthesia.
Methods
Context-sensitive monitoring
The mathematical process of detecting trend changes has
been extensively studied in many areas such as price-trend
prediction and weather forecasting.5 A similar mathemat-
ical process can be used to identify trend changes in phys-
iological monitoring. For context-sensitive monitoring, the
threshold for providing an alert to the clinician is not
fixed, but is dynamically dependent on the information
available before an observation. Predictions based on
information from all the previous observations are made
for the next observation or observations further in the
future. When the prediction accurately describes the obser-
vation, the trend is unchanged. Alternatively, when the
prediction differs from the observation, the trend is chan-
ging. The accumulation of differences between the predic-
tions and the observations is used to identify whether a
significant change in trend has occurred. The point at
which the accumulation of differences exceeds a threshold
is called a change point.
Physiological models of change
Physiological observations fluctuate due to the natural
rhythms of the human body. The changes in trends, which
differ from natural fluctuations that occur over time, are
constrained by the physiological processes that produce
the observations. For example, heart rate (HR) cannot
increase or decrease by 50 beats min21 in 1 s. The pattern
of change for each physiological variable differs (e.g.
changes in HR have different patterns than those seen in
oxygen saturation). The patterns of change can be captured
using statistical models and can be used to distinguish a
true change from a normal physiological fluctuation or an
artifact.
Change-point detection algorithms
The algorithms in this study predict future observations
and update the Cumulative Sum (CUSUM) threshold in
real-time. The CUSUM test, a method widely adopted in
process control and monitoring, is a technique where the
differences between successive observed values and a
target value (a predicted value in this case) are accumu-
lated. The standard CUSUM test compares the CUSUM
values with a threshold. A change point is reached when
the accumulated value breaks the threshold. Change points
in mean non-invasive arterial pressure (mean NIAP) and
oxygen saturation (SpO2) were detected using CUSUM
testing on a filtered residual from an Exponentially
Weighted Moving Average filter as described in detail
elsewhere.6 The algorithms used for HR, end-tidal carbon
dioxide (E0CO2), exhaled minute ventilation (MVexp), and
respiratory rate (RR) treat the observations as a series of
linear segments, the true levels of which are contaminated
by noise such as measurement error, artifact, and physio-
logical variations (e.g. respiration).7 8 When the current
observation follows the previous pattern, the mean and
slope from the previous observations are similar to the
current observations. However, when the observations start
to change into a different pattern, either the mean or the
slope will be different. The probability of these relation-
ships is described by a matrix-formatted model called the
dynamic linear model.9 On the basis of this model, an
adaptive Kalman filter (a minimum-variance stochastic
estimation procedure) is used to estimate the mean and the
incremental rate from all previous observations. The stan-
dard Kalman filter provides the optimal prediction by
minimizing the squared estimation error. Furthermore, it
can recursively update the estimates when a new obser-
vation 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 adap-
tive Kalman filter extracts the level of the noise and the
degree of signal variability, and then adjusts the weights
assigned to the observations. For example, if the recent
observations are highly corrupted by artifact, less weight
will be assigned to the recent observations with more
influence given to earlier observations. Since the Kalman
filter captures the full monitoring history and assigns more
weight to the more recent or more reliable observations,
the estimate from the Kalman filter is robust against
disturbances.
Dosani et al.
Page 2 of 6