264 Int. J. Biomedical Engineering and Technology, Vol. 3, Nos. 3/4, 2010
Copyright © 2010 Inderscience Enterprises Ltd.



Clinical decision support in physiological monitoring
J. Mark Ansermino*
and Stephan K.W. Schwarz
Faculty of Medicine,
Department of Anesthesiology, Pharmacology & Therapeutics,
The University of British Columbia,
2176 Health Sciences Mall, Vancouver,
British Columbia, V6T 1Z3, Canada
Fax: 604-822-6012
E-mail: anserminos@yahoo.ca
E-mail: stephan.schwarz@ubc.ca
*Corresponding author
Guy A. Dumont, Chris Brouse
and Yang Ping
Department of Electrical and Computer Engineering,
The University of British Columbia,
2332 Main Mall, Vancouver,
British Columbia, V6T 1Z4, Canada
E-mail: guyd@ece.ubc.ca
E-mail: chris.brouse@gmail.com
E-mail: pingy@ece.ubc.ca
Joanne Lim, Dustin Dunsmuir
and Jeremy Daniels
Department of Anesthesia,
British Columbia Children’s Hospital
1L7 – 4480 Oak Street, Vancouver,
British Columbia, V6H 3V4, Canada
E-mail: jlim2@cw.bc.ca
E-mail: ddunsmuir@cw.bc.ca
E-mail: jdaniels2@cw.bc.ca
Abstract: Effective use of the information produced by current and future
physiological sensors can be used to improve monitoring, diagnosis, and
treatment of patients. The successful introduction of intelligent monitoring and
automated control promises to harness this information to enhance safety,
as it has in aviation. Intelligent data analysis requires a multifaceted approach.
A range of techniques that include statistical characterisation, modelling,
feature extraction, and prediction augmented with expert knowledge can be
used to present the salient information and combine information from multiple
sources. We present a collaborative effort to embrace intelligent monitoring
and automated control for physiological monitoring.

Clinical decision support in physiological monitoring 265




Keywords: clinical decision support; expert system; physiological monitoring;
automation; anaesthesia; knowledge authoring; adaptive feature extraction;
patient safety; intelligent data analysis; controlled drug delivery.
Reference to this paper should be made as follows: Ansermino, J.M.,
Schwarz, S.K.W., Dumont, G.A., Brouse, C., Ping, Y., Lim, J., Dunsmuir, D.
and Daniels, J. (2010) ‘Clinical decision support in physiological monitoring’,
Int. J. Biomedical Engineering and Technology, Vol. 3, Nos. 3/4, pp.264–286.
Biographical notes: J. Mark Ansermino is an Assistant Professor in the
Department of Anesthesiology, Pharmacology & Therapeutics, The University
of British Columbia (UBC). He holds the position of Director of Research
for the Department of Anesthesia at British Columbia Children’s Hospital.
He has graduate training in Health Informatics from City University,
London, UK. He holds a Scholar award from the Michael Smith Foundation for
Health Research.
Stephan K.W. Schwarz is an Assistant Professor of Anesthesiology,
Pharmacology & Therapeutics at the UBC and member of the anesthesia
consultant staff at Vancouver’s St. Paul’s Hospital, where he serves as
Anesthesia Research Director. He obtained his Medical Doctorate from the
University of Göttingen, Germany. Subsequently, he completed a PhD in
pharmacology & therapeutics at UBC. His interests have focused on
anaesthetic pharmacology, pain control, and patient safety.
Guy A. Dumont is a Professor and Distinguished University Scholar at
the UBC. He is co-founder of the Electrical and Computer Engineering
in Medicine (ECEM) research group. His primary research interests are in
biosignal processing and automated drug delivery for critical care. He chaired
the 2008 IEEE Engineering in Biology and Medicine Conference and is
an Associate Editor for the IEEE Transactions on Information Technology in
Biomedicine.
Chris Brouse is currently a PhD student with the UBC’s Electrical and
Computer Engineering for Medicine (ECEM) research group. He received an
MASc in Electrical Engineering from the UBC in 2007. His research interests
include biomedical signal processing, software architecture, and health
informatics.
Yang Ping is currently a PhD candidate in the Department of Electrical and
Computer Engineering at the UBC, Vancouver, Canada. She received a
Bachelor’s Degree in Automatic Control and a Master’s Degree in System
Engineering from Xi’an Jiaotong University, China, in 1999 and 2002,
respectively. Her research interests include trend detection, feature extraction,
sensor fusion and decision-support systems in anaesthesia and critical care.
Joanne Lim is a Research Manager in the Department of Pediatric Anesthesia
at British Columbia Children’s Hospital and the Electrical and Computer
Engineering in Medicine group at the UBC. She received her MASc
in Mechanical Engineering from UBC after completion of a BSc (Eng) in
Biological Engineering from the University of Guelph. Her research interests
are broad and span a wide variety of areas including medical simulation and
interface design.

266 J.M. Ansermino et al.




Dustin Dunsmuir is currently a Software Designer in the Pediatric Anesthesia
Research group at British Columbia Children’s Hospital. He is involved in
ongoing work in knowledge management and decision-support systems
designed to aid clinicians. Also a new graduate student from Simon Fraser
University, British Columbia, his primary research interests include visual
analytics, expert systems, and human-computer interaction.
Jeremy Daniels is a second year medical student at the University of British
Columbia, and a Part Time Research Engineer at British Columbia Children’s
Hospital, Department of Pediatric Anesthesia Research. He received his BASc
in Mechanical Engineering from the University of Waterloo in 2005.
1 Introduction
1.1 The promise of Clinical Decision-Support Systems
in physiological monitoring
Advances in technology have resulted in an exponential growth in the amount of
physiological data collected in many healthcare environments including the operating
theatre, the intensive care unit, the ward, and even at home. Improved knowledge
(evidence) is the foundation of improved decision-making; more data has the potential
to significantly improve knowledge.
Physiological monitoring in the future likely will rely on miniaturised, wireless
sensors recording multiple physiological processes non-invasively. This vast array of
data will inform clinical decisions but, similar to aviation, will also initiate increased
automation, especially for safety enhancement. The integration of technology with
human expertise will complement clinician performance by improving knowledge and
optimising tasks humans perform sub optimally, such as continuous observation.
As an example of a complex, high-intensity healthcare environment, the modern
operating theatre is an arena of continuously competing noises, alarms, signals, and
patient data, to which the anaesthetist must instantly attend and respond appropriately.
The monitors collect and display an array of physiological data and are configured in an
attempt to direct the attention of attending clinicians to relevant changes in patient status.
Typically, if the value of a monitored variable strays outside a preset range, an alarm will
be triggered. Despite major advances in sensor technology, the means of advising
clinicians of significant events is still underdeveloped. Clinicians and patients would
benefit significantly from development of a Clinical Decision-Support System (CDSS)
that can harness the potential of new technology in this complex environment.
Outside of medicine, the archetypical example of a successful introduction of
automated control and intelligent monitoring is represented by cockpit automation and
autopilots in the aviation industry. Whereas this technology initially was met with
resistance by pilots, analysis of commercial jet airplane accidents from 1959 to 2006
shows that as the level of automation in the cockpit significantly increased, the number
of hull losses significantly decreased despite an exponential growth in the number of
annual flight departures and flight hours (Boeing, 2007). Although this improvement in
safety is due to many factors beyond cockpit automation and intelligent monitoring,
they have played a part in this improvement in safety. It is also worth noting that this new

Clinical decision support in physiological monitoring 267




technology was introduced without randomised studies to prove its impact on safety.
We hope that automation in anaesthesia will result in similar safety improvements.
Improving the performance of skilled medical professionals is unlikely to be achieved
by exhorting them to work more carefully, more cheaply or more quickly. Clinicians are
already likely to make decisions, plan their time, and remember key information to the
highest degree human ability will allow. There is a compelling argument that advances in
patient monitoring have actually increased the chances of human error – by adding
greater complexity to the monitoring task. In addition, the monitor may become
a distraction when things ‘go wrong’. One of the main causes of this paradox is that,
while a current monitoring system is capable of detecting abnormalities, it may not
ensure that this information is transferred to the clinician in a meaningful manner.
This cognitive burden is expected to increase with the ongoing advent of new monitoring
devices, unless a new means of transforming data into information for the clinician to
use is developed.
1.2 Expert systems for clinical monitoring
Expert systems are well established for the prevention of errors in the aviation and
atomic energy industries (Nabeshima et al., 2003). Human reliability analysis has been
extensively studied and significant gains demonstrated with the use of decision-support
tools (Swain and Guttmann, 1983). The implementation of expert systems in healthcare
has been limited by the complexity of the environment and lack of validation in the
clinical setting.
Expert systems, including CDSS, can assist clinicians in making decisions with
respect to patient care. Unfortunately, clinicians thus far have failed to adopt expert
systems into routine practice. Reasons for this include a lack of understanding of the
systems, peer pressure, physical unavailability, or other barriers such as reduced linkages.
Poor quality outputs, unconvincing information, inability to change practice, or a belief
that performance was already optimal were cited as additional common reasons
for failure of expert systems (Liu et al., 2006a). Key challenges for the future
implementation of a CDSS include the dissemination of best practices in design,
development, and implementation; patient-level information summarisation; prioritisation
of recommendations to the user; architecture for sharing the CDSS knowledge;
combining recommendations with co-morbidities; prioritising content development and
implementation; creating clinical decision-support repositories; the use of free text
information to drive clinical decision support; mining of large clinical databases to create
new knowledge (Sittig et al., 2008).
Physiological monitoring is a data-intensive environment that provides the ideal
setting for demonstrating the clinical benefits of implementing a CDSS. The main
advantages are improved clinician performance by monitoring and analysing
physiological parameters at a faster speed and with more vigilance than a clinician.
This includes improved automated trend detection with early intervention and rescue of
life-threatening events. In addition, a CDSS provides a rich source of knowledge that
can be used to enforce clinical guidelines (Mersmann and Dojat, 2004), provide cognitive
aids (Neily et al., 2007) and encourage learning with ‘just-in-time’ information
(Dunsmuir et al., 2008). A CDSS can promote documentation or even trigger automated
events.

268 J.M. Ansermino et al.




1.3 Key success factors for Decision-Support Systems in clinical monitoring
We have identified a number of limitations with current devices that will become the key
factors in design, implementation and adoption of a CDSS for clinical monitoring.
1.3.1 The balance between sensitivity and specificity
Current alerting modalities deployed in physiological monitors err on the side of caution
by triggering an alarm each and every time a value could possibly be abnormal.
No latitude is given to missing an event. This results in highly sensitive devices with very
low specificity. False alarm prone monitoring degrades performance significantly more
than miss prone monitoring. The false alarms degrade both compliance and reliance
whereas missed events only reduce reliance (Dixon et al., 2007), which may be
advantageous in keeping the human ‘in the loop’. Although a highly sensitive alerting
scheme can be argued as being safe, from a regulatory perspective, it has a very negative
impact on human performance.
1.3.2 More information produces better decisions
The key to success in an intelligent monitoring device is to use all possible sources of
information. The complexity in the variation in normal human physiology, combined
with contamination from clinical interventions and artefacts (such as electocautery and
movement) makes the ability to precisely identify abnormal clinical conditions extremely
challenging. To mathematically reliably differentiate a normal from an abnormal value
in a single variable is an intractable problem. The identification of an abnormal state
requires much more than just selecting the correct threshold or developing the perfect
filter. Additional information, such as contextual information, information from related
variables, and expert knowledge all can significantly improve the ability to discriminate
a normal from an abnormal value. Intelligent alarm systems are possible due to advances
in knowledge-based systems. Domain expert rules can enhance the performance of
clinical alarms (Laramee et al., 2006).
1.3.3 Monitoring context
The implementation of a clinical monitoring expert system will be highly dependent on
the setting of the implementation. The probability of an abnormal event as well as the
probability of signal noise or a planned clinical event must be considered in the
implementation. In anaesthesia, for example, the induction and emergence phases are
much more likely to be contaminated by clinical events compared with the stable period
between these two endpoints. Attempts to bridge these multiple settings with the same
implementation are likely to fail. Providing relevant information about a planned event to
the inference system will improve performance.
1.3.4 System complexity
Clinical adoption of an expert system, like the adoption of technology for any
application, requires that the complexity of the system and its end-user interface
are matched to the value derived from the support provided. Reliable adoption will
require a system that is reliable as well as plausible. Clinicians will need a sound grasp of

Clinical decision support in physiological monitoring 269




the functioning of a system with the ability to interpret output based on additional
information. A ‘black-box’ approach or the use of overly complex reasoning
methodologies can distance the user from the application, inhibiting adoption and
preventing the user from applying additional expert knowledge or reasoning.
1.3.5 Resonable expectations
The requirement to engineer an expert system in a dynamic software environment to the
same technical standards as required for medical hardware devices will make the cost of
systems prohibitive. The fact that some of the information may be imperfect must be
accepted and included in the final decision-making process used by the clinician.
The liability for appropriate treatment must be shared by the system and the clinician.
Even expert clinicians cannot always agree on the optimal way to treat patients.
The clinician must be fully aware of the capabilities and limitations of the system to
ensure the patient is not unnecessarily harmed. Designing systems to share knowledge
between users and the creation of CDSS repositories will enhance efforts to generate
knowledge. Local, regional, national and international systems for governance over the
validity of knowledge are required.
1.4 The human factors in clinical monitoring
The design and development of a CDSS must consider the interaction between the human
and the device providing support. A clear understanding of the varied clinical settings,
competing tasks and human limitations is essential. Support is required in acute critical
situations but also in the identification of rare but critical events in an extended period
of monitoring.
1.4.1 Vigilance
In a vigilance task such as clinical monitoring, operators must detect changes in a steady
stream of data over an extended period of time. Human Factors research has shown that
vigilance decrements over time, and that this drop typically happens during the first
30 minutes of the task (Parasuraman, 1986). Rather than being understimulating,
vigilance tasks are exacting, capacity-draining assignments that are associated with a
considerable degree of mental demand and frustration. Automation may lead to further
decrement in vigilance. Decision-support systems can be designed to reduce operator
memory load and increase sensitivity at vigilance tasks (Schoenfeld and Scerbo, 1999).
Additionally, systems that increase the target salience through feedback can reduce the
vigilance decrement.
1.4.2 Attention
In complex work environments, operators make errors in selecting the focus of attention.
A clinician monitoring a complex patient may fall victim to a cognitive tunnelling error
where inappropriate attention is paid to one malfunctioning subsystem, whereas other
unnoticed and more critical subsystems fail. Decision-support systems designed to direct
attention to where it is needed most at any given time promise to reduce such cognitive
tunnelling errors.

270 J.M. Ansermino et al.




1.4.3 Memory limitations: duration and capacity
Human memory is a finite resource subject to many limitations. Working memory is used
for conscious deliberations about a topic of concern (such as a patient’s current/past vital
signs), and is limited to approximately seven chunks of information at any given time
(Miller, 1956). Long-term memory is used to retrieve information previously learned into
working memory, when it is needed. A decision-support system can support limitations in
human memory by automatically processing vast amounts of information and only
presenting limited chunks of information improving a clinician’s situational awareness
(Endsley and Garland, 2000). It can also cue important information to the clinician’s
working memory when needed.
1.4.4 Risks and benefits of Decision-Support Systems
One of the costs associated with any technology is the complexity it may add to the work
domain, particularly if it malfunctions. A well-functioning computer program may
contain millions of lines of computer code, which may result in errors or ‘bugs’ not
expected by the designers. In fact, in many complex computer programs, it is not possible
to test all possible configurations for system safety (Leveson, 1995). Additionally,
because computerisation allows for the development of algorithms more complex than
those used by humans to arrive at conclusions, decision-support systems may give
optimal but surprising results to the clinicians using them (automation-induced surprises)
(Sarter and Woods, 1997).
Decision-support systems that do not inform the clinicians of how conclusions were
derived may evoke low trust, detracting from their potential to achieve their goals
(Kessel, 2005). On the other hand, decision-support systems may lead to complacency on
behalf of the clinicians using them. Complacent users are at a risk for being unfamiliar
with the system of interest due to being ‘out of the loop’.
Decision-support systems that are introduced into the work environment without due
care will not be used effectively and will put patients at risk. Improper training and
poorly usable interfaces can contribute to poor adoption of technological support.
These systems are then rejected, consequently representing a very poor investment for
the healthcare organisation and clinicians alike (Lapointe and Rivard, 2006).
2 Design of a Clinical Decision-Support System for monitoring
This section describes our research team’s efforts to develop and evaluate a CDSS for
clinical monitoring. We have chosen a multi-faceted approach to the design of a robust
CDSS that has a significant focus on how the information will be integrated into the
workflow of the clinician (Figure 1). The intelligent data analysis includes statistical
characterisation, modelling, feature extraction and prediction augmented with expert
knowledge. Evaluation is emphasised as the essential element to move a CDSS from the
laboratory to the bedside.

Clinical decision support in physiological monitoring 271




Figure 1 A multiple level, integrated process for transforming data into knowledge to be
presented to the clinician (see online version for colours)
2.1 Signal processing
To ensure that a CDSS is effective, it needs high-quality data. A significant problem
when measuring low-voltage physiological signals is noise. This measurement noise
should be removed or at least identified to ensure optimal performance of the CDSS.
Here, we describe examples of the application of advanced signal processing techniques
that may be utilised in conjunction with CDSS for physiological monitoring.
2.1.1 Electrocautery detection
Electrocautery noise is a significant cause of artefact in important physiological signals,
such as the electrocardiogram (ECG), in the operating theatre. The noise is caused by the
electrosurgical unit used to cut tissue and cauterise blood vessels. Electrocautery noise
manifests as high-frequency sinusoids superimposed on the signal in electrical potentials
recorded at the skin surface, such as the ECG (Figure 2). The high power of
electrocautery can also transiently interfere with many other physiological measurements.
Electrocautery noise is a leading cause of false alarms and can impair the
performance of an expert system reliant on data from these sources. To address this
problem, we have developed an algorithm that detects electrocautery activation from the
ECG. It analyses the recorded ECG using the wavelet transform (Brouse et al., 2006).
The vanishing moments, characteristic of wavelets, makes them sensitive to sharp signal
changes (discontinuities) whereas remaining blind to slow changes. As the ECG is mostly
smooth and electrocautery noise is discontinuous, wavelets allow us to discriminate
between clean and noisy ECG signals (Figure 3).

272 J.M. Ansermino et al.




Figure 2 ECG corrupted by electrocautery noise of (a) weak; (b) medium and (c) high power
(see online version for colours)
Figure 3 (a) ECG partially corrupted by electrocautery noise and (b) the threshold wavelet
coefficients (see online version for colours)

Clinical decision support in physiological monitoring 273




The CDSS employs the electrocautery noise detection algorithm to identify when
data quality becomes unreliable, both for the ECG and other physiological signals.
For example, the algorithm analyses the ECG signal to assess signal quality. If the ECG
is ‘clean’ the algorithm provides the CDSS with the ECG-based heart rate. Otherwise,
it provides the heart rate from an alternate source, such as the pulse oximeter or invasive
arterial blood pressure monitoring line, or suppresses the triggering of an alert until the
signal quality improves. The algorithm thus acts as a switch (Figure 4). It can be used,
within limits, to suppress output from a CDSS until the data quality has improved.
Figure 4 The electrocautery noise detector chooses a heart rate source
2.1.2 Sensor fusion
Redundancy exists in the current sensor network, with some signals measured by more
than one independent sensor, or sensors that are related by physiological mechanism or
measurement principle. For example, heart rate is routinely measured from the R wave
peaks of the ECG, the pulse oximetry waveform peaks, or by detecting the peaks
in the invasive blood pressure traces. Each sensor is susceptible to interference from
independent sources. Sensor fusion aims to combine measurements from multiple
sensors, clinical knowledge, and historical observations to provide more robust estimates
of the true levels of the variable than would be possible from a single-sensor source,
especially in the presence of noise.
Sensor fusion starts with detecting artefacts by testing the measurements against prior
knowledge. The residuals are generated by fitting the measurements to the relationships
between the sensors and the trend models. Statistics are designed to test if the deviations
fall within the allowable range. If all the constraints can be described by linear models
and background noise follows a Gaussian distribution, there are well-studied statistical
tools available for artefact detection, such as the chi-square test (a global scheme to
detect the existence of subtle artefacts), the Shewhart chart, or the Cumulative Sum
(CUSUM) test.
The corrupted variables identified in the artefact detection process should be treated
as unmeasured data and rectified with information from related measurements so that
the estimates are consistent with known constraints. This step is referred to as Data
Reconciliation (DR). Constrained optimisation provides a natural framework for DR.
In constrained optimisation, estimation quality is defined as a distance from the original
measurements, referred to as an objective function. Prior knowledge of the static and
dynamic characteristics of the variable and the interrelationships with other variables are
formulated into equality or inequality constraints. Depending on the formation (linearity)

274 J.M. Ansermino et al.




of the constraints and the objective function, the constrained optimisation can be
implemented analytically or iteratively using numerical algorithms.
The performance of a sensor fusion process depends on the amount of redundancy in
the monitoring system in relation to the combined number of unmeasured variables and
corrupted measurements. The amount of redundancy can be increased by adding
independent sensors to the network, or introducing more clinical knowledge into the
constraints. If the number of artefacts is larger than the degree of redundancy,
the credibility of reconciled measurements will be significantly reduced. The CDSS
should be informed of this deterioration to incorporate this into the reasoning process.
2.1.3 Smart sensors
Smart sensors are created with the use of a single or multiple sensor sources to produce
a new data source. The sensor data is interpreted with expert knowledge about the
underlying physiological process. For example, the use of scaled, processed
electroencephalographic (EEG) data can be used to provide an indication of the level of
unconsciousness and guide the administration of anaesthesia (Bibian et al., 2001).
The representation of variability in a signal, such as heart rate, blood pressure, or ‘tidal’
(exhaled) respiratory volume (Barbour et al., 2004) can provide valuable information
about autonomic system function (Winchell and Hoyt, 1996) or the likelihood that
cardiac output will increase with intravenous fluid administration (Tran et al., 2007).
Pulse transit time, the interval between electrical activation in the heart and the arrival of
the pulse wave in the periphery, measured with pulse oximetry, is used as a continuous
non-invasive index of cardiac output (Fung et al., 2004).
2.1.4 Adaptive feature extraction (context-sensitive monitoring)
The context in which data is collected can provide significant insights into the
interpretation of any physiological measurement. The context can include information
about the patient or clinical setting but the temporal dimension is the most important.
Identifying the changes in physiological variables over time (time varying information) is
a fundamental step in intelligent patient monitoring. Tracking the tendencies of
physiological variables can lead to earlier diagnosis and rescue, should an adverse event
occur. The temporal features of each measurement reflect the dynamics in a patient’s
status. These features should be extracted and conveyed to the clinician or translated to
semantic inputs for a higher-level inference module. We have previously described
our work in adaptive feature extraction for patient monitoring (Yang et al., 2005,
2006a, 2006b).
The simplest level of feature extraction is change point detection, which detects the
location of a change in the trend direction. After describing the signal as an exponentially
weighted moving average of historical data, we test the CUSUM of the deviations
between successive measurements and the model predictions, from the last time the
CUSUM was restarted to the current moment, using a V-mask. The V-mask is a visual
procedure proposed by Barnard (1959) to perform change detection with the CUSUM
statistic. The performance in change point detection can be improved by adapting the
setting of the CUSUM test to the characteristics of the monitored signal. We have
proposed an adaptive change point detection method based on a dynamic linear model of
the signal (Yang et al., 2006a). In this model, both the signal level and the speed of trend

Clinical decision support in physiological monitoring 275




change are treated as dynamic processes. An adaptive Kalman filter is used to estimate
the mean and the incremental rate from the observations in the presence of noise.
The adaptive Kalman filter extracts the level of noise and the degree of signal variability.
These extracted features are used to adapt the thresholds of the CUSUM test and
influence the change detection online. The method has been implemented into a software
system and applied to the clinical monitoring of heart rate (Yang et al., 2006a) and
respiratory variables (Yang et al., 2005).
Basic change point detection, as a hypothesis test, answers the question of whether
a trend has changed its direction. The result reduces a high volume of numerical
information to binary points, therefore losing a significant amount of detail about the
trend features. Furthermore, trends of similar features may have different degrees of
clinical significance depending on the pattern of previous trend segments. Clinicians not
only base their decision on the quantitative value of the current measurements but also
intuitively resort to their experience to evaluate their dynamic significance in the context
of the signal history.
The temporal patterns of the trend signals can be sorted into several categories
according to the trend features such as direction, duration, and rate of change. The trend
signals can be viewed as a sequence of segments of different patterns and described by
a two-layer model called the Generalised Hidden Markov Model (GHMM). The Markov
Chain as the top layer describes the influence of the features of the previous segment on
the significance of the current trend. The lower layer describes the variation within each
segment by a stochastic dynamic model. Bayesian reasoning, together with a fixed-point
Kalman smoothing process based on the GHMM, estimates the probability of occurrence
for each pattern dynamically. The algorithm has been tested on non-invasive mean blood
pressure data, and shows improvement in vital sign monitoring (Yang et al., 2006b).
The extracted trend features and significance scores can provide inputs to a
higher-level diagnostic system.
2.2 Data integration
The signal processing steps described allow the extraction of key features from the data.
Data integration is the process of intelligently combining these key features to reach a
conclusion about the processes that produced the data. Once provided the necessary
knowledge, an expert system can perform data integration more accurately and efficiently
than a clinician.
2.2.1 Knowledge authoring
Clinicians combine clinical experience, published evidence, training, and, most
importantly, the observed patterns of multiple variables over time to identify significant
clinical events. The explicit knowledge (formal theoretical learning) and tacit (personal)
knowledge are integrated to make rapid clinical decisions in the interest of patient safety.
However, as with most experts, clinicians are not aware of how tacit knowledge is
gained or used. Initiatives to facilitate the evolution of tacit knowledge into explicit
knowledge are integral to the advancement of a CDSS (Pope et al., 2003). There are
few expert clinicians who are able or willing to translate clinical knowledge into
a computerised form using traditional methods as they lack the training or expertise.
Typically, this knowledge must be gathered through time-consuming interviews and then

276 J.M. Ansermino et al.




encoded by those with the technical skills. We have developed a knowledge authoring
tool, iKnow (Figure 5), to simplify the process of knowledge transfer from experts to the
expert system by providing a computer interface tailored to the clinician (Dunsmuir et al.,
2008). Clinicians must grasp only how knowledge is modelled, and not the computer
code or instructions underlying the model. Through this model, clinicians may specify
patterns in patient data and then define a patient status that can be concluded from these
patterns. The resulting model is then used in the real-time CDSS to inform a clinician
of the patient status when the data matches the specified patterns. A collaborative
development of the knowledge base by local, national and international groups of experts
is anticipated. This will facilitate the sanctioning of rules and avoid the unnecessary
duplication of effort in developing rule sets. Users from different communities may easily
merge separate rule sets to provide a wider range of knowledge to the decision-support
system.
Figure 5 Screenshot of the knowledge authoring tool (see online version for colours)
2.2.2 Rule-based DSS
For a knowledge authoring tool to be usable for the everyday clinician, the data model
must take a form simple enough to understand and at the same time must be complex
enough to adequately model the knowledge. A rule-based model provides a suitable
balance of these criteria. The knowledge base consists of a set of rules and each rule is a
conditional if-then statement based on a range of physiological or demographic inputs.

Clinical decision support in physiological monitoring 277




Each rule contains a list of patterns and an outcome (Tables 1 and 2). Each pattern is a
statement about the value of a physiological variable. The expert system optimises the
comparison of each value to the pattern statements. Rule chaining allows for the encoding
of complex knowledge in simple increments.
Table 1 Sample patient data for rules of Table 2
Parameter Value
Patient age 24 years
HR 120 bpm
NIBPsys 76 mmHg
T1 37.8°C
EtCO2 62 mmHg
Table 2 Sample rules showing those patterns matched (shown as checkboxes) and those
outcomes asserted that are used again in patterns
Rule name Patterns Outcome
Adult Age ≥ 17 years Adult
AdultAdult Bradycardia
20 bpm ≤ HR ≤ 40 bpm
Bradycardia
AdultAdult Tachycardia
100 bpm ≤ HR ≤ 300 bpm
Tachycardia
AdultAdult Hypotension
20 mmHg ≤ NIBPsys ≥ 80 mmHg
Hypotension
Hyperthermia 37.5 C ≤ T1 ≤ 41.0°C Hyperthermia
Hypercapnia 45 mmHg ≤ EtCO2 ≤ 80 mmHg Hypercapnia
Tachycardia
Hypotension
Hyperthermia
Malignant Hyperthermia
Hypercapnia
Malignant Hyperthermia
The current application uses an optimised rule-based reasoning system to encapsulate the
clinician’s expertise. Artificial intelligence methods that have been used in a CDSS
include knowledge-based machine learning, fuzzy logic, neural networks and Bayesian
networks (Imhoff and Kuhls, 2006). Though these methods show promise for future
applications, they have not been utilised in the current implementation. The challenge in
adopting these methods will be to enable the clinician to understand the use of these
reasoning methodologies.
2.2.3 Just-in-time information
In addition to providing an alert to the clinician, the decision rules provide information
to the clinician exactly when it is needed (‘just-in-time’). A clear explanation to the
clinician is essential to support clinical adoption. The explanation provides the clinician

278 J.M. Ansermino et al.




an opportunity to follow the logic the system used to arrive at its conclusion.
Cognitive aids (Neily et al., 2007), check lists, clinical guidelines or care pathways
can also be triggered by the rules.
2.3 Output
2.3.1 Visual
Visual communication remains to be the predominant method currently used to relay
information to the clinician. Current physiological patient monitoring displays follow the
single-sensor, single indicator paradigm, showing one waveform or numeric for each
sensor. Most importantly, all available monitors still require healthcare providers to
integrate multiple sources of pertinent information to make an appropriate clinical
decision. The outcome of a rule in the CDSS typically triggers an alert on the visual
display. This alert may be enhanced with colour or animation. The visual display allows
for the display of large volumes of information or for navigation to further sources
of information. Current displays do not display contextual information that could be used
as cues to the immanent outcome.
We have developed and tested a display that includes context-relevant information
(Tappan et al., 2008). This display resulted in a more rapid response to critical events.
We are hopeful that this benefit will carry over to improved cueing for the CDSS.
Additional design features of the interaction such as salience or the use of multiple
sources (monitors) require further investigation.
2.3.2 Auditory
The sonification (transfer of information with the use of sound) of heart rate (by a change
in the interval between tones) and oxygen saturation (by a change in pitch) is routinely
used to augment information transfer during physiological monitoring. This multi-modal
information transfer assists clinical performance during periods of high cognitive
workload (Ford et al., 2008). The continuous, uninterrupted display of information is a
significant advantage of auditory displays.
The transfer of precise values of oxygen saturation using sonification has been shown
to be poor; a change of 8% in oxygen saturation was required before 95% of anaesthetists
noticed a change (Santamore and Cleaver, 2004). We have recently investigated
the anaesthetist’s ability to detect changes and accuracy in estimating the heart rate
(Chou et al., 2008). No difference was found between the anaesthetists and non-experts in
detecting heart rate changes from the auditory tone with or without a distraction.
Despite these limitations, sonification offers significant potential to deliver
information during physiological monitoring, especially for maintaining peripheral
awareness during other tasks (Sanderson et al., 2008). Sound, ideally more than just an
alarm, can play a role in communicating information from a CDSS to the clinician.
2.3.3 Tactile
In addition to the visual and auditory domains, we have pioneered the relatively
underutilised sense of touch to transmit information during physiological monitoring.
Sensory receptors are stimulated, via a tactile device, to provide pulses of vibration that
correspond to physiological changes in a patient. This has the advantage of using the

Clinical decision support in physiological monitoring 279




body’s largest sensory organ, which responds to stimuli with a high degree of precision,
to provide subtle cues that do not detract from patient observation and will not disturb
other individuals in the clinical environment.
In previous studies, we have investigated the effect of different tactile display
modalities (vibration and electrical stimulation) at different body locations including the
forearm (Ng et al., 2007), wrist (Ng et al., 2005), and abdomen on the impact of encoding
on the reception process, and have proposed an optimal method to design rhythm-based
stimuli.
The sense of touch offers a unique, underutilised resource to communicate
information from the CDSS to the clinician.
2.3.4 Automation
Whereas monitoring is the process of observing a system, control is the process
of guiding a system to a desirable state. Whereas control technology has been applied to
most fields of human activity, it has yet to have a significant impact on the practice
of anaesthesia, despite a number of experimental and clinical studies on closed-loop
control of depth of anaesthesia (Liu et al., 2006b; Struys et al., 2001; Puri et al., 2007).
Automated drug delivery allows exact matching between drug administration and
individual patient’s needs and response (Bibian et al., 2005). Additional potential
advantages include reduced workload for the anaesthetists, allowing them to spend more
time on higher-level tasks, and minimisation of the risk of drug errors. It should be
noted that although published studies have demonstrated the feasibility of automated
drug delivery, they have not yet shown clear benefits in terms of patient outcomes.
The initial systems that we developed, based on feedback control methodology, were
used as advisory systems, i.e., the system recommends an action to the clinician who then
decides whether to implement the action or not. For example, the administration of
phenylephrine to prevent hypotension during spinal anaesthesia for Caesarean section
(Fung et al., 2004). Using pulse transit time to assess blood pressure, the system relied on
adaptive model predictive control to recommend an adequate vasopressor drug
(phenylephrine) dose according to the identified internal patient model. Subsequently,
we developed the Neuromuscular Blockade Advisory System (NMBAS), which is based
on patients’ historical electromyographic responses and adaptive model predictive control
to provide clinicians with advice on the timing and dose of neuromuscular blocking drugs
(Gilhuly et al., 2008a). The NMBAS was limited to bolus administration at no less than
20 minute intervals. No notification was given until the proposed dose corresponded to
a minimum size bolus. Despite the limitations owing to such infrequent control actions,
the system produced significant positive clinical outcomes. NMBAS-guided care was
associated with improved neuromuscular blockade quality and improved indices of
recovery from blockade at the end of surgery, potentially reducing the risk of residual
blockade and improving perioperative patient safety (Gilhuly et al., 2008b).
Advisory systems suffer from a number of drawbacks, however, including limits
on the frequency of drug dose adjustments and the fact that clinicians’ attention must be
directed to the device so as not to miss important interventions. The optimal solution
is a full closed-loop continuous infusion-based system. The problem of reliability of the
closed-loop system, particularly in view of the large intra- and inter-patient variability,
is a concern. These systems need to meet stability and robustness criteria for a realistic
range of variations such as will be met in practice. Robust control design theory is the

280 J.M. Ansermino et al.




only practical way to meet such criteria. Hybrid systems theory can be used to design
verifiably safe systems.
An advantage of closed-loop control in anaesthesia will come from design and
implementation of multi-variable control systems, i.e., systems that take into account the
simultaneous drug interactions, a difficult task for a human operator. Studies in the
aerospace and nuclear energy industries have demonstrated that though automation can
reduce mental workload, it does not eliminate the vigilance decrement discussed earlier.
The benefits of such a system, for both clinicians and patients, can only be confirmed by
rigorous evaluation.
2.4 Integration framework
We have developed an integrated software framework (iAssist) to facilitate clinical
evaluation of the integrated system for real-time physiological monitoring.
The framework accommodates input of data with multiple sources, processing by
multiple signal processing algorithms, integration with a real-time inference engine
(JBoss) and outputs to visual, tactile or auditory data displays or some automated action.
The software can be used both offline (using pre-recorded data) and online (in a real-time
clinical environment). The software’s user interface has been designed for use
with a touch screen, and allows clinicians to give feedback about the performance
of components (Figure 6). Feedback is used to refine performance of components.
The framework is based on Java® SE 6.0, for its cross-platform compatibility and rich set
of foundation classes. The framework is architected around four aspects: extensibility,
flexibility, scalability, and interoperability.
Figure 6 The iAssist user interface. The interface is designed for use with a touch screen and
allows clinicians to give feedback about the performance of individual components
(see online version for colours)

Clinical decision support in physiological monitoring 281




2.4.1 Extensibility
Extensibility is achieved using a framework architecture. Plug-ins, which are small
custom-written components designed to serve a targeted purpose, can be added or
removed as required, without making any changes to the core framework (Figure 7).
Figure 7 The iAssist extensible framework is analogous to a puzzle. Different pieces (plug-ins)
can be connected, provided they have the correct shape (interface)
A high level of abstraction is maintained between the framework and plug-ins.
This serves the dual purpose of keeping the framework versatile, and simplifying plug-in
development. Abstraction is achieved via common interfaces, which define interaction
points between the framework and plug-ins.
2.4.2 Flexibility
Flexibility serves as a mechanism to model plug-in relationships. The plug-ins are linked
together to form user-defined patterns of workflow. Plug-ins can have a Multiple Input,
Multiple Output (MIMO) relationships and interact dynamically. This flexibility supports
complex relationships among signal processing plug-ins with the results of one algorithm
used as input for another algorithm.
The signal processing plug-ins are modelled as a Directed Acyclic Hypergraph
(DAHG), allowing very general arrangements or specific interaction (Gallo et al., 1993).
Interconnections and configurations are stored in XML deployment descriptors.
At run-time, the deployment descriptors dictate plug-in DAHG structure.
2.4.3 Scalability
Scalability allows the software to gracefully accommodate a large number of plug-ins,
and takes advantage of future advances in computing hardware. The architecture is well
suited to highly parallel processing with each instance of a plug-in executed in a
dedicated thread.

282 J.M. Ansermino et al.




2.4.4 Interoperability
The software framework is designed for a high degree of interoperability with other
medical devices (Figure 8). It employs the ISO/IEEE 11073 (X73) international standard
for modelling point-of-care medical device data (ISO/IEEE, 2004a, 2004b, 2004c).
To acquire data from proprietary devices the incoming data is translated into the X73
standard for integration within the framework and linkage to other data sources.
Figure 8 iAssist can interoperate with other medical devices. Custom-written input and output
plug-ins translate between X73 and proprietary standards
2.5 Evaluation
As with the design and development of any human-technology interaction, evaluation of
the performance, benefits, and risks is essential. For a CDSS, the evaluation should
initially concentrate on performance and safety but must also include usability and
interaction evaluations. Evaluating a CDSS in the early stage of development is an
extremely challenging task. It needs to be performed over much iteration with the domain
experts’ imperfect opinion used as a benchmark. The clinically important improvements
(e.g., length of hospital stay, morbidity, or mortality) will require large clinical trials.
A framework for conducting usability evaluations has been described (Daniels et al.,
2007). The software framework described earlier allows rapid prototyping with iterative
phases of testing.
2.5.1 Offline
The initial evaluation of the CDSS should be carried out offline using simulated or
historical data. The software framework facilitates the evaluation, tuning and
improvement of algorithms with the ability to rapidly institute changes and provide
robust evaluation of many thousand iterations per minute.

Clinical decision support in physiological monitoring 283




2.5.2 Simulation
Objective measurements of performance are more readily available in a simulated
environment as compared with obtaining measurements in a real-life clinical
environment. The performance of new instrumentation and technologies can be evaluated
by providing precise control of the simulated environment and avoiding potential hazards
to patients. The evaluation of the CDSS is an iterative process of evaluation and
improvement that is most effectively, efficiently, and safely performed in a simulation
environment.
Simulated scenarios can be high fidelity (e.g., virtual reality, human patient
simulators) or low fidelity (e.g., laboratory, office space, or work environment). The
fidelity level and setting are dependant on the purpose and scope of the evaluation.
Examples of simulation evaluations include auditory (Chou et al., 2008), vibrotactile
(Ng et al., 2007) and visual displays (Tappan et al., 2008). In a recent study using a
high-fidelity human patient simulator mannequin, we designed a scenario containing a
critical incident of anaphylaxis (i.e., a life-threatening allergic reaction) to evaluate the
benefits of a vibrotactile display (Ford et al., 2008). The participants of this study rated
the simulation to be very realistic, and an improved performance was demonstrated in the
group of clinicians wearing the Vibrotactile display compared with routine physiological
monitoring.
2.5.3 Clinical trials
The gold standard for the evaluation of the CDSS, as with any therapeutic modality or
intervention, is the Randomised Controlled Trial (RCT) (Young and Griffiths, 2006).
However, the introduction of current technology for physiological monitoring, including
pulse oximetry and capnometry, was introduced with minimal or no RCT-based evidence
of its benefit (Pedersen et al., 2003). RCTs on technological devices, compared with
drugs, are associated with considerable design and implementation challenges.
These include difficulties in defining appropriate and meaningful study outcome
variables as critical incidents often represent rare occurrences; difficulties to achieve
adequate sample sizes; challenges with blinding and concealment of group allocation
(complete blinding of the attending anaesthetist may pose unacceptable risks); and a lack
of relevant adequate control groups (related, for example, to the difficulty to develop a
true placebo CDSS). It is obviously vitally important to keep the anaesthetist ‘in the loop’
to introduce human wisdom to the process and provide a crucial safety check.
Whereas these difficulties may seem daunting, some of these challenges can be overcome
by appropriate surrogate outcome variable selection, careful blinding of data collectors
and additional ‘study’ clinicians not directly responsible for patient care, and careful
attention to selection of appropriate control groups.
3 Conclusion
Improved use of the data collected by physiological monitors will require an integrated
approach for data reduction and integration. The intervention required to rescue or reduce
harm will require the intervention of a human operator or robust automation.
The anaesthetist is no longer able to monitor all the data produced by the increased

284 J.M. Ansermino et al.




number of sensors at the bedside. Augmenting the performance of the clinician,
by improved communication of information in critical settings, is the primary goal of the
CDSS. Automation can further improve reliability in an increasingly data-rich
environment, as it has in other arenas such as aviation. We have described a multi-level
proof-of-principle CDSS for physiological monitoring that includes advanced signal
processing techniques, a rule-based expert system, and a range of options for human
interaction. A software framework facilitates rapid iterative development, evaluation and
improvement in the laboratory and in real-time clinical settings. The demonstration
of improved patient safety will be required before widespread clinical adoption.
References
Barbour, S.J., Vandebeek, C.A. and Ansermino, J.M. (2004) ‘Increased tidal volume variability in
children is a better marker of opioid-induced respiratory depression than decreased respiratory
rate’, J. Clin. Monit. Comput., Vol. 18, pp.171–178.
Barnard, G.A. (1959) ‘Control charts and stochastic processes’, J. R. Stat. Soc. [Ser. B], Vol. 21,
pp.239–271.
Bibian, S., Ries, C.R., Huzmezan M. and Dumont, G.A. (2005) ‘Introduction to automated drug
delivery in clinical anesthesia’, Eur. J. Control., Vol. 11, pp.535–557.
Bibian, S., Zikov, T., Dumont, G.A., Ries, C.R., Puil, E., Ahmadi, H., Huzmezan, M. and
Macleod, B.A. (2001) ‘Estimation of the anesthetic depth using wavelet analysis of
electroencephalogram’, Conf. Proc. IEEE Eng. Med. Biol. Soc., Vol. 1, pp.951–955.
Boeing Commercial Airplanes (2008) Statistical Summary of Commercial Jet Airplane Accidents.
Aviation Safety, Worldwide Operations 1959–2007, Boeing Commercial, Airplanes, Seattle,
WA, USA.
Brouse, C., Dumont, G.A., Herrmann, F.J. and Ansermino, J.M. (2006) ‘A wavelet approach to
detecting electrocautery noise in the ECG’, IEEE Eng. Med. Biol. Mag., Vol. 25, pp.76–82.
Chou, E., Lim, J., Brant, R., Ford, S. and Ansermino, J.M. (2008) ‘Heart rate in a simulated
operating accuracy of detecting changes in auditory room environment’, Anaesthesia, Vol. 62,
pp.1181–1186.
Daniels, J., Fels, S., Kushniruk, A., Lim, J. and Ansermino, J.M. (2007) ‘A framework for
evaluating usability of clinical monitoring technology’, J. Clin. Monit. Comput., Vol. 21,
pp.323–330.
Dixon, S.R., Wickens, C.D. and McCarley, J.S. (2007) ‘On the independence of compliance
and reliance: Are automation false alarms worse than misses?’, Hum. Factors, Vol. 49,
pp.564–572.
Dunsmuir, D., Daniels, J., Brouse, C., Ford, S. and Ansermino J.M. (2008) ‘A knowledge authoring
tool for clinical decision support’, J. Clin. Monit. Comput., Vol. 22, pp.189–198.
Endsley, M.R. and Garland, D.J. (2000) Situation Awareness, Lawrence Erlbaum Associates,
Mahwah, NJ, USA.
Ford, S., Daniels, J., Lim, J., Koval, V., Dumont, G.A., Schwarz, S.K.W. and Ansermino, J.M.
(2008) ‘A novel vibrotactile display to improve the performance of anesthesiologists in a
simulated critical incident’, Anesth. Analg., Vol. 106, pp.1182–1188.
Fung, P., Dumont, G.A., Ansermino, J.M., Huzmezan, M. and Kamani, A. (2004)
‘Toward an advisory system for cesarean section spinal anesthesia’, Proceedings of the 2004
American Control Conference, Vol. 1, pp.952–957.
Fung, P., Dumont, G.A., Ries, C., Mott, C. and Ansermino, J.M. (2004) ‘Continuous noninvasive
blood pressure measurement by pulse transit time’, Conf. Proc. IEEE Eng. Med. Biol. Soc.,
Vol. 1, pp.738–741.

Clinical decision support in physiological monitoring 285




Gallo, G., Longo, G., Pallottino, S. and Nguyen, S. (1993) ‘Directed hypergraphs and applications’,
Discrete Appl. Math., Vol. 42, pp.177–201.
Gilhuly, T.J., Hutchings, S.R., Dumont, G.A. and Macleod, B.A. (2008a) ‘Development and pilot
testing of the Neuromuscular Blockade Advisory System’, Comput. Methods Programs
Biomed., Vol. 89, pp.179–188.
Gilhuly, T.J., MacLeod, B.A., Dumont, G.A., Bouzane, A. and Schwarz, S.K.W. (2008b)
‘Improved neuromuscular blockade using a novel Neuromuscular Blockade Advisory System
(NMBAS): a randomized controlled clinical trial’, Anesth. Analg, Vol. 107, pp.1609–1617.
Imhoff, M. and Kuhls, S. (2006) ‘Alarm algorithms in critical care monitoring’, Anesth. Analg.,
Vol. 102, pp.1525–1537.
ISO/IEEE (2004a) Health Informatics – Point-of-Care Medical Device Communication – Part
10101: Nomenclature, ISO/IEEE 11073-10101: 2004 (E), IEEE.
ISO/IEEE (2004b) Health Informatics – Point-of-Care Medical Device Communication – Part
10201: Domain Information Model, ISO/IEEE 11073-10201: 2004 (E), IEEE.
ISO/IEEE (2004c) Health Informatics – Point-of-Care Medical Device Communication – Part
20101: Application Profile-base Standard, ISO/IEEE 11073-20101: 2004 (E), IEEE.
Kessel, R.T. (2005) Apparent Reliability: Conditions for Reliance on Supervised Automation,
Defence R&D Canada – Atlantic, Dartmouth, NS, Canada.
Lapointe, L. and Rivard, S. (2006) ‘Getting physicians to accept new information technology:
insights from case studies’, CMAJ, Vol. 174, pp.1573–1578.
Laramee, C.B., Lesperance, L., Gause, D. and Mcleod, K. (2006) ‘Intelligent alarm processing into
clinical knowledge’, Conf. Proc. IEEE Eng. Med. Biol. Soc., Supplement, pp.6657–6659.
Leveson, N.G. (1995) Safeware: System Safety and Computers, Addison-Wesley, New York.
Liu, J., Wyatt, J.C. and Altman, D.G. (2006a) ‘Decision tools in health care: focus on the problem,
not the solution’, BMC Med. Inform. Decis. Mak., Vol. 6, pp.1–7.
Liu, N., Chazot, T., Genty, A., Landais, A., Restoux, A., McGee, K., Laloe, P-A., Trillat, B.,
Barvais, L. and Fischler, M. (2006b) ‘Titration of propofol for anesthetic induction and
maintenance guided by the Bispectral Index: closed-loop versus manual control: a prospective,
randomized, multicenter study’, Anesthesiology, Vol. 104, pp.686–695.
Mersmann, S. and Dojat, M. (2004) ‘SmartCare™ – automated clinical guidelines in critical care’,
ECAI, Vol. 16, pp.745–749.
Miller, G.A. (1956) ‘The magical number seven’, Psychol. Rev., Vol. 63, pp.81–97.
Nabeshima, K., Suzudo, T., Seker, S., Ayaz, E., Barutcu, B., Turkcan, E., Ohno, T. and Kudo, K.
(2003) ‘On-line neuro-expert monitoring system for Borssele nuclear power plant’,
Prog. Nucl. Energ., Vol. 43, pp.397–404.
Neily, J., Derosier, J.M., Mills, P.D., Bishop, M.J., Weeks, W.B. and Bagian, J.P. (2007)
‘Awareness and use of a cognitive aid for anesthesiology’, Jt. Comm. J. Qual. Patient Saf.,
Vol. 33, pp.502–511.
Ng, G., Barralon, P., Dumont, G., Schwarz, S.K.W. and Ansermino, J.M. (2007) ‘Optimizing the
tactile display of physiological information: vibro-tactile vs. electro-tactile stimulation, and
forearm or wrist location’, Conf. Proc. IEEE Eng. Med. Biol. Soc., pp.4202–4205.
Ng, J.Y., Man, J.C., Fels, S., Dumont, G. and Ansermino, J.M. (2005) ‘An evaluation of a
vibro-tactile display prototype for physiological monitoring’, Anesth. Analg., Vol. 101,
pp.1719–1724.
Parasuraman, R. (1986) ‘Vigilance, monitoring, and search’, Handbook of Perception and Human
Performance, Vol. 2, pp.43–51.
Pedersen, T., Moller, A.M. and Pedersen, B.D. (2003) ‘Pulse oximetry for perioperative
monitoring: Systematic review of randomized, controlled trials’, Anesth. Analg., Vol. 96,
pp.426–431.
Pope, C., Smith, A., Goodwin, D. and Mort, M. (2003) ‘Passing on tacit knowledge in anaesthesia:
a qualitative study’, Med. Educ., Vol. 37, pp.650–655.

286 J.M. Ansermino et al.




Puri, G.D., Kumar, B. and Aveek, J. (2007) ‘Closed-loop anaesthesia delivery system (CLADS)
using bispectral index: a performance assessment study’, Anaesth. Intens. Care, Vol. 35,
pp.357–362.
Sanderson, P.M., Watson, M.O., Russell, W.J., Jenkins, S., Liu, D., Green, N., Llewelyn, K.,
Cole, P., Shek, V. and Krupenia, S.S. (2008) ‘Advanced auditory displays and head-mounted
displays: advantages and disadvantages for monitoring by the distracted anesthesiologist’,
Anesth. Analg., Vol. 106, pp.1787–1797.
Santamore, D.C. and Cleaver, T.G. (2004) ‘The sounds of saturation’, J. Clin. Mon. Comput.,
Vol. 18, pp.89–92.
Sarter, N.B. and Woods, D.D. (1997) ‘Team play with a powerful and independent
agent: operational experiences and automation surprises on the Airbus A-320’, Hum. Factors,
Vol. 39, pp.553–569.
Schoenfeld, V.S. and Scerbo, M.W. (1999) ‘The effects of search differences for the presence and
absence of features on vigilance performance and mental workload’, Automation Technology
and Human Performance: Current Research and Trends, pp.177–182.
Sittig, D.F., Wright, A., Osheroff, J.A., Middleton, B., Teich, J.M., Ash, J.S., Campbell, E.
and Bates, D.W. (2008) ‘Grand challenges in clinical decision support’, J. Biomed. Inform.,
Vol. 41, pp.387–392.
Struys, M., de Smet, T., Versichelen, L.F.M., van de Velde, S., van den Broecke, R. and
Mortier, E.P. (2001) ‘Comparison of closed-loop controlled administration of propofol
using Bispectral Index as the controlled variable versus ‘standard practice’ controlled
administration’, Anesthesiology, Vol. 95, pp.6–17.
Swain, A.D. and Guttmann, H.E. (1983) Handbook of Human-Reliability Analysis with Emphasis
on Nuclear Power Plant Applications, Final Report, NUREG/CR-1278, Sandia National Labs,
Albuquerque, NM, USA.
Tappan, J., Daniels, J., Slavin, B. and Lim, J. (2008) ‘Improved vigilance with context relevant
physiological monitoring’, Society for Technology in Anesthesia Annual Meeting, San Diego,
USA, p.S-19.
Tran, H., Froese, N., Dumont, G., Lim, J. and Ansermino, J.M. (2007) ‘Variation in blood pressure
as a guide to volume loading in children following cardiopulmonary bypass’, J. Clin. Monit.
Comput., Vol. 21, pp.1–6.
Winchell, R.J. and Hoyt, D.B. (1996) ‘Spectral analysis of heart rate variability in the ICU:
a measure of autonomic function’, J. Surg. Res., Vol. 63, pp.11–16.
Yang, P., Dumont, G. and Ansermino, J.M. (2006a) ‘Adaptive change detection in heart rate trend
monitoring in anesthetized children’, IEEE Trans. Biomed. Eng., Vol. 53, pp.2211–2219.
Yang, P., Dumont, G. and Ansermino, J.M. (2006b) ‘An adaptive Cusum test based on a hidden
semi-Markov model for change detection in non-invasive mean blood pressure trend’,
Conf. Proc. IEEE Eng. Med. Biol. Soc., Vol. 1, pp.3395–3398.
Yang, P., Dumont, G.A., Lim, J. and Ansermino, J.M. (2005) ‘Adaptive change point detection
for respiratory variables’, Conf. Proc. IEEE Eng. Med. Biol. Soc., Vol. 1, pp.780–783.
Young, D. and Griffiths, J. (2006) ‘Clinical trials of monitoring in anaesthesia, critical care and
acute ward care: a review’, Br. J. Anaesth., Vol. 97, pp.39–45.