ShareThe unique challenges in developing early
warning for children have led to a patient-specific early warning approach. The
methodology, which is based on the altered patterns of physiological
derangement, associated with compensation and decompensation in clinical
deterioration rather than population normal distributions, is suitable for
patients of all ages. A software platform that is used for Formula One telemetry,
which is able to analyse continuous data in real-time and produce predictive
models for the future, has been adapted for use in critically ill patients.
This enables real-time principal component analysis (PCA) and predictive
modelling, which are promising solutions for developmental physiological
changes and patient specific variations, whilst avoiding false alarms.
Introduction
Formula One motor racing is all about winning. Healthcare is all
about reducing distress and saving lives. Both are most successful when
underpinned by detailed knowledge of the context and behaviour involved as well
as the drive for safety and quality. In addition, they both require prediction
of likely outcomes, pre-planning to mitigate
anticipated complications and rapid, real-time decision-making to achieve the
best results.
Children provide unique challenges in early warning that have led to
developing patient specific monitoring. The wide range of “normal” physiology
which develops as children grow, and the complex “at risk” patients that have
their own specific range of normality, for example pulse oximetry of 75% in
cyanotic heart disease, has provided the drive to develop early warning that
learns the normal and abnormal physiological patterns for that patient. Once
successfully developed, this age-agnostic approach will almost certainly be
equally applicable to adult and geriatric patients. We describe an approach where
the patterns generated by individual susceptibility, acute physiological
response, compensation and decompensation thresholds create early warning.
Early Warning Systems in Adults and Children
Early warning scores are substantially better than relying on
chance identification of deterioration, achieving a 25% reduction in
life-threatening illness and death (Priestley
et al. 2004; Tibballs et al. 2009). However, there are two good reasons why we
need to explore beyond the current aggregate systems.
Firstly,
current early warning systems rely on categorising a patient as normal or
increasingly abnormal in comparison to a population normal distribution. Patients
have individual normal physiology and resilience and responses to illness are
based on their background health or illness, medication and age (Bion 2000;
Smith et al. 2008). We need to know what is normal for that specific patient. A
well 10-year-old, an immunocompromised 20-year-old with leukaemia and an
elderly 80-year-old patient with heart failure will each respond completely
differently to abdominal sepsis. Current early warning systems are not tailored
to this background-dependent resilience or susceptibility but focus only on the
acute, generic physiological changes. In addition, some patients deteriorate
and decompensate within the normal range for their age and for them. This is
because the effects of deterioration aren’t limited to changes in high and low
thresholds, but are related to the pattern of variation within physiological
parameters. The patient shows a clear change in pattern of variability prior to
cardiac arrest, but would not have triggered high or low alarm threshold until
the acute life-threatening event. Some of this altered pattern is measured in
heart rate variability: a marker that has benefit in identifying sepsis in
adults, neonates and foetal distress (Ahmad et al. 2009; Moorman et al. 2011;
Van Laar et al . 2008).
Secondly, current early warning systems rely on relatively infrequent
(every one to 12 hours) and varied vital sign measurements, from a choice of up
to 36 parameters. Furthermore, the warning can in some cases be associated with
missing measurements (Royal College of Physicians 2012; Duncan 2007; Royal College
of Nursing 2011). There is no information on the optimal frequency of observations
for a specific patient or population and this decision is often left to relatively
inexperienced, busy bedside nurses and healthcare assistants. The assumption is
that the bedside staff will recognise a deteriorating change in trends and
alter the observation frequency, but this opportunity is often missed. There is
also variation in which vital signs should be measured.
In
2008, a paediatric early warning system (PEWS) was introduced in all wards at
Birmingham Children’s Hospital (Duncan et al. 2006; Parshuram et al. 2011).
This is a paper-based aggregate score embedded as colour-coded, age-dependent
thresholds on the four standardised observation charts. It is associated with
simulation-based training on taking routine observations and appropriate
decision-making relating to deterioration. The type and frequency of observations
are guided by a comprehensive evidence- and expert-based observation, monitoring
and escalation policy. All life-threatening events are tracked, in keeping with
international recommendations (Devita et al. 2006), and they are classified
into timely and untimely for intensive care referral and admission, and whether
or not they are predictable and/or potentially preventable for acute life
threatening events. It is this detailed forensic review of all episodes of
critical deterioration that has provided the insight into how best to approach
early warning.
Since the introduction of early warning systems, in-hospital cardiac
arrests have reduced and more patients are receiving optimal pre-intensive
care. These are direct indicators of more timely treatment of acute illness.
But it could be better: measurements and observations could be more frequent,
processing the data could be automated, data entry mistakes could be avoided
and warnings or alarms could be tailored more specifically to individual patients.
Birmingham Children’s Hospital cares for children from birth to 16 years old, with weights ranging from 450g to 120kg. Our patients are frequently complex with cyanotic heart disease, chronic lung disease, neuro developmental disorders, multi- organ involvement and they epitomise individuality. Four age-appropriate observation charts are needed to accommodate ab/normal physiological parameters for the age-groups: birth to one year, one to five years, five to 12 years and older than 12 years. Infants, in particular, and older children can deteriorate very quickly; in between infrequent observations. These situations can erroneously be interpreted as unpredictable; however, parents are often adamant that a change had occurred in the child’s condition prior to an acute lifethreatening event that routine or even enhanced monitoring did not detect. These challenges have led to our exploration of Real-time continuous, Adaptive patientspecific, Predictive Indicators of Deterioration (RAPID).
How Does Formula One Help Solve These Problems?
Based on the problems identified so far, we determined that what is required is a system with the following requirements:
• Real-time analysis - to identify changes in patterns of physiological compensation and decompensation, and predict or form a model for the future;
• Adaptive - to have real-time analytical ability to learn normal for that specific patient;
• Continuous - until an optimal observation frequency can be determined; and
• Scalable and not reliant on expensive individual monitors - to measure as many at risk patients as possible.
A
solution is not yet available for medical monitoring, but is routinely used in
motor racing telemetry. You
need a fast car, great driver and good strategy to win races in the complex and
highly competitive world of Formula One. It is human endeavour at its most extreme,
characterised by relentless development with new innovations appearing
throughout the year in everchanging forms. Each must be anticipated, its
influence evaluated, and then put into action quickly.
The cars are changing continually to make them faster, stronger and
safer, and they go into intense competition every two weeks between March and
November. Hence, it is unsurprising that the world of Formula One is underpinned
by data. Quickly making sense of what you see and hear is often the difference
between winning and losing. In this respect, healthcare is little different. Recognising
problems quickly is the first step towards effective treatment, but each patient
is different and early signs can be subtle and complex. Nonetheless, they are
usually there to be found in the data. Recognising deterioration early provides
a real opportunity for reducing distress and saving lives.
McLaren Electronics Systems provides telemetry for all Formula One teams so they can measure, visualise and respond to changes during development of the cars as well as during the time-critical race situation. The Formula One realtime data system comprises SQL-Race, an application processing interface (API) that manages a large population of individual sources of time-series and associated data; vTAG server, a data logging and processing platform upon which real-time models run; and ATLAS, the data analysis and viewing software used by teams and engine makers throughout Formula One.
Each car is fitted with over a hundred sensors. Live health and
performance data is sent back via telemetry to the garage and over the Internet
to the team’s factory, often on the other side of the world. Over 750 million
numbers from each car are processed in real-time during a two-hour race. Over
the race weekend the data is used to make the cars better and faster. The data
tells the engineers how much life remains in the engine, how quickly the tyres
are degrading and how much fuel is being used (as well as how much is left in
the tank). The data tells them whether setup changes are effective or not, and
the system has the ability to run thousands of models simultaneously to predict
the consequences of different treatment strategies.
In healthcare, it is not feasible to have the equivalent of a Formula One team’s engineers focusing on just two patients, but it is possible to use the data platform to analyse patient-specific data in realtime, and to predict the future. If such analysis of changing physiology and variation could be visible to bedside or remote clinicians, then a much higher incidence of subtle signs of compensation and decompensation could trigger more sophisticated alerts, and we could see the predicted consequences of treatment and observation strategies.
Saving Young Lives
In 2011,
Birmingham Children’s Hospital and McLaren Electronic Systems installed a
real-time data system to gather and process live physiological data from all
beds in paediatric intensive care and from the trolley in one of the specialist
child transport ambulances. Through the “Young Lives” project, supported by the
Health Foundation SHINE programme and applied mathematics academics from Aston
University, we developed a system that would stream data from all of the
bedside monitors and quickly tease out patterns in the data, with a purpose of
alerting doctors and nurses to changing conditions.
The reason for starting in paediatric intensive care was twofold: it
is where the sickest children are treated with 1:1 nursing and it is where the
physiological data was already routinely collected (but previously overwritten
after 96 hours). In the first twelve months of running the system, we collected
physiological data from more than 1000 different patients. By streaming the
data into the Formula One data system, we have been able to provide a richer
display and manipulation of data and store it longer for the purposes of
clinical review and research.
The bedside instruments provide data from a range of sensors, but
initial focus was placed on pulse oximetry (SpO2) because it is readily measured and is rich in information about
respiration and cardiac activity. The real-time data platform can gather and
process data from a large population of individual patients. The data processing
can be applied to any of the physiological sensors and uses principal component
analysis (PCA) to extract the characteristic patterns from the data as it changes
with time. This technique is used for analysis and prediction in financial, environmental,
military and aeroplane engineering applications. A patient who is stable
exhibits patterns that change little over time. Plotting two principal
components against each other creates a model “distance”. Deterioration is
reflected in an increase in the model “distance”, a parameter which characterises
how well the principal components correlate with the evolving data.
The PCA approach not only teases out characteristic patterns, but
also provides the means to predict how the data should look in the future. It
does this by extrapolating and then reconstructing the physiological data for a
later time. Currently, we predict about two minutes ahead. The importance of
the prediction is that it enables quite tight margins to be applied in testing
for divergence from normality. This can lead to much earlier reliable detection
of change for individual patients.
We are testing the PCA model distance alongside an automated version of a modified paediatric early warning score (mPEWS). Early indications show that changing conditions are apparent in the PCA distance and scatter plots well before the mPEWS or raw data are seen to change. Further clinical interpretation is needed before changes may be characterised in terms of deterioration.
What Does the Future Hold?
Formula One has been using telemetry data to develop and race cars for over 25 years. The engineers and drivers believe and act upon the information they see, using it to understand and continually improve their cars and race-craft. Analytical techniques and fidelity checking between parameters has managed false alarms out of the system. Much of the work done in setting up the race car and developing a winning strategy takes place away from the track using live data sent across the world via standard fibre and wireless networks. It is no longer always necessary for the engineer and car to be in the same location in order to make a difference.
However, exploiting this approach in healthcare involves more than
simply transferring technology. The healthcare environment is less structured,
people can be more complicated and less predictable, clinical interventions can
be frequent and varied and the culture in secondary (and primary) health is not
always one that is immediately receptive to change. The next stages of
development at Birmingham Children’s Hospital will be to:
(1) Establish more rigorous clinical interpretation of the changing patterns;
(2)
Ensure that false detection of deterioration cannot happen;
(3)
Move the system beyond the walls of intensive care and into the high dependency
and general wards through-out the hospital; and
(4)
Create new patient pathways and resourcing models that make use of the better
clinical cues.
A lot has been achieved, but there is much more to do (Nangalia et al. 2010; Bion 2008). Embedding knowledge into the system of what constitutes normality, how characteristic changes in patterns relate to treatment and outcomes, and how alarm thresholds could and should be set, will all come with detailed clinical scrutiny of the data. Properly engineered, our approach will present physiological data clearly, immediately and in context, so that every patient, regardless of age, might be seen by the right people, in the right place and at the right time. Ideally, the changing conditions of the population of individual patients would help inform the most appropriate allocation of nurses throughout the hospital and direct doctors and other clinicians to the sickest patients. There is no reason, however, why an approach like RAPID should be confined to the hospital. Once developed, the applications that detect deterioration could operate remotely or be embedded in local devices, such as smart phones or tablets. Patients with acute and chronic conditions could be monitored at home with the reassurance that expert help could be informed quickly should a condition suddenly worsen.
Acknowledgements
The real-time Principal Component Analysis application that is used to extract and display changing patterns was developed by Dr. Rajeswari Matam of Birmingham Children’s Hospital. She also liaised directly with the software and support engineers from McLaren Electronic Systems throughout the development and commissioning of the system. The development of the analysis approach, and ensuring that it was both practical and relevant to this new application, was supported strongly by Prof. David Lowe of Aston University
