=Paper=
{{Paper
|id=None
|storemode=property
|title=A process mining driven framework for clinical guideline improvement in critical care
|pdfUrl=https://ceur-ws.org/Vol-765/paper5.pdf
|volume=Vol-765
}}
==A process mining driven framework for clinical guideline improvement in critical care==
https://ceur-ws.org/Vol-765/paper5.pdf
A Process Mining Driven Framework for Clinical
Guideline Improvement in Critical Care
Carolyn McGregor1, Christina Catley1 and Andrew James2,3
1
University of Ontario Institute of Technology, Oshawa, Canada
2
The Hospital for Sick Children, Toronto, Canada
3
Department of Paediatrics, University of Toronto, Toronto, Canada
{c.mgregor, ccatley}@ieee.org, andrew.james@sickkids.ca
Abstract. This paper presents a framework for process mining in critical care.
The framework uses the CRISP-DM model, extended to incorporate temporal
and multidimensional aspects (CRISP-TDMn), combined with the Patient
Journey Modeling Architecture (PaJMa), to provide a structured approach to
knowledge discovery of new condition onset pathophysiologies in physiological
data streams. The approach is based on temporal abstraction and mining of
physiological data streams to develop process flow mappings that can be used
to update patient journeys; instantiated in critical care within clinical practice
guidelines. We demonstrate the framework within the neonatal intensive care
setting, where we are performing clinical research in relation to
pathophysiology within physiological streams of patients diagnosed with late
onset neonatal sepsis. We present an instantiation of the framework for late
onset neonatal sepsis, using CRISP-TDMn for the process mining model and
PaJMa for the knowledge representation.
Keywords: Critical care, physiological data streams, process mining,
knowledge discovery, clinical guidelines, patient journey modeling
1 Introduction
The neonatal intensive care unit (NICU) is a complex critical care environment that
must support collaborative decision making amongst different care provider roles
based on information received from high volumes of heterogeneous real-time
synchronous medical device data, as well as asynchronous data from hospital
information systems. Information must be organized to enable efficient and effective
clinical decision making on the many complicated diagnoses present in today’s
NICUs. This includes supporting both known clinical guidelines and the ability to
modify and extend current clinical guidelines based on new research, for example the
retrospective analysis of physiological data streams to establish new and earlier
pathophysiological behaviours evident in seemingly unrelated physiological data
streams is a growing research area.
This paper presents a framework for process mining in critical care. It provides a
structured approach to the knowledge discovery of new condition onset
pathophysiologies of physiological data streams and constructs process flow
mappings that can be used to update patient journeys as instantiated within clinical
practice guidelines. We demonstrate the flexibility and multidimensionality of this
framework within the NICU setting where there are multiple streams of physiological
data being captured from multiple patients, being watched for the onset of multiple
potential conditions that can occur concurrently with patients located in multiple
NICU locations. We are currently performing clinical research in relation to
pathophysiology within physiological streams of patients who were diagnosed with
late onset neonatal sepsis (LONS). We present an instantiation of the framework for
LONS, showing the interactions between the physiological data streams and the high
level clinical decision. While this case study is presented within a NICU context, its
flexible and adaptive nature makes it equally applicable to any critical care
environment acquiring data streams and performing similar knowledge discovery.
The remainder of the paper is structured as follows: Section 2 presents background
information, including an overview of patient journey modeling and process mining.
Section 3 introduces our framework and Section 4 demonstrates the framework within
the NICU context. The paper concludes in section 5 where limitations of this research
and future research directions are discussed.
2 Background
2.1 Process Mining
Process mining is an emerging research area motivated by the goal of extracting
knowledge from event logs recorded by an information system; an event log contains
data on the order in which the events take place [1]. While event-based data has
potential to provide new knowledge ‘process mining provides a new means to
improve processes in a variety of application domains’ [2], in most cases this
potential is unrealized. By providing techniques for discovering process, control, data,
organizational, and social structures from event logs, process mining attempts to
overcome this limitation [3]; for a current process mining review please see [2].
Event log data can originate from diverse systems, including hospital information
systems (HISs), and support diverse applications, including health care [4]. As an
example, the NICU’s HIS records data on many clinical processes, including
admission, investigations, diagnoses, treatment and follow-up. The current research
on process mining in health care is in its early stages. Mans et al. provide several case
studies proving the applicability of process mining to health care. In one, they
demonstrate that process mining can be used for obtaining insights related to
careflows, also referred to as patient journeys, for gynecological oncology patients
where each event refers to a service delivered to a patient [4]. In that work, process
mining was used to derive understandable models for large groups of patients. They
identified that process mining is limited in its ability to work with unstructured
processes often found in hospital environments. In a separate study, Mans et al. study
two datasets for stroke patients, one of which encompasses the clinical course from
admission to discharge, and the second focuses on pre-hospital behaviour [5]. They
found that process mining can be used both to construct models for a whole data set,
or for only aspects that are of interest; for example, in the NICU context, only the
investigation, diagnosis, or treatment process pathways. By focusing on specific
process paths and using discovery process mining, they were able to identify
differences in treatment strategies between different hospitals.
We see strong linkages between the real-time physiological data streams
continually recorded in critical care settings and the event logs used in process
mining; in critical care the processes of interest are the implementation of clinical
guidelines, for which we seek to support process improvement and obtain new
insights. However, process mining in health care has not yet been considered within
the context of critical care and towards the improvement of care pathways and clinical
guidelines through the inclusion of new pathophysiological behaviour analysis from
complex and interrelated physiological data streams analysis.
2.2 Patient Journey Modeling
The Patient Journey Modelling Architecture (PaJMa) was designed specifically for
health care, providing a visual representation of the processes, information and
technology involved in a patient journey for different clinical scenarios. Recent work
has implemented PaJMa in diverse hospital settings, proving that this technique is a
highly applicable method of health care modeling for use within critical care [6].
PaJMa makes use of horizontal information layers containing independent
information on different aspects of the patient’s journey; layers are read vertically
from left to right, with each process step moving across the page [7].
Working from the top-down, each layer provides important information related to
the patient journey, specifically: 1) the patient’s point of entry into the journey and
subsequent interactions with the roles and processes contained in the journey; 2) the
roles involved; 3) processes and possible decisions that comprise the action items of
the journey; 4) information required or obtained by each process; 5) technology used;
6) underlying infrastructure support; 7) patient needs, policies, guidelines, and
strategic objectives that are relevant to the various processes; and 8) optional metrics.
An example PaJMa model instantiated for the diagnosis of LONS is provided in
Section 4’s case study. Experience indicates that PaJMa is effective at modeling
different health care roles and their associated processes, as well as representing the
use of technology to support the processes. However, to date PaJMa has not been
used to support the event based activities that constitute clinical guideline models,
where responses to events can trigger ad hoc pathways of patient journeys. In this
work, we extend the current PaJMa modeling technique to show how it can support
process mining of temporally abstracted physiological data streams.
2.3 CRISP-TDM
The CRoss Industry Standard Process for Data Mining (CRISP-DM) [8] was
developed in 1996, with the goal of being industry, tool and application-neutral.
Repeated references to the methodology by analysts have established it as the de facto
standard for data mining and KDD [9]. The model breaks the life cycle of a data
mining project into six phases: business understanding, data understanding, data
preparation, modeling, evaluation and deployment.
Analysis of the current CRISP-DM 1.0 Process and User Guide [8] identified that
phases 1 (business understanding), 2 (data understanding), 4 (data modeling), 5
(evaluation) and 6 (deployment) of the current CRISP 1.0 model were all limited in
their ability to describe: 1) clinically relevant and population-based information in the
business understanding phase; 2) temporal aspects of the multidimensional data and
the clinical study in the data understanding phase; 3) temporal abstraction of relevant
details and knowledge management in the data modeling phase; 4) system integration
in the data modeling phase, where CRISP 1.0 concentrates on applying several data
mining techniques to arrive at one which offers the best results, rather than providing
support for integrating techniques, such as data mining and temporal abstraction; 5)
assessment of process mining results based on temporally abstracted data in the
evaluation phase; and 6) storage issues, knowledge sharing and representation issues
in the deployment phase, including mechanisms used for knowledge representation,
such as adherence to health care standards [10].
Fig. 1 represents the enriched CRISP-DM model for reporting results of clinical
investigations, for a detailed discussion please see [10]; the extended model is labeled
CRISP-TDMn to highlight its support of temporal (T) and multidimensional (n)
clinical data. The first branch following a phase definition represents a task output for
a CRISP-TDMn phase; all lower branches in the hierarchy represent attributes of a
task output for a CRISP-TDMn phase. Tasks and attributes from the original CRISP
1.0 model are not presented in these figures, only those extensions relevant to the
CRISP-TDMn methodology are shown; extended tasks are marked with * and
extended attributes with **. In this work, CRISP-TDMn is used in conjunction with
PaJMa to develop a framework for process mining of physiological data streams.
3 Framework for Process Mining and Process Representation in
Critical Care (CRISP-TDMnPaJMa)
In this section we introduce our framework to enable process mining in critical care.
The framework follows the principles of the CRISP-TDMn approach to knowledge
discovery and utilizes the Service based Multidimensional Temporal Data Mining
(STDMn0) framework [11] to perform the process mining that enables knowledge
discovery of new condition onset pathophysiologies from physiological data streams.
It enables the construction of process flow/patient journey pathways within an
extended PaJMa model based on the knowledge gained that can be used to update
patient journeys as instantiated within clinical practice guidelines. The first subsection
describes the components for knowledge discovery and the second details the
representation of the new knowledge within the PaJMa model.
Fig. 1. Extended CRISP-TDMn model [10]
3.1 Process Mining in Critical Care – Knowledge Discovery
This framework utilizes the extensions proposed to CRISP-DM for a standardised
approach to temporal data mining. Following the principles of the CRISP-TDMn, for
process mining to occur a goal must first be determined as part of the Business
Understanding Phase. The goal will be to determine whether new pathophysiological
behaviours can be found in retrospective analysis of physiological data streams prior
to the diagnosis of a selected clinical condition or clinical event of interest, where in
each instance the condition or event will be quantified. During this phase ethics
approval would normally be sought for the research.
The second phase of CRISP-TDMn is to perform Data Understanding. Through the
use of STDMn0 the data will have already been gathered by the STDMn0 processing
agents and contained within the data management layer of the architecture. In
addition, several temporal abstractions against that data will have already been run.
The rules for which are contained in the temporal rules table, and the derived patient
data is contained within the temporal abstraction tables. New temporal abstractions
related to the study can also be performed on the data at anytime through the study
process. As a result, CRISP-TDMn Data Preparation is also completed.
The next step is to establish a study data subset for the process mining. Within
STDMn0 this is performed by defining a new study. When a new study is defined, the
clinical condition event of interest is defined and, associated with that, the rules for
the selection of patients to be included in that retrospective study. As part of the data
subset establishment and in support of patient anonymity, the data is relatively aligned
to represent a relative negative distance back in time from the point of the clinical
condition or event of interest; this is performed by the Relative Agent. The rules for
the study are stored within the study table and the data within the relative alignment
table, encoded with the study id.
The modeling component of CRISP-TDMn is performed within STDMn0 by the
Processing Agents. The evaluation component of CRISP-TDMn is supported through
the provision of support for null hypothesis testing. From this process mining activity,
temporal abstractions within one or many isolated or interrelated physiological
streams will be found to have strong correlations as pathophysiological indictors for
the given clinical condition. A major strength of our approach for knowledge
discovery from process mining is its multidimensional nature. We can support the
instantiation of several studies around several clinical conditions, exploring different
physiological streams, for data located in different locations, and of interest to
different care provider roles. This multidimensional approach is represented in Fig. 2,
illustrating the view for a neonatologist studying apnoea as a clinical manifestation of
LONS. The next phase is to propose a mechanism for translation and representation
of that new knowledge within a process model as detailed in the next subsection.
Fig. 2. Multi-dimensional nature of process mining environment – neonatologist’s view
3.2 Process Mining in Critical Care – Knowledge Translation and
Representation
For the representation of the new temporal abstraction knowledge within the process
model we have chosen to use the PaJMa model. The information for the source
physiological streams and the location of that data as it will be used in real-time is
easily represented within the PaJMa model within the Technology Used and
Information layers. In addition, the derived temporal abstractions that will need to be
performed in real-time can also be represented within the Information layer of the
PaJMa model. The clinical events that these abstractions work to support are
represented within the Process layer of the model and the staff roles who will
participate in these processes are included in the Roles layer.
Fig. 3. Relationship between PaJMa layers and process mining in critical care components
Elements in the Information and Technology Used layer are colour coded to
provide linkage to the roles. The relationship between PaJMa model layers and
process mining in critical care is shown in Fig. 3. In previous work, we have
demonstrated how the new knowledge in the form of temporal abstraction rules can
be translated into a workflow engine designed to support the ingestion and analysis or
real-time streaming sensor data and show its application to critical care. This
instantiation is through the Artemis platform [12-14].
4 Demonstration of the Framework within a NICU Case Study
NICUs provide critical care to newborn children, often prematurely born infants that
are especially susceptible to infection [15]. In this case study we focus on process
mining centred around two clinically significant and related neonatal outcomes:
LONS and central apnoea. We instantiate the layers of the CRISP-TDMn technique
and present extended PaJMA models, illustrating the framework’s approach to
process mining and knowledge translation for new onset diagnosis of LONS.
Neonatal sepsis is a life threatening condition, with early diagnosis being
significant; infants are often diagnosed only when seriously ill which decreases the
probability for prompt, complete recovery with antibiotic therapy [16]. LONS is
typically defined as sepsis acquired as early as four days after birth and as late as 28
days after birth [15]; for the purpose of this work we use the definition that LONS
refers to sepsis acquired on or after the fifth day of life. LONS occurs in
approximately 10% of all neonates and in more than 25% of very low birth weight
infants who are hospitalized in NICUs [17]. Numerous studies have linked sepsis to
reduced heart rate variability (HRV) [16], [18-21]. Previous work has discussed our
approach to analyzing HRV for use as an early indicator of neonatal sepsis [22].
Central apnoea is another clinically significant condition, episodes are first
recognized as a lapse in respiration, identified either by a decrease in respiration rate
below a predefined threshold, or by a drop in frequency of the impedance respiratory
wave. Many studies have shown that apnoea duration is positively correlated with
both a decrease in blood oxygen saturation and heart rate [23-26]; this pathological
change is used as a marker to indicate the possible presence of apnoea in the absence
of a confirmed clinical observation of apnoea. Previous work has presented our
implementation of clinical guidelines for the real-time detection of central apnoea
using Artemis [27]. Episodes of apnoea can be a pathophysiological onset behaviour
for several of conditions, one of which is LONS.
Given the body of research showing the worth of knowledge discovery in
physiological streams, there is a need to enable a multidimensional environment that
enables knowledge discovery of new pathophysiological behaviours in physiological
streams prior to the onset of a clinical event and a multidimensional mechanism to
watch for these in real-time.
4.1 CRISP-TDMn Business Understanding
The Clinical Application Domain is that of neonatal intensive care involving two
collaborating NICUs, The Hospital for Sick Children, Toronto (SickKids) and
Women and Infants Hospital, Providence, Rhode Island (WIHRI).
The Clinical Objective is the proposition of new physiological stream
pathophysiological behaviours evident prior to the diagnosis of LONS with a higher
sensitivity and specificity as compared to current clinical guideline
pathophysiological measures. A secondary objective is that these new behaviours are
evident earlier than the current clinical guideline pathophysiological measures.
Research ethics board approval was received from the hospitals of the two
collaborating NICUs.
The Data Mining TA Goals and hence the process mining goals are then to perform
population-based retrospective supervised data mining of temporally abstracted
behaviours of physiological streams of patients who had developed LONS. Within
this paper, for the purposes of this demonstration we will limit our temporal data
mining to the temporal abstractions of reduced HRV and mild, moderate and severe
instances of central apnoea. Temporal abstraction involves transforming time stamped
data into an interval-based representation of the data by extracting the most relevant
features [28], such as identifying states, trends and temporal relationships. Then
sections of data that hold true for certain criteria can be summarized with a start time
and end time.
4.2 CRISP-TDMn Data Understanding and Data Preparation
Data for this research was collected utilizing the Artemis platform, and the study into
LONS represents the primary use of this data. Data collection from SickKids
commenced in August 2009 and in April 2010 from WIHRI [12], [14]. Heart rate
(HR), respiration rate (RR), and blood oxygen saturation (SpO2) are collected from
patients at SickKids at the rate of one reading a second (1Hz) for the duration of their
stay in the NICU from the bedside Philips MP70 devices. In addition, an impendence
respiratory wave (IRW) is also collected measuring the chest wall impedance to
determine the presence of breathing, with a frequency of 125 Hz. Currently Artemis
has been enabled at 8 bed spaces while the platform has undergone testing, but a
hospital infrastructure upgrade will now enable us to utilize network ports at each bed
space to collect from all 38 beds.
The Space Labs bedside devices are used at WIHRI and a cloud computing
implementation of Artemis [14] enables the transmission of a number of physiological
streams including heart rate, respiration rate and blood oxygen saturation readings
every minute to the Artemis cloud implementation located at the University of
Ontario Institute of Technology (UOIT). Currently Artemis cloud has been collecting
data from between 10-15 patients daily during performance testing, but the
architecture could collect from as many bed spaces as the clinical research requires.
4.3 CRISP-TDMn Data Modeling
This research performs quantitative analysis of the temporal abstractions performed
on physiological streams as detailed in Table 1. The details of the temporal
abstraction calculations are outside the scope of this paper. Our approach for reduced
HRV temporal abstractions is described in [22] and the abstractions for neonatal
apnoea are detailed in [27]. The flexibility of this approach enables various data
mining algorithms, such as machine learning, that accommodate temporal data to
propose and analyse correlations.
Table 1. Temporal abstractions for LONS Process Mining
Temporal Abstraction Physiological Stream
Reduced HRV HR
Mild Apnoea IRW
Moderate Apnoea IRW, SpO2
Severe Apnoea IRW, SpO2, HR
4.3 CRISP-TDMn Evaluation
For the purposes of this demonstration for process mining we will assume that the
results of the data mining have demonstrated superior sensitivity and specificity when
mining the temporal abstractions relating to reduced HRV and those that classify
apnoea. The next step is to model results within the PaJMa model.
4.4 PaJMa Knowledge Translation and Representation
The PaJMa model provides the representation of the knowledge within the process
model to support the enactment of this new knowledge in real-time. Each form of
temporal abstraction that represents a clinical event found to be relevant in the process
mining phase is represented within the PaJMa model as a vertical slice. For each
vertical slice, the clinical event is represented in the Processes layer. The medical
devices creating the physiological data are shown in the Technology Used layer. The
temporal abstractions that relate to that clinical event are represented in the
Information Creation/Movement layer. The roles interested in this event are shown in
the Roles layer and the elements of the Information Creation/Movement and
Technology Used layers are colour coded to link them with the roles. The
Infrastructure Used layer demonstrates where equipment is connected to the hospital
network and where that could be connected to a secure gateway for connection
through the internet. The Technology Used layer demonstrates that Artemis is used for
the temporal abstractions and role notifications to different user interface devices.
Fig. 4. PaJMa process model for LONS
5 Conclusions and Future Work
This research has presented a generalized framework to support process mining in
critical care that enables knowledge discovery of new condition onset
pathophysiologies using temporal data mining of physiological data streams and
constructs process flow mappings that can be used to update patient journeys as
instantiated within clinical practice guidelines. The research was demonstrated within
the context of neonatal intensive care and specifically as it relates to LONS and the
potential of reduced heart rate variability and apnoea to be pathophysiologies for that
clinical condition. In this way, we have demonstrated the ability for this framework to
be used by clinical researchers as a platform to perform knowledge discovery, to store
that new knowledge, and to translate that knowledge to updated clinical guidelines.
While we have proven that PaJMa can easily represent the distinct activities and
provide a rich set of information needed for the instantiation within real-time clinical
care, the ad hoc nature of the event sequence was a challenge to represent, as it is with
any business process modeling technique. We have commenced work on extensions
to the PaJMa model to better support the event based nature of the ICU setting. The
Guidelines Element Model (GEM) [29] has been proposed as a mechanism for
knowledge representation of clinical guidelines. In future work we will show how the
process mining phase can generate GEM representations which can then be
represented within an extended PaJMa model.
To assess quality improvement in health care, the impact of changes to the clinical
guidelines need to be assessed through either clinical trial or other metrics, such as in
the case of LONS, a reduction in average length of stay. Formal models for evaluation
need to be proposed for the analysis of these process mining driven changes to
clinical guidelines.
Acknowledgments. This research is funded by the Canada Research Chairs program
and a Canadian Institutes for Health Research (CIHR) Operating Grant. We would
also like to thank Dr James Padbury and his research team from the NICU, Women
and Infants Hospital, Providence Rhode Island.
References
1. van der Aalst, W.M.P., Weijters, A.J.M.M.: Process Mining: a Research Agenda. Comput.
Ind. 53, 231-244 (2004)
2. Van der Aalst, W.M.P.: Process Mining: Discovery, Conformance and Enhancement of
Business Processes. Springer (2011)
3. Queensland University of Technology & Eindhoven University of Technology: Process
Mining. http://is.tm.tue.nl/staff/wvdaalst/BPMcenter/process%20mining.htm
4. Mans, R.S., Schonenberg, M.: Application of Process Mining in Healthcare–A Case Study
in a Dutch Hospital. In: International Joint Conference on Biomedical Engineering Systems
and Technologies, pp 425-438. Springer-Verlag, Berlin Heidelberg (2008)
5. Mans, R., Schonenberg, H., Leonardi, G., Panzarasa, S., Cavallini, A., Quaglini, S., van der
Aalst, W. Process Mining Techniques: an Application to Stroke Care. In: Proceedings of
Medical Informatics Europe, pp 573-578. IOS Press (2008)
6. MacDougall, C., McGregor, C., Percival, J. The Fusion of Clinical Guidelines with
Technology: Trends and Challenges. eJHI. 5, 2 (2010)
7. Percival, J., Catley, C., McGregor, C., James A. Modeling the Impact of Technology on
Information Flows in Neonatal Intensive Care Units. Stud. Comp. Intell. 189, 147-169
(2009)
8. Chapman, P., Clinton, J., Kerber, R., Khabaza, T., Reinartz, T., Shearer, C. CRISP-DM 1.0,
http://www.crisp-dm.org/download.htm
9. Shearer, C. The CRISP-DM Model: The New Blueprint for Data Mining. J. Data
Warehousing. 5, 13-22 (2000)
10. Catley, C., Smith, K., McGregor, C., Tracy, M. Extending CRISP-DM to Incorporate
Temporal Data Mining of Multidimensional Data Streams: a Neonatal Intensive Care Unit
Case study. In: Computer-Based Medical Systems, pp. 1-5 (2009)
11. McGregor, C. System, Method and Computer Program for Multi-Dimensional Temporal
Data Mining. Patent # 089705-0009; Canada, Gatineau Quebec (2010)
12. Blount, M., Ebling, M.R., Eklund, J.M., James, A.J., McGregor, C., Percival, et al. Real-
Time Analysis for Intensive Care: Development and Deployment of the Artemis Analytic
System. IEEE Eng Med Biol Mag. 29, 2, 110-118 (2010)
13. McGregor, C., Sow, D., James, A., Blount, M., Ebling, M., Eklund, J., et al. Collaborative
research on an intensive care decision support system utilizing physiological data streams.
In: Proc AMIA Annu. Symp. pp 1124-6 (2009)
14. McGregor, C., Catley, C., James, A., Padbury, J. Next Generation Neonatal Health
Informatics with Artemis. Accepted for presentation at MIE 2011, Informatics Europe,
Oslo, Norway.
15. Ganatra, H. a., Stoll, B.J., Zaidi, A.K.M. International Perspective on Early-Onset Neonatal
Sepsis. Clin. Perinatol. 37, 2, 501-523 (2010)
16. Griffin, M.P., OʼShea, T.M., Bissonette, E.A., Harrell, F.E., Lake, D.E., Moorman, J.R.
Abnormal Heart Rate Characteristics Preceding Neonatal Sepsis and Sepsis-like Illness.
Pediatr Res. 53, 6, 920-926 (2003).
17. Beck-Sague, C.M., Azimi, P., Fonseca, S.N., Baltimore, R.S., Powell, D.A., Bland, L.A., et
al. Bloodstream Infections in Neonatal Intensive Care Unit Patients: Results of a
Multicenter Study. Pediatr Infect Dis J. 13, 12, 1110-1116 (1994)
18. Flower, A.A., Moorman, J.R., Lake, D.E., Delos, J.B. Periodic Heart Rate Decelerations in
Premature Infants. Exp. Biol. Med. 235, 4, 531-538 (2010)
19. Griffin, M.P., Lake, D.E., O’Shea, T.M., Moorman, J.R. Heart Rate Characteristics and
Clinical Signs in Neonatal Sepsis. Pediatr Res. 61, 2, 222-227 (2007).
20. Griffin, M.P., Lake, D.E., Moorman, J.R. Heart Rate Characteristics and Laboratory Tests
in Neonatal Sepsis. Pediatrics. 115, 4, 937-41 (2005)
21. Griffin, M.P., Moorman, R. Using Novel Heart Rate Analysis. Pediatrics, 107, 97-104
(2001)
22. McGregor, C., Catley, C., James, A. Simplifying Complexity: Variability Analysis with
Analytics Applied to Physiological Data Streams from the Neonatal Intensive Care Unit.
Submitted for presentation at Annual Fall Symposium, AMIA (2011), Washington DC.
23. Finer, N.N., Barrington, K.J., Hayes, B.J., Hugh, A. Obstructive, Mixed, and Central Apnea
in the Neonate: Physiologic Correlates. J. Pediatr. 121, 943-950 (1992)
24. Finer, N.N., Higgins, R., Kattwinkel, J., Martin, R.J. Summary Proceedings from the
Apnea-of-Prematurity Group. Pediatrics. 117, S47-51 (2006)
25. Muttitt, S.C., Finer, N.N., Tierney, A.J., Rossmann, J. Neonatal Apnea: Diagnosis by Nurse
Versus Computer. Pediatrics. 85, 713-720 (1988)
26. Upton, C.J., Milner, A.D., Stokes, G.M. Apnoea, Bradycardia, and Oxygen Saturation in
Preterm Infants. Arch Dis Child. 66, 381-385 (1991)
27. Catley, C., McGregor, C., Smith, K., James, A., Eklund, J.M. A Framework for Multi-
Dimensional Real-Time Data Analysis: a Case Study for the Detection of Apnoea of
Prematurity. J. Comput. Mod. Alg. Med. (in press 2011)
28. Post, A.R., Harrison, J.H. Temporal Data Mining. Clin. Lab. Med. 28, 83-100 (2008)
29. Guidelines Element Model, http://gem.med.yale.edu/default.htm