ENABLING UBIQUITOUS DATA MINING IN INTENSIVE CARE
Features Selection and Data Pre-processing
Manuel Santos and Filipe Portela
Departamento de Sistemas de Informação, Universidade do Minho, Guimarães, Portugal
Keywords: Ubiquitous Data Mining, Real-time Intelligent Decision Support Systems, Organ failure prediction, Clinical
Data Mining, Intensive Care Environment.
Abstract: Ubiquitous Data Mining and Intelligent Decision Support Systems are gaining interest by both computer
science researchers and intensive care doctors. Previous work contributed with Data Mining models to
predict organ failure and outcome of patients in order to support and guide the clinical decision based on the
notion of critical events and the data collected from monitors in real-time. This paper addresses the study of
the impact of the Modified Early Warning Score, a simple physiological score that may allow improvements
in the quality and safety of management provided to surgical ward patients, in the prediction sensibility. The
feature selection and data pre-processing are also detailed. Results show that for some variables associated
to this score the impact is minimal.
1 INTRODUCTION
The decision making process is a key factor in critical
environments such as intensive care, either in
prognosis or diagnosis of the patients’ condition, for
their lives may be at risk. According to some studies,
medical error may be the eighth cause of death in
industrialized countries (Kohn, Corrigan, &
Donaldson, 2000). The development of Intelligent
Decision Support Systems (IDSS) by means of Data
Mining (DM) prediction models for mortality and
organ failure may contribute to the reduction of
medical error.
Presently, there is an IDSS for organ failure
prediction by means of DM techniques. INTCare is
being tested in the Intensive Care Unit (ICU) of
Hospital Geral de Santo António (HGSA) (Santos et
al., 2009; Vilas-Boas, Santos, Portela, Silva, & Rua,
2010; Villas Boas et al., 2010).
The high volume of clinical information that
doctors have to deal with every day in ICU’s explains
the need to rely on Decision Support Systems (DSS).
ICU’s are a complex medical environment where
patients’ condition is critical and often their lives
depend on the care provided in those units.
Furthermore, there are so many data relating to a
patient, that one person can’t effectively process it
(Donchin & Seagull, 2002). Safer, less expensive, and
higher-quality health care can be achieved using
clinical DSS (Menachemi & Brooks, 2006).
Computer-based decision support tools are supposed
to help practitioners to avoid errors, ensure quality
and improve efficiency in healthcare. Health care
providers can manually enter patient characteristics
into the computer system, but ideally, information
retrieval, registration and display should be as
automated as possible.
Hence the interest by intensive care clinicians and
researchers in the development of prediction models
based on DM to support clinical decision. The
emergence of network-based computing
environments established a new dimension to this
area, regarding distributed sources of data and
computing. The growing use portable devices and
communication networks is making ubiquitous access
to large quantity of distributed data a reality and is
changing our relationship with a real-time and
distributed environment (Kargupta, 2001).
This paper presents the experiments regarding the
INTCare system, an IDSS for ICU in real-time, and
its pervasive and ubiquitous features. It focuses on the
features selection and data pre-processing for mining
intensive care data.
2 BACKGROUND AND RELATED
WORK
Pervasive computing has been pointed as a new
261
Santos M. and Portela F..
ENABLING UBIQUITOUS DATA MINING IN INTENSIVE CARE - Features Selection and Data Pre-processing.
DOI: 10.5220/0003507302610266
In Proceedings of the 13th International Conference on Enterprise Information Systems (ICEIS-2011), pages 261-266
ISBN: 978-989-8425-53-9
Copyright
c
2011 SCITEPRESS (Science and Technology Publications, Lda.)
paradigm for the 21st century. Significant hardware
developments, as well as advances in location
sensors, wireless communications, and global
networking, have advanced towards technical and
economic viability (Saha & Mukherjee, 2003).
Ubiquitous computing, pervasive computing and
ambient intelligence are concepts evolving in a
growing number of applications in health care and are
increasingly influencing it (Orwat, Graefe, &
Faulwasser, 2008). However, the majority of these
systems are in their prototype stages and there is a
need for advanced research on their deployment,
mainly clinical studies, economic and social analyses
and user studies. Ubiquitous computing can transform
some key features of user interaction: their location,
the scope of the service, and its duration and
frequency (Fano & Gershman, 2002).
3 PERVASIVENESS IN
INTENSIVE CARE
3.1 Requirements of HealthCare
Environments and Computer
Applications
To our knowledge, there is no IDSS that uses DM
models and addresses the important aspects of critical
environments, mainly the need for acting in real-time
in a fast, reliable, secure and ubiquitous way.
To achieve that, it is essential to make clinical
information available by ubiquitous devices
(Varshney, 2007), making possible to avoid medical
errors that occur due to lack of the correct
information when and where it is required (Kohn, et
al., 2000). The lack of correct and complete
information may lead to decision errors in 50%
(Bergs, Rutten, Tadros, Krijnen, & Schipper, 2005).
Ubiquity in electronic medical records that contain
detailed data about patients leverage their analysis by
whom is authorized, anytime and anywhere
(Varshney, 2009), allowing physicians and nurses to
act in real-time and as fast as possible.
Pervasive healthcare addresses this challenge and,
according to Varshney (Varshney, 2007), it may be
defined as “healthcare to anyone, anytime, and
anywhere by removing locational, time and other
restraints”. A pervasive IDSS may bridge most of
this gaps and help to coordinate the various activities
underlying patients care and survival (Scicluna,
Murray, Xiao, & Mackenzie, 2008), relying on
prediction and decision models that may be accessed
anywhere. Based on the work done during the
research for the INTCare project, and after a thorough
analysis of the problem, some features were identified
as being essential to the development of an IDSS.
Within this context it was determined to be essential
for an IDSS that uses DM and acts in real-time the
following features:
a) Online Learning: The system should have the
ability to operate in online mode, i.e. the DM
models must be induced using online data, in
contrast to the existing approaches that
emphasize offline learning mode.
b) Real-Time: The system must be able to act in
real-time. The process of acquisition and storage
of data should occur immediately after the
events take place;
c) Adaptability: The system must have the ability
to automatically optimize the models with the
latest data;
d) Optimization: Consists in optimizing the values
obtained by the prevision models generated,
analyzing all the results and selecting the best.
e) Data Mining Models: The success of this IDSS
depends on the accuracy of DM models, i.e. the
degree of reliability of these models should be
high;
f) Decision Models: Achieving the best results is
highly dependent on the decision models
created. These models are based on various
factors such as the decision and differentiation
that are used in predicting models, so that they
can help doctors choose the best decision during
the process of decision making;
g) Intelligent Agents: This type of agents allows
the system to work through autonomous actions
that perform some essential tasks. These tasks
support some system modules: Data
Acquisition, Management of Knowledge,
Inference and interface;
h) Pervasive / Ubiquitous: The technology of
ubiquitous computing aims to improve the living
and working environments and to build
advanced devices, infrastructure and network
operating systems.
i) Safety: This type of system that operates in
critical environments must guarantee safety for
both patients and users. By using ubiquity, some
protocols need to be followed like access,
authentication, communication (Langheinrich,
2001; Nigel, 2002; Piramuthu, 2003)
4 UBIQUITOUS DATA MINING IN
INTENSIVE CARE
Advanced analysis of distributed data for extracting
ICEIS 2011 - 13th International Conference on Enterprise Information Systems
262
useful knowledge is the next expected step in the
increasingly connected world of ubiquitous
computing (Neaga & Harding, 2005). In this context,
UDM is attracting a great amount of attention. Much
of the DM tasks which are being done currently focus
on databases or data warehouses whose data is
physically located in one place. However, the
scenario arises where information may be located in
different physical locations. Therefore, the goal is to
effectively mine distributed data streams which are
located in heterogeneous sites (Hsu, 2002). The
challenge and the novelty of this approach is to bring
ubiquity to DM, which is the process of selecting,
exploring and modeling large amounts of data in
order to discover unknown patterns or relationships
that provide a clear and useful result to the data
analyst (Giudici, 2003). The abundance of mobile
devices, such as PDA’s and cellular phones, coupled
with the progress made in wireless communication,
has made possible to perform mobile DM (Horovitz,
Gaber, & Krishnaswamy, 2005). Consequently,
considerable effort has focused on research and
development in this area.
The main functions performed by these systems
are: (1) data acquisition of biological sensors, (2) data
analysis, to detect some abnormal situations, and if
they are detected, then (3) generation alarm and
notifying doctors so they can make the right decision
(Goñi et al., 2009). Next, we will describe the
experimental settings we worked on regarding the
data description, data pre-processing, features
selection, DM techniques used and results. We
developed prediction models for dysfunction/failure
of five organic systems - cardiovascular, respiratory,
renal, coagulation and liver - as well as the outcome.
4.1 Data Description and Pre-processing
The data were gathered in the ICU of HGSA and
were collected in the first five days of stay of thirty
two patients.
They can be divided in five groups, described below.
1) SOFA Scores:
In intensive care, there are some scores to assess
severity of illness, like the Sequential Organ Failure
Assessment (SOFA), which is commonly used in ICU
on a daily basis to score the degree of
dysfunction/failure of six organic systems –
Cardiovascular, Respiratory, Renal, Liver,
Coagulation and Neurological (Vincent et al., 1996).
SOFA is scored in a scale from 0 (normality) to 4
(failure) for each organic system. In this experiment,
we transformed the SOFA scores in binary variables,
where 0 describes normality and 1 describes
dysfunction/failure and comprises the original SOFA.
SOFA
Cardio, Resp, Renal, Liver, Coagulat, neuro
= {0,1}
The variables required to calculate de SOFA
scores derive from heterogeneous sources, with
different frequencies, as shown in table 1.
Table 1: Data sources for sofa score calculation.
SOFA
Variables Source Frequency
Cardiovascular
Blood Pressure
BM Minute
Dopamine, dobutamine,
noradrenaline
EHR Day
Respiratory PaO
2
/FiO
2
NR Day
Renal Creatinine
EHR Day
Liver
Bilirubin EHR Day
Coagulation
Blood platets EHR Day
Neurological
Glasgow Coma Score NR Hour
In the data collecting phase, we have realized that the
Glasgow Come Score is rarely registered, so it was
not possible to generate models for the neurological
system due to the missing data regarding its
evaluation. As we can see in table 1, not only
variables are registered in different sources, but also
they have different frequencies. E.g. Blood Pressure
is registered every minute, whereas creatinine is only
measure and registered once a day. Taking this into
account, for the construction of the final dataset, we
transformed every SOFA score to an hourly register,
for we are making predictions in an hourly basis.
2) Critical Events:
Despite being commonly used, the SOFA score is
controversial, for it can’t predict the precisely
individual outcome (Vincent et al., 1998). As a
consequence, research has evolved towards the use of
intermediate outcomes such as Critical Events (CE)
(Silva, Cortez, Santos, Gomes, & Neves, 2008). CE
was defined by a panel of experts and relates to four
physiological variables – Blood Pressure (BP), Heart
Rate (HR), Urine Output (UR) and Oxygen
Saturation (O2). Whenever these parameters are out
of normal ranges for 10 minutes, it is considered a CE
for the corresponding parameter. They were
calculated hourly and, subsequently, it was derived a
new variable – Accumulated Critical Events (ACE) –
to reflect the patients’ clinical evolution/severity of
illness.
3) Case Mix = {
Age, Admission type, Admission from}
4) Ratios = {ACE
BP
, ACE
SO2
, ACE
HR
, ACE
Ur
}
5) MEWS = {MS
BP
, MS
HR
, MS
RR
, MS
Temp
, MS
AVPU
}
ENABLING UBIQUITOUS DATA MINING IN INTENSIVE CARE - Features Selection and Data Pre-processing
263
The Modified Early Warning Score (MEWS) is a
simple, physiological score that may allow
improvements in the quality and safety of
management provided to surgical ward patients
(Gardner-Thorpe, Love, Wrightson, Walsh, &
Keeling, 2006). This score uses data derived from four
physiological parameters - Systolic Blood Pressure,
Heart Rate, Respiratory Rate and Body Temperature -
and levels of consciousness - Alert, Voice, Pain,
Unresponsive (AVPU) - as shown in table 2. The
MEWS (MS) uses a scale from 0 (normal) to 3
(extreme risk) to classify the danger of each score.
With the data collected in the ICU, we were able to
calculate the score for only two variables for the
MEWS: Blood Pressure and Heart Rate.
Table 2: variables and data sources for the mews
calculation.
Score
Scale
Source Frequency
Blood Pressure
BM Minute
Heart Rate
BM Minute
Respiratory Rate
BM Minute
Body temperature
BM Minute
AVPU
NR Daily
To use it in final dataset, with hourly data, it was
calculated the number of occurrences for each score in
each hour, i.e., how many minutes the values were
out of normal range (danger) and its score.
In conclusion, the data required to calculate the
variables derive from heterogeneous sources -
Bedside Monitors (BM), Nursing Records (NR) and
Electronic Health Records (EHR) - as shown in Table
2. By exploring different and incremental scenarios
we intend to attest the importance of new variables,
other than the critical events. The inclusion of ratios
points to the severity of the patient’s clinical
evolution and the inclusion of SOFA scores relates to
the multi-organ failure perspective.
4.2 Features - Incremental Approach
For the prediction of each organic system and
outcome, it was explored five scenarios regarding the
inclusion of the variables mentioned before – M1,
M2, M3, M4 and M5. Previous research (Villas Boas,
et al., 2010) has shown that the inclusion of ratios and
the SOFA scores of other organic system for the
prediction of dysfunction/failure of some systems
presented better results when comparing to the
approach on the CE alone. Accordingly, we have
added a new variable – MS - as input and compared
the results to conclude if its inclusion leads to better
results. Also, we have tried to use exclusively the MS
to predict organ failure/dysfunction of the
cardiovascular system. We used the MS only for the
prediction of this system because the only available
data relates to. Data for the other organic systems, as
shown in table 3, was not registered.
M1 = {Case Mix, CE}
M2 = {Case Mix, CE, Ratios}
M3 = {Case Mix, CE, Ratios, SOFA}.
M4 = {Case Mix, CE, Ratios, SOFA, MEWS}
M5 = {Case Mix, MEWS}
4.2.1 Modelling
The DM techniques used for the experiments were
Artificial Neural Networks (ANN), Decision Trees
(DT), Regression and Ensembles. To avoid
overfitting due to their biased distribution, we
performed normalization by using the logarithmic
function. We used fully connected multilayer
perceptron’s with 5 hidden neurons and logistic
activation function. The training technique applied
was the backpropagation algorithm. To assure
statistical significance, 30 runs were applied to all
tests. For the DT, we used Classification and
Regression Tree (CART) algorithm. The splitting
method used for partitioning the data was the Gini
reduction. The default algorithm splits a node into 2
branches and, to avoid overfitting, the maximum
number of branches from a node was set to 10 and the
splits were evaluated as a reduction in impurity (Gini
index). We used logistic regression since the targets
have a binary measurement. The selection method
used was the Stepwise. The ensemble method used a
combined mode of the ANN, DT and Regression with
the Mean probability function.
4.3 Results
Previous research has concluded that, for intensive
care, is preferable to have models that favor
sensitivity (Vilas-Boas, et al., 2010). Hence, the
assessment of the models, presented in table 3, is
made in terms of sensitivity, derived from the
confusion matrices. For each system, we present the
DM technique that had the best results for each
scenario, in terms of sensitivity.
As we explained before, MEWS could only be
used to relate to the cardiovascular system, so there
are no results for the other systems regarding M4 and
M5. Table 4 sums the results and presents, for each
organic system, the best scenario and the technique
with the best results, in terms of sensitivity.
ICEIS 2011 - 13th International Conference on Enterprise Information Systems
264
Table 3: Results by scenario and technique for each target.
M
Target
in terms of sensivity (%)
Cardio Respirat Renal Liver Coag Out
M
1
DT
92.4
ANN
89.2
ANN
97.7
DT
94.3
ANN
94.2
ANN
98.3
M
2
ANN
91.3
ANN
96.2
ANN
98
ANN
95.4
ANN
97.5
ANN
98
M
3
ANN
93.4
DT
95
ANN
98.1
ANN
98.3
ANN
95.6
ANN
97.6
M
4
ANN
73.2
- - - - -
M
5
DT
90.0
- - - - -
Table 4: Best results for each target by scenario and
technique.
Target Scenario Sens Technique
Cardio M3
93.4 ANN
Respirat
M2 96.2 ANN
Renal M3
98.1 ANN
Liver
M3 98.3 ANN
Coag
M2 97.5 ANN
Out
M1 98.3 ANN
5 DISCUSSION
This experiment focused on features selection and
data pre-processing for the construction of the final
dataset for building the DM models for prediction of
organ dysfunction/failure. The experiments about the
features selection had the objective of analyzing the
impact on models’ sensitivity of adding more
variables. Accordingly, we conclude that adding the
score MEWS not only did not result in better models,
but actually worsened their sensitivity for all systems
and outcome. E.g. prediction of the cardiovascular
system has better results using CE, ratios and SOFA
scores (93.4%), however, when we added the MEWS
scores, sensitivity decayed to 73.2%. We can’t fairly
terminate that one shouldn’t use MEWS to predict
organ dysfunction/failure.
We also concluded that, regarding features
selection, the variables to be included are highly
dependent on the organic system; there isn’t a
scenario that is better for all systems, as shown in
table 3. E.g. for the prediction of the liver system, it
should be used CE, ratios and the SOFA of the other
systems (M3), whereas for the respiratory system, the
best results are achieved by not including the SOFA
(M2). The models have achieved great results in
terms of sensitivity and the best technique for all
systems and outcome was the ANN.
The INTCare system is still being tested,
however, in the future, the tasks regarding data
preprocessing will be automatically performed by the
application, with minimal human intervention and
awareness, with an Electronic Nursing Record
(ENR).
6 CONCLUSIONS AND FUTURE
WORK
In this paper we presented the work undergoing the
INTCare system, a real-time IDSS for intensive care.
We developed DM models for prediction of organ
dysfunction/failure and outcome and we presented the
scenarios and techniques we experimented.
As shown in table 3, the inclusion of the new
variable MEWS did not contribute to a better
performance of the models in terms of sensitivity.
However, one mustn’t discard this line of
investigation because in this experiment we only used
two out of six scores for MEWS. It would be
interesting to, in future research, rebuild the models
with all the six variables of MEWS and compare the
results. Also in future work, we will continue towards
ubiquity and pervasiveness of the system, with the
features presented in section 3 regarding DM and
decision models and intelligent agents and some
problems related to web traffic need to be studied and
predicted (Piramuthu, 2003).
ACKNOWLEDGEMENTS
The authors would like to thank FCT (Foundation of
Science and Technology, Portugal) for the financial
support through the contract PTDC/EIA/72819/2006.
REFERENCES
Bergs, E. A. G., Rutten, F. L. P. A., Tadros, T., Krijnen, P.,
& Schipper, I. B. (2005). Communication during
trauma resuscitation: do we know what is happening?
Injury, 36(8), 905-911.
Donchin, Y., & Seagull, F. J. (2002). The hostile
environment of the intensive care unit. Current opinion
in critical care, 8(4), 316.
Fano, A., & Gershman, A. (2002). The future of business
services in the age of ubiquitous computing.
Communications of the ACM, 45(12), 87.
Gardner-Thorpe, J., Love, N., Wrightson, J., Walsh, S., &
Keeling, N. (2006). The value of Modified Early
Warning Score (MEWS) in surgical in-patients: a
prospective observational study. Annals of The Royal
ENABLING UBIQUITOUS DATA MINING IN INTENSIVE CARE - Features Selection and Data Pre-processing
265
College of Surgeons of England, 88(6), 571.
Giudici, P. (2003). Applied data mining: Statistical methods
for business and industry: John Wiley & Sons, Inc.
Goñi, A., Burgos, A., Dranca, L., Rodríguez, J.,
Illarramendi, A., & Bermúdez, J. (2009). Architecture,
cost-model and customization of real-time monitoring
systems based on mobile biological sensor data-
streams. Computer Methods and Programs in
Biomedicine, 96(2), 141-157.
Horovitz, O., Gaber, M. M., & Krishnaswamy, S. (2005).
Making sense of ubiquitous data streams - A fuzzy
logic approach. In R. Khosla, R. J. Howlett & L. C. Jain
(Eds.), Knowledge-Based Intelligent Information and
Engineering Systems, Pt 2, Proceedings (Vol. 3682, pp.
922-928). Berlin: Springer-Verlag Berlin.
Hsu, J. (2002, 2002). Data mining trends and
developments: The key data mining technologies and
applications for the 21st century.
Kargupta, H. (2001, 2001). CAREER: Ubiquitous
Distributed Knowledge Discovery from Heterogeneous
Data.
Kohn, L. T., Corrigan, J., & Donaldson, M. S. (2000). To
Err Is Human: Building a Safer Health System:
National Academy Press.
Langheinrich, M. (2001). Privacy by Design - Principles of
Privacy-Aware Ubiquitous Systems. Paper presented at
the Proceedings of UbiComp 2001.
Menachemi, N., & Brooks, R. G. (2006). Reviewing the
benefits and costs of electronic health records and
associated patient safety technologies. Journal of
Medical Systems, 30(3), 159-168.
Neaga, E. I., & Harding, J. A. (2005). An enterprise
modeling and integration framework based on
knowledge discovery and data mining. International
Journal of Production Research, 43(6), 1089-1108.
Nigel, D. (2002). Beyond Prototypes: Challenges in
Deploying Ubiquitous Systems, 1, 26-35.
Orwat, C., Graefe, A., & Faulwasser, T. (2008). Towards
pervasive computing in health care–A literature review.
BMC Medical Informatics and Decision Making, 8(1),
26.
Piramuthu, S. (2003). On learning to predict web traffic.
Decision Support Systems, 35(2), 213-229.
Saha, D., & Mukherjee, A. (2003). Pervasive computing: a
paradigm for the 21st century. Computer, 25-31.
Santos, M. F., Portela, F., Vilas-Boas, M., Machado, J.,
Abelha, A., Neves, J., et al. (2009). Information
Modeling for Real-Time Decision Support in Intensive
Medicine. In S. Y. Chen & Q. Li (Eds.), Proceedings of
the 8th Wseas International Conference on Applied
Computer and Applied Computational Science -
Applied Computer and Applied Computational Science
(pp. 360-365). Athens: World Scientific and
Engineering Acad and Soc.
Scicluna, P., Murray, A., Xiao, Y., & Mackenzie, C. F.
(2008). Challenges to Real-Time Decision Support in
Health Care. Agency for Healthcare Research and
Quality
.
Silva, Á., Cortez, P., Santos, M. F., Gomes, L., & Neves, J.
(2008). Rating organ failure via adverse events using
data mining in the intensive care unit. Arificial
Intelligence in Medicine, 43(3), 179-193.
Varshney, U. (2007). Pervasive healthcare and wireless
health monitoring. Mobile Networks and Applications,
12(2), 113-127.
Varshney, U. (2009). Pervasive Healthcare Computing:
EMR/EHR, Wireless and Health Monitoring: Springer-
Verlag New York Inc.
Vilas-Boas, M., Santos, M. F., Portela, F., Silva, Á., & Rua,
F. (2010). Hourly prediction of organ failure and
outcome in intensive care based on data mining
techniques. Paper presented at the 12th International
Conference on Enterprise Information Systems.
Villas Boas, M., Gago, P., Portela, F., Rua, F., Silva, Á., &
Santos, M. F. (2010). Distributed and real time Data
Mining in the Intensive Care Unit. Paper presented at
the 19
th
European Conference on Artificial Intelligence
- ECAI 2010.
Vincent, J. L., de Mendonca, A., Cantraine, F., Moreno, R.,
Takala, J., Suter, P. M., et al. (1998). Use of the SOFA
score to assess the incidence of organ
dysfunction/failure in intensive care units: results of a
multicenter, prospective study. Critical care medicine,
26(11), 1793.
Vincent, J. L., Moreno, R., Takala, J., Willatts, S., De
Mendonca, A., Bruining, H., et al. (1996). The SOFA
(Sepsis-related Organ Failure Assessment) score to
describe organ dysfunction/failure. Intensive care
medicine, 22(7), 707-710.
ICEIS 2011 - 13th International Conference on Enterprise Information Systems
266