=Paper=
{{Paper
|id=Vol-1098/aih2013_Complete
|storemode=property
|title=None
|pdfUrl=https://ceur-ws.org/Vol-1098/aih2013_Complete.pdf
|volume=Vol-1098
}}
==None==
https://ceur-ws.org/Vol-1098/aih2013_Complete.pdf
The Third Australasian Workshop on
Artificial Intelligence in Health
AIH 2013
Fourth International Workshop on Collaborative Agents --
REsearch and Development (CARE) 2013
“CARE for a Smarter Society”
held in conjunction with the
26th Australasian Joint Conference on Artificial Intelligence (AI 2013)
Tuesday, 3rd December 2013
University of Otago, Dunedin, New Zealand
JOINT
WORKSHOP
PROCEEDINGS
Editors :
Sankalp Khanna1, Christian Guttmann2,
Abdul Sattar3, David Hansen1, Fernando Koch4
1 The Australian e-Health Research Centre, CSIRO Computational Informatics, Australia
2 IBM Research, Australia
3 Institute for Integrated and Intelligent Systems, Griffith University, Australia
4 Samsung Research Institute, Brazil
AIH 2013 - ACKNOWLEDGEMENTS
Program Chairs
• Abdul Sattar (Griffith University, Australia)
• David Hansen (CSIRO Australian e-Health Research Centre, Australia)
Workshop Chair
• Sankalp Khanna (CSIRO Australian e-Health Research Centre, Australia)
Senior Program Committee
• Aditya Ghose (University of Newcastle, Australia)
• Jim Warren (University of Auckland, New Zealand)
• Wayne Wobcke (University of New South Wales, Australia)
• Mehmet Orgun (Macquarie University, Australia)
• Yogesan (Yogi) Kanagasingam (CSIRO Australian e-Health Research Centre, Australia)
Program Committee
• Simon McBride (CSIRO AEHRC) • Kewen Wang (Griffith University)
• Adam Dunn (University of New South Wales) • Vladimir Estivill-Castro (Griffith University)
• Stephen Anthony (University of New South • John Thornton (Griffith University)
Wales) • Bela Stantic (Griffith University)
• Lawrence Cavedon (Royal Melbourne Institute • Byeong-Ho Kang (University of Tasmania)
of Technology / NICTA) • Justin Boyle (CSIRO AEHRC)
• Diego Mollá Aliod (Macquarie University) • Guido Zuccon (CSIRO AEHRC)
• Michael Lawley (CSIRO AEHRC) • Hugo Leroux(CSIRO AEHRC)
• Anthony Nguyen (CSIRO AEHRC) • Alejandro Metke (CSIRO AEHRC)
• Amol Wagholikar (CSIRO AEHRC)
• Bevan Koopman (CSIRO AEHRC)
Key Sponsors
• CSIRO Australian e-Health Research Centre
• Institute for Integrated and Intelligent Systems, Griffith University
Supporting Organisations
• The Australasian College of Health Informatics
• The Australasian Medical Journal
• The Australasian Telehealth Society
CARE 2013 - ACKNOWLEDGEMENTS
General Chairs
• Christian Guttmann (IBM Research -- Australia)
• Fernando Koch (Samsung Research Institute -- Brazil)
ABSEES Track Chairs:
• Maryam Purvis, maryam.purvis@otago.ac.nz
• Takayuki Ito, ito.takayuki@nitech.ac.jp
Program Committee
• Andrew Koster • Marcelo Ribeiro
• Anthony Patricia • Martin Purvis
• Bastin Tony Roy Savarimuthu • Meritxell Vinyals
• Benjamin Hirsch • Michael Thielscher
• Carlos Cardonha • Neil Yorke-Smith
• Cristiano Castelfranchi • Priscilla Avegliano
• David Morley • Rainer Unland
• Diego Gallo • Ryo Kanamori
• Frank Dignum • Sankalp Khanna
• Franziska Klügl • Sarvapali Ramchurn
• Gordon McCalla • Sascha Ossowski
• Ingo J. Timm • Shantanu Chakraborty
• Inon Zuckerman • Sherief Abdallah
• Jaime Sichman • Simon Thompson
• Kobi Gal • Simon Goss
• Lars Braubach • Toby Walsh
• Lawrence Cavedon • Wayne Wobcke
• Leonardo Garrido • Wei Chen
• Liz Sonenberg • Zakaria Maamar
• Magnus Boman
Joint Proceedings - AIH 2013 / CARE 2013
TABLE OF CONTENTS
KEYNOTE ADDRESS
Health Informatics and Artificial Intelligence solutions: Addressing the
Challenges at the Frontiers of Modern Healthcare 3
Professor Michael Blumenstein
AIH 2013 FULL PAPERS
Classification Models in Intensive Care Outcome Prediction-can we improve on
current models? 5
Nicholas Barnes, Lynette Hunt, and Michael Mayo
Towards a visually enhanced medical search engine
22
Lavish Lalwani, Guido Zuccon, Mohamed Sharaf and Anthony Nguyen
Using Fuzzy Logic for Decision Support in Vital Signs Monitoring
29
Shohas Dutta, Anthony Maeder and Jim Basilakis
A Novel Approach for Improving Chronic Disease Outcomes using Intelligent
Personal Health Records in a Collaborative Care Framework 34
Amol Wagholikar
AIH 2013 SHORT PAPERS
Partially automated literature screening for systematic reviews by modelling non-
relevant articles 43
Henry Petersen, Josiah Poon, Simon Poon, Clement Loy and Mariska Leeflang
CARE 2013 FULL PAPERS
Optimizing Shiftable Appliance Schedules across Residential Neighbourhoods for
Lower Energy Costs and Fair Billing 45
Salma Bakr and Stephen Cranefield
Proposal of information provision to probe vehicles based on distribution of link
travel time that tends to have two peaks 53
Keita Mizuno, Ryo Kanamori, and Takayuki Ito
Page 1
Joint Proceedings - AIH 2013 / CARE 2013
Page 2
Joint Proceedings - AIH 2013 / CARE 2013
Health Informatics and Artificial Intelligence solutions: Addressing
the Challenges at the Frontiers of Modern Healthcare
Keynote Address
Professor Michael Blumenstein
Griffith University, Australia
m.blumenstein@griffith.edu.au
Speaker Profile
Michael Blumenstein is a Professor and Head of
the School of Information and Communication
Technology at Griffith University, where he previously
served as the Dean (Research) in the Science,
Environment, Engineering and Technology Group. In
addition, Michael currently serves as the Leader for
the Health Informatics Flagship Program at the
Institute for Integrated and Intelligent Systems.
Michael is a nationally and internationally
recognised expert in the areas of automated Pattern
Recognition and Artificial Intelligence, and his current
research interests include Document Analysis, Multi-
Script Handwriting Recognition and Signature
Verification. He has published over 132 papers in refereed books, conferences and
journals. His research also spans various projects applying Artificial Intelligence to
the fields of Engineering, Environmental Science, Neurobiology, Coastal
Management and Health. Michael has secured internal/nationally competitive
research grants to undertake these projects with funds exceeding AUD$4.3 Million.
Components of his research into the predictive assessment of beach conditions have
been commercialised for use by local government agencies, coastal management
authorities and in commercial applications.
Following his achievements in applying Artificial Intelligence to the area of bridge
engineering (where he has published widely and has been awarded federal funding),
he was invited to serve on the International Association for Bridge and Structural
Engineering’s Working Commission 6 to advise on matters pertaining to Information
Technology. Michael is the first Australian to be elected onto this committee. In
addition, he was previously the Chair of the Queensland Branch of the Institute for
Electrical and Electronic Engineers (IEEE) Computational Intelligence Society. He is
also the Gold Coast Chapter Convener and a Board Member of the Australian
Page 3
Joint Proceedings - AIH 2013 / CARE 2013
Computer Society's Queensland Branch Executive Committee as well as the
Chairman of the IT Forum Gold Coast and a Board Member of IT Queensland.
Michael currently serves on the Australian Research Council's (ARC) College of
Experts on the Engineering, Mathematics and Informatics (EMI) panel. In addition,
he has recently been elected onto the Executive of the Australian Council of Deans of
Information and Communication Technology (ACDICT). Michael also serves on a
number of Journal Editorial Boards and has been invited to act as General Chair,
Organising Chair, Program Chair and/or Committee member for numerous
national/international conferences in his areas of expertise.
In 2009 Michael was named as one of Australia’s Top 10 Emerging Leaders in
Innovation in the Australian’s Top 100 Emerging Leaders Series supported by
Microsoft. Michael is a Fellow of the Australian Computer Society and a Senior
Member of the IEEE.
Abstract
Numerous challenges currently exist in the Health Sector such as effective
treatment of patients with chronic diseases, early diagnosis and prediction of health
conditions, patient data administration and adoption of electronic health records,
strategic planning for hospitals and engagement of health professionals in training.
This presentation focuses on these challenges and examines some innovative Health
Informatics solutions with prospective deployment of automated artificial
intelligence tools to augment current practices.
Some challenges are examined at a brand new University Hospital in Queensland,
whereby a number of automated solutions are investigated using technology and
intelligent approaches such as mobile devices for understanding patient chronic
health conditions over time, image analysis and pattern recognition for the early
diagnosis and treatment of such brain disorders as Parkinson's disease, social media
analytics for patient engagement in the adoption of electronic health records, on-
line collaborative tools for strategic planning in the hospital and the use of 3D virtual
worlds for realistic training and professional development for medical staff.
Finally, the presentation will conclude with a discussion about the emerging
"Research Triangle" present at the Gold Coast, in Queensland, which includes the
new Gold Coast University Hospital and is directly adjacent to Griffith University's
Gold Coast campus with proximity to the emerging Health and Knowledge Precinct.
This special zone presents a unique opportunity to nurture cutting edge health-
related research intersecting information technology in collaboration with industry
and government, which may have a profound impact on the future landscape of
Health Informatics innovation in the region.
Page 4
Joint Proceedings - AIH 2013 / CARE 2013
Classification Models in Intensive Care Outcome
Prediction-can we improve on current models?
Nicholas A. Barnes,
Intensive Care Unit,
Waikato Hospital, Hamilton, New Zealand.
Lynnette A. Hunt,
Department of Statistics,
University of Waikato, Hamilton, New Zealand.
Michael M. Mayo,
Department of Computer Science,
University of Waikato, Hamilton, New Zealand.
Corresponding Author: Nicholas A. Barnes.
Abstract
Classification models (“machine learners” or “learners”) were developed using
machine learning techniques to predict mortality at discharge from an intensive
care unit (ICU) and evaluated based on a large training data set from a single
ICU. The best models were tested on data on subsequent patient admissions.
Excellent model performance (AUCROC (area under the receiver operating
curve) =0.896 on a test set), possibly superior to a widely used existing model
based on conventional logistic regression models was obtained, with fewer per-
patient data than that model.
1 Introduction
Intensive care clinicians use explicit judgement and heuristics to formulate prog-
noses as soon as reasonable after patient referral and admission to an intensive care
unit [1].
Models to predict outcome in such patients have been in use for over 30 years [2]
but are considered to have insufficient discriminatory power for individual decision
making in a situation where patient variables that are difficult or impossible to meas-
ure may be relevant. Indeed even variables that have little or nothing to do with the
patient directly (such as bed availability or staffing levels [3]) may be important in
determining outcome.
There are further challenges for model development. Any model used should be
able to deal with the problem of class imbalance, which refers in this case to the fact
Page 5
Joint Proceedings - AIH 2013 / CARE 2013
that mortality should be much less common than survival. Many patient data are
probably only loosely or indeed not related to outcome and many are highly corre-
lated. For example, elevated measurements of serum urea, creatinine, urine output,
diagnosis of renal failure and use of dialysis will all be closely correlated.
Nevertheless, models are used to risk adjust for comparison within an institution
over time or between institutions, and model performance is obviously important if
this is to be meaningful. It is also likely that a model with excellent performance
could augment clinical assessment of prognosis. Furthermore, a model that performs
well while requiring fewer data would be helpful as accurate data acquisition is an
expensive task.
The APACHE III-J (Acute Physiology and Chronic Health Evaluation revision III-
J [4]) model is used extensively within Australasia by the Centre for Outcomes Re-
search of the Australian and New Zealand Intensive Care Society (ANZICS) and a
good understanding of its local performance is available in the published literature
[4]. It should be noted that death at hospital discharge is the outcome variable usually
considered by these models. Unfortunately the coefficients for all variables for this
model are no longer in the public domain so direct comparison with new models is
difficult. The APACHE (Acute Physiology and Chronic Health Evaluation) models
are based largely on baseline demographic and illness data and physiological mea-
surements taken within the first day after ICU admission.
This study aims to explore machine learning methods that may outperform the lo-
gistic regression models that have previously been used.
The reader may like to consult a useful introduction to the concepts and practice of
machine learning [5] if terms or concepts are unfamiliar.
2 Methods
The study is comprised of three parts:
1. An empirical exploration of raw and processed admission data with a variety of
attribute selection methods, filters, base classifiers and metalearning techniques
(which are overarching models that have other methods nested within them) that
were felt to be suitable to develop the best classification models. Metamodels and
base classifiers may be nested within other metamodels and learning schemes can
be varied in very many ways .These experiments are represented below in Figure 1
where we used up to two metaclassifiers with up to two base classifiers nested
within a metaclassifier.
Page 6
Joint Proceedings - AIH 2013 / CARE 2013
Choose
Dataset Metamodel 1 Metamodel 2
Evaluate
Base Classifier (s) Classifier
Results
Fig. 1. Schematic of phase 1 experiments. Different color arrows indicate that one or more
metamodels and base classifiers may optionally be combined in multiple different ways. One or
more base classifiers are always required.
2. Further testing with the best performing data set (full unimputed training set) and
learners with manual hyperparameter setting. A hyperparameter is a particular
model configuration that is selected by the user, either manually or following an
automatic tuning process. This is represented in a schematic below:
Fig. 2. Schematic of phase 2 experiments. As in phase 1, one or more metamodels may be
optionally combined with one or more base classifiers.
3. Testing of the best models from phase 2 above on a new set of test data to better
understand generalizability of the models. This is depicted in Figure 3 below.
Page 7
Joint Proceedings - AIH 2013 / CARE 2013
Four Best Models
Matching based on 4 Evaluate Classifier
Test Set Evaluation Results on Test Set
Measures
Fig. 3. Schematic of phase 3
The training data for adult patients (8122 patients over 16 years of age) were ob-
tained from the database of a multidisciplinary ICU in a tertiary referral centre from a
period between July 2004 and July 2012.Data extracted were comprised of a demo-
graphic variable (age), diagnostic category (with diagnostic coefficient from the
APACHE III-J scoring system, including ANZICS modifications), and an extensive
list of numeric variables relating to patient physiology and composite scores based on
these, along with the classification variable: either survival, or alternatively, death at
ICU discharge (as opposed to death at hospital discharge as in the APACHE models).
Much of the data collected is used in APACHE III-J model mentioned above, and
represents a subset of the data used in that model. Training data, prior to the imputa-
tion process, but following discretization of selected variables are represented in Ta-
ble 1. Test data for the identical variable set were obtained from the same database for
the period July 2012 to March 2013.
Of particular interest is that the data is clearly class imbalanced with mortality dur-
ing ICU stay of approximately 12%. This has important implications for modelling
the data.
There were many strongly correlated attributes within the data sets. Many of the
model variables are collected as highest and lowest measures within twenty four
hours of admission to the ICU. Correlated variables may bring special problems with
conventional modelling including logistic regression. The extent of correlation is
demonstrated in Figure 4.
Page 8
Joint Proceedings - AIH 2013 / CARE 2013
Fig. 4. Pearson correlations between variables are shown using colour. Blue colouration indi-
cates positive correlation. Red colouration indicates negative correlation. The flatter the ellipse,
the higher the correlation. White circles indicate no significant correlation between variables.
Patterns of missing data are indicated in Table 1 and represented graphically in
Figure 5.
Page 9
Joint Proceedings - AIH 2013 / CARE 2013
Fig. 5. Patterns of missing data in the raw training set. Missing data is represented by red
colouration.
Missing numeric data in the training set was imputed using multiple imputation
with the R program [6] and the R package Amelia [7], which utilises bootstrapping of
non-missing data followed by imputation by expectation maximisation. We initially
used the average of five multiple imputation runs.
Using the last imputed set was also trialled, as it may be expected to be the most
accurate based on the iterative nature of the Amelia algorithm. No categorical data
were missing. Date of admission was discretized to the year of admission, age was
converted to months of age, and the diagnostic categories were converted to five to
eight (depending on study phase) ordinal risk categories by using coefficients from
the existing APACHE III-J risk model.
A summary of data is presented below in Table 1.
Table 1. Data Structure
Variable Type Missing Distinct Min. Max.
values
CareUnitAdmDate numeric 0 9 2004 2012
AgeMonths numeric 0 880 192 1125
Sex pure factor 0 2 F M
Risk pure factor 0 8 Vlow High
Page 10
Joint Proceedings - AIH 2013 / CARE 2013
CoreTempHi numeric 50 89 29 42.3
CoreTempLo numeric 53 102 25.2 40.7
HeartRateHi numeric 25 141 38.5 210
HeartRateLo numeric 26 121 0 152
RespRateHi numeric 38 60 8 80
RespRateLo numeric 40 42 2 37
SystolicHi numeric 27 161 24 288
SystolicLo numeric 55 151 11 260
DiastolicHi numeric 27 105 19 159
MAPHi numeric 28 124 20 200
MAPLo numeric 43 103 3 176
NaHi numeric 46 240 112 193
NaLo numeric 51 245 101 162
KHi numeric 46 348 2.7 11.7
KLo numeric 51 275 1.4 9.9
BicarbonateHi numeric 218 322 3.57 48
BicarbonateLo numeric 221 319 2 44.2
CreatinineHi numeric 130 606 10.2 2025
CreatinineLo numeric 134 552 10 2025
UreaHiOnly numeric 232 433 1 99
UrineOutputHiOnly numeric 184 3501 0 15720
AlbuminLoOnly numeric 281 66 5 65
BilirubinHiOnly numeric 1579 183 0.4 618
GlucoseHi numeric 172 255 1.95 87.7
GlucoseLo numeric 177 198 0.1 60
HaemoglobinHi numeric 54 153 1.8 25
HaemoglobinLo numeric 59 151 1.1 25
WhiteCellCountHi numeric 131 470 0.1 293
WhiteCellCountLo numeric 135 393 0.08 293
PlateletsHi numeric 149 653 7 1448
PlateletsLo numeric 153 621 0.27 1405
OxygenScore numeric 0 8 0 15
pHAcidosisScore numeric 0 9 0 12
GCSScore numeric 0 11 0 48
ChronicHealthScore numeric 0 6 0 16
Status at ICU Discharge pure factor 0 2 A D
Phase 1 consisted of an exploration of machine learning techniques thought suit-
able to this classification problem, and in particular those thought to be appropriate to
a class imbalanced data set. Attribute selection, examining the effect of using imputed
and unimputed data sets and application of a variety of base learners and metaclassifi-
ers without major hyperparameter variation occurred in this phase. The importance of
Page 11
Joint Proceedings - AIH 2013 / CARE 2013
attributes was examined in multiple ways including using random forest methodology
for variable selection, using improvement in Gini index using particular attributes.
This information is displayed in figure 6.
Fig. 6. Variable importance as measured by Gini index using random forest methodology. A
substantial decrease in Gini index indicates better classification with variable inclusion. Va-
riables used in the study are ranked by their contribution to Gini index.
A comprehensive evaluation of all techniques is nearly impossible given the
enormous variety of techniques and the ability to combine up to several of these at a
time in any particular model. Techniques were chosen based on the likely success of
their application. WEKA [8] was used to apply learners and all models were eva-
luated with tenfold cross validation. WEKA default settings were commonly used in
phase 1 and the details of these defaults are widely available [9]. Unless otherwise
stated all settings in all study phases were the default settings of WEKA for each clas-
sifier or filter. Two results were used to judge overall model performance during
phase 1. These were:
1. Area under the receiver operating curve (AUC ROC)
2. Area under the precision recall curve (AUC PRC)
The results are presented in Table 3 in the results section.
Page 12
Joint Proceedings - AIH 2013 / CARE 2013
Phase 2 of our study involved training and evaluation on the same data sets with
learners that had performed well in phase 1. Hyperparameters were mostly selected
manually, as automatic hyperparameter selection in any software is limited and ham-
pered by a lack of explicitness. Class imbalance issues were addressed with appropri-
ate WEKA filters (spread subsample and SMOTE, a filter which generates a synthetic
data set to balance the classes [10]), or the use of cost sensitive learners [11]. Unless
otherwise stated in Table 3, WEKA default settings were used for each filter or classi-
fier. Evaluation of these models proceeded with tenfold cross-validation and the re-
sults were examined in light of four measures:
1. Area under the receiver operating curve with 95% confidence intervals by the
method of Hanley and McNeill [12]
2. Area under the precision recall curve
3. Matthews correlation coefficient and,
4. F-measure
Additionally, scaling the quantitative variables by standardizing or normalizing the
data was explored as this is known to sometimes improve model performance [13].
The results of phase 2 are presented in Table 2 in the results section.
Phase 3 involved evaluating the accuracy of the best classification models from phase
2 on a new test set of 813 patient admissions. Missing data in the test set were not
imputed. Results are shown in Table 3.
3 Results
Table 2 presents the results following tenfold cross validation on a variety of
techniques thought suitable for trial in the modelling problem. These are listed in
order of descending area under the curve of the receiver operating curve and the area
under the precision recall curve is also presented.
Table 2. Phase 2 of study.
Base
Meta Meta model Meta Base classi-
Data Preprocess classifier ROC PRC
Model 1 2 model 3 fier 1
2
Cost
Sensitive Random
Unimputed
NA Classifier NA NA Forest 500 NA 0.895 0.629
all variables
matrix trees
0,5;1,0
Cost
Sensitive Random
Unimputed
NA Classifier NA NA Forest 200 NA 0.894 0.416
all variables
matrix trees
0,5;1,0
Cost
Sensitive
Unimputed
NA Classifier NA NA Naïve Bayes NA 0.864 0.418
all variables
matrix
0,5;1,0
Page 13
Joint Proceedings - AIH 2013 / CARE 2013
Attribute
selected
Spread- classifier 20
Unimputed Filtered Naïve
subsample variables Vote J4.8 tree 0.854 0.439
all variables Classifier Bayes
uniform selected on
info. Gain
and ranked
Spread-
Imputed ten Filtered Logistic Logistic
subsample NA NA 0.766 0.283
variables Classifier regression Regression
uniform
Spread-
Imputed ten Filtered SimpleLogis-
subsample NA NA NA 0.766 0.28
variables Classifier tic
uniform
Spread-
Imputed ten Filtered Random
subsample NA REP tree NA 0.753 0.259
variables Classifier Comm
uniform
Imputed ten Filtered
NA NA NA Naïve Bayes NA 0.742 0.248
variables Classifier
Spread-
Imputed ten Filtered
subsample Adaboost M1 NA J48 NA 0.741 0.254
variables Classifier
uniform
Spread- Random
Imputed ten Filtered Naïve
subsample Vote NA Forest 10 0.741 0.252
variables Classifier Bayes
uniform trees
Spread-
Imputed ten Filtered
subsample Bagging NA J48 NA 0.736 0.258
variables Classifier
uniform
Spread-
Imputed ten Filtered
subsample Decorate NA Naïve Bayes NA 0.735 0.238
variables Classifier
uniform
Attribute
selected
Spread- classifier 20
Imputed all Filtered Naïve
subsample variables Vote J4.8 tree 0.735 0.238
variables Classifier Bayes
uniform selected on
info. Gain
and ranked
Spread-
Imputed ten Filtered
subsample NA NA J4.8 tree NA 0.734 0.234
variables Classifier
uniform
Spread- Random
Imputed ten Filtered
subsample NA NA Forest 10 NA 0.713 0.221
variables Classifier
uniform trees
Spread-
Imputed ten Filtered
subsample SMO NA SMO NA 0.5 0.117
variables Classifier
uniform
ROC-area under receiver operating characteristic curve
CI-confidence interval
PRC-area under precision-recall curve
NA-not applicable
Table 3 presents the results of tenfold cross validation on the best models from
phase 1 trained on the training set in phase 2 of our study. Models are listed in des-
cending order of AUC ROC. The data set used in the modelling is indicated, along
with any pre-processing of data, base learners, metalearners if applicable, and other
evaluation tools as listed in the methods section above. The model which performs
Page 14
Joint Proceedings - AIH 2013 / CARE 2013
best of all models on any of the four classification methods is shaded in red to empha-
sise that no one performance measure dominates a classifier’s overall utility.
Table 3. Phase 2 results
Base F-
Preprocess Metamodel1 Metamodel2 Base Model 1 ROC ROC 95% CI's PRC MCC
model 2 measure
Spread Rotation Alternating
Filtered
subsample forest 100 decision tree NA 0.903 (0.892,0.912) 0.622 0.47 0.51
classifier
uniform iterations 100 iterations
Cost sensi-
Rotationforest
NA tive classi- NA J 48 0.901 (0.881,0.921) 0.625 0.482 0.481
500 iterations
fier 0,5;1,0
Spread
Filtered Rotationforest
subsample NA J 48 0.897 (0.888,0.906) 0.606 0.452 0.494
classifier 200 iterations
uniform
Spread
Filtered Rotationforest
subsample NA J 48 0.897 (0.888,0.906) 0.608 0.45 0.493
classifier 500 iterations
uniform
Spread Rotation
Filtered J48
subsample NA forest 500 0.897 (0.888,0.906) 0.611 0.456 0.5
classifier graft
uniform iterations
Spread Rotation Alternating
Filtered
subsample forest 50 decision tree NA 0.896 (0.887,0.905) 0.608 0.452 0.495
classifier
uniform iterations 50 iterations
Spread Rotation
Filtered
subsample NA forest 100 J 48 0.895 (0.886,0.904) 0.602 0.443 0.488
classifier
uniform iterations
Random
Cost sensi- forests (RF)
NA tive classi- NA 1000 trees 2 NA 0.893 (0.879,0.907) 0.599 0.506 0.561
fier 0,5;1,0 features each
tree
Cost sensi- RF 500 trees 2
NA tive classi- NA features each NA 0.892 (0.878.0.906) 0.598 0.511 0.567
fier 0,5;1,0 tree
Cost sensi- RF 500 trees 2
NA tive classi- NA features each NA 0.891 (0.867,0.915) 0.602 0.416 0.398
fier 0,1;1,0 tree
Cost sensi- RF 1000 trees
NA tive classi- NA 2 features NA 0.891 (0.867,0.915) 0.603 0.422 0.391
fier 0,1;1,0 each tree
Cost sensi- RF 500 trees 2
NA tive classi- NA features each NA 0.891 (0.878,0.904) 0.594 0.497 0.558
fier 0,10;1,0 tree
Cost sensi- Rotation
NA tive classi- NA Forest 50 J48 0.891 (0.871,0.911) 0.606 0.479 0.485
fier 0,5;1,0 iterations
Spread Filtered Bagging 150 J 48 C 0.25 M
NA 0.89 (0.869,0.911) 0.609 0.474 0.471
subsample classifier iterations 2
Spread Filtered Bagging 200 J 48 C 0.25 M
NA 0.889 (0.868,0.910) 0.61 0.474 0.473
subsample classifier iterations 3
Cost sensi- RF 200 trees 2
NA tive classi- NA features each NA 0.889 (0.865,0.913) 0.598 0.425 0.395
fier 0,1;1,1 tree
Spread Filtered Bagging 100 J 48 C 0.25 M
NA 0.888 (0.867,0.909) 0.605 0.47 0.467
subsample classifier iterations 2
Page 15
Joint Proceedings - AIH 2013 / CARE 2013
Cost sensi- RF 100 trees 2
NA tive classi- NA features each NA 0.888 (0.864,0.912) 0.594 0.42 0.396
fier 0,5;1,0 tree
Spread Random
Filtered Random
subsample NA committee 0.887 (0.879,0.895) 0.578 0.373 0.409
classifier tree
uniform 500 iterations
Spread Filtered Adaboost M1 J 48 C 0.25 M
NA 0.886 (0.865,0.907) 0.584 0.48 0.476
subsample classifier 150 iterations 2
Spread Filtered Adaboost M1 J 48 C 0.25 M
NA 0.884 (0.863,0.905) 0.577 0.469 0.467
subsample classifier 100 iterations 2
Spread Filtered Bagging 50 J 48 C 0.25 M
NA 0.883 (0.862,0.904) 0.597 0.465 0.465
subsample classifier iterations 2
Spread Random
Filtered REP
subsample NA subspace 100 0.877 (0.868,0.886) 0.563 0.423 0.473
classifier tree
uniform iterations
Spread
Filtered Multiboost AB
subsample NA J 48 0.874 (0.864,0.884) 0.428 0.435 0.482
classifier 50 iterations
uniform
RF-random forest
REP-representative
NA-not applicable
MCC-Matthews correlation coefficient
Normalizing or standardizing the data did not improve model performance and in-
deed tended to moderately worsen it.
Table 4 presents the results of applying four of the best models from phase 2 on a
test data set of 813 patient admissions which should be from the same population
distribution (if date of admission is not a relevant attribute). Evaluation is based on
AUC ROC, AUC PRC, Matthews’s correlation coefficient and F-measure. These
evaluations were obtained by WEKA’s knowledge flow interface.
Table 4. Model results with new test set in Phase 3
Base
Data prepro- Metamo- Metamo- Base 95% CI
Clas- ROC PRC MCC F-meas
cessing del 1 del 2 Classifer 1 ROC
sifier 2
Alternat-
Spread Rotation ing
Filtered (0.854,0
subsample forest 100 decision NA 0.896 0.592 0.401 0.426
classifier .938)
uniform iterations tree 100
iterations
Spread Rotation
Filtered (0.863,0
subsample forest 200 NA J 48 0.893 0.571 0.525 0.534
classifier .923)
uniform iterations
Cost
Rotation
sensitive (0.821,0
NA NA forest 500 J 48 0.887 0.561 0.386 0.411
classifier .953)
iterations
0,5;1,0
Random
Cost
forest 500
sensitive (0.855,0
NA NA trees, 2 NA 0.885 0.551 0.51 0.555
classifier .915)
features
0,5;1,0
each tree
Page 16
Joint Proceedings - AIH 2013 / CARE 2013
ROC-area under receiver operating characteristic curve
CI-confidence interval
PRC-area under precision-recall curve
MCC-Matthews correlation coefficient
F-meas-F-measure
4 Discussion
It is unrealistic to expect models to perfectly represent such a complex reality as
that of survival from critical illness. Perfect classification is impossible because of the
limitations of any combination of currently available measurements made on such
patients to accurately reflect survival potential. Patient factors such as attitudes to-
wards artificial support and presumably health practitioner and institution related
factors are important. Additionally non-patient related factors which may be purely
logistical will continue to thwart perfect prediction by any future model. For instance,
a patient may die soon after discharge from the ICU if a ward bed is available and
conversely will die within the ICU if a ward bed is not available and transfer cannot
proceed. Models currently employed generally consider death at hospital discharge,
but new factors that increase randomness can enter in the hospital stay following ICU
discharge, so problems are not necessarily decreased with this approach.
The best models we have studied have excellent performance when evaluated fol-
lowing tenfold cross validation in the single ICU setting with use of fewer data points
than the current gold standard model. Machine learning techniques usually make few
distributional assumptions about the data when compared with the traditional logistic
regression model. Missing data are often dealt with effectively with machine learning
techniques while complete cases are generally used in traditional general linear mod-
elling such as logistic regression. Clinical data will never be complete, as some data
will not be required for a given patient, while some patients may die prior to collec-
tion of data which cannot subsequently be obtained. Imputation may be performed on
data prior to modelling but has limitations. It is interesting that models trained on
unimputed data tend to perform better than imputed data, both in phase 2 and with the
test set in phase 3.
The best comparison we can make in the published literature is the work of Paul et
al [4] which demonstrates that the AUC ROC of the APACHE-III-J model has varied
between 0.879 and 0.890 when applied to over half a million adult admissions to Aus-
tralasian ICUs between 2000 and 2009. Routine exclusions in this study included
readmissions, transfers to other ICUs, and missing outcome and other data, and ad-
mission post coronary artery bypass grafting prior to introduction of the ANZICS
modification to APACHE-III-J for this category. None of these were exclusions in our
study. The Paul et al paper looks at outcome at hospital discharge, while ours ex-
amines outcome at ICU discharge. For these reasons the results are not directly com-
Page 17
Joint Proceedings - AIH 2013 / CARE 2013
parable but our results for AUC ROC of up to 0.896 on a separate validation set clear-
ly demonstrate excellent model performance.
The techniques associated with the best performance involve addressing class im-
balance (i.e. pre-processing data to create a dataset with similar numbers of those who
survive and those that die). This class imbalance is a well-known problem in classifi-
cation. Mortality data from any healthcare setting tend to be class imbalanced. Our
study shows that any approach to class imbalance in the data greatly enhance model
performance. Cost sensitive metalearners [11], synthetic minority generation tech-
niques (SMOTE [10]) and creating a uniform class distribution by subsampling across
the data all improve model performance.
A cost sensitive learner indicates a technique that reweights cases according to a
cost matrix that the user sets to reflect differing “cost” of misclassification of positive
and negative cases. This intuitively lends itself to the intensive care treatment process
where such a framework is likely implemented at least subconsciously by the inten-
sive care clinician. For instance the cost of clinically “misclassifying” a patient may
be substantial and clinicians would likely try hard to avoid this situation.
In our study, the ensemble learner random forests [14] with or without a technique to
address class imbalance tends to outperform many more complex metalearners, or
enhancements of single base classifiers such as bagging [15] and boosting [16]. Ran-
dom forests involve generation of many different tree models, each of which splits the
cases based on different variables and a criterion to increase information gain. Voting
then occurs across the “forest” to decide on the best way to split the cases and this
produces the model. The term ensemble simply represents the fact that multiple learn-
ers are involved, rather than a single tree. As many as 500 or 1000 trees are com-
monly required before the error of the forest is at a minimum. The number of vari-
ables to be considered by each tree may also be set to try and improve performance.
The other techniques that produced excellent results were rotation forests either alone,
with a cost sensitive classifier, or in combination with a technique known as alternat-
ing decision tree. Alternating decision tree takes a “weak” classifier (such as a tree
classifier) and uses a technique similar to boosting to improve performance.
The reason extensive experimentation may be required to produce the best model is
attributed to Wolpert [17] and described as the “no free lunch theorem”, meaning that
there is no one single technique that will model the best in every given scenario. Of
course the same is true of any conventional statistical technique applied to multidi-
mensional problems. Data processing and model selection are crucial to performance
although if prediction alone is important, a pragmatic approach can be taken to the
usual statistical assumptions. Machine learning techniques are generally not a “black
box” approach however and deserve the same credibility as any older method, if ap-
plication is appropriate.
Similarly, no single evaluation measure can summarize a classifier’s performance and
different model strengths and weaknesses may be more or less tolerable depending on
the circumstances of model use and hence a range of measures are usually presented
as we have done.
There are several weaknesses to our study. It is clearly from a single centre and
may not generalize to other ICUs in other healthcare systems. Mortality remains a
Page 18
Joint Proceedings - AIH 2013 / CARE 2013
crude measure of ICU performance but remains simple to measure and of great rele-
vance nevertheless. The existing gold standard models usually measure classification
of survival or death at hospital discharge, so are not necessarily directly comparable
to our models which measures survival or death at ICU discharge.
We are unable to directly compare our models with what may be considered gold
standards as some of these (e.g. APACHE IV) are only commercially available, and
as mentioned before, even the details of APACHE-III-J are not in the public domain.
The best comparison involving Australasian data using APACHE-III-J comes from
the paper of Paul et al. [4] but as with all APACHE models, this predicts death at
hospital discharge. Additionally, re-admissions were excluded which may be a sig-
nificant factor beyond what are often relatively small numbers of re-admissions in any
given ICU, as re-admissions suffer a disproportionately high mortality.
Exploration of the available hyperparameters of the many models examined has
been relatively limited. The ability to do this automatically, and explicitly or in a re-
producible way in WEKA and indeed any available software is limited although this
may be changing [18]. Yet minor changes to these hyperparameters may produce
meaningful enhancements in model performance. Tuning hyperparameters runs the
risk of overfitting a model, but we have tried to guard against this by testing the data
on a separate validation set.
Likewise, the ability to combine models with the best characteristics [19], which is
becoming more common in prediction of continuous variables [20] is not yet easily
performed with the available software.
We have not examined the calibration of our models. Good calibration is not re-
quired for accurate classification. Accurate performance across all risk categories is
highly desirable in a model. Similarly, performance including calibration for different
diagnostic categories that may become more significant in an ICU’s case mix is not
accounted for.
Modelling using imputed data in every phase of our study tends to show inconsis-
tent or suboptimal performance. It may be that imputation could be applied more
accurately by another approach that would improve model performance.
The major current use of these scores is in quality improvement activities. Once a
score is developed which accurately quantitates risk, the expected number of deaths
may be compared to those observed [21]. The exact risk for a given integer valued
number of deaths may be derived from the Poisson binomial distribution and com-
pared to the number observed [22]. A variety of risk adjusted control charts can be
constructed with confidence intervals [23].
5 Conclusions
We have presented alternative approaches to the classification problem involving
prediction of mortality at ICU discharge using machine learning techniques. Such
techniques may hold substantial advantage over traditional logistic regression ap-
proaches and should be considered to replace these. Complete clinical data may be
unnecessary when using machine learning techniques, and in any case are frequently
Page 19
Joint Proceedings - AIH 2013 / CARE 2013
not available. Out of the techniques studied, random forests seems to be the model-
ling approach with the best performance and has an advantage that it is relatively easy
to conceptualise and implement with open source software. During model training a
method to address class imbalance should be used.
6 Bibliography
[1]. Downar, J. (2013, April 18). Even without our biases, the outlook for prognos-
tication is grim. Available from ccforum:
http://ccforum.com/content/13/4/168
[2]. Knaus WA, W. D. (1981). APACHE-acute physiology and chronic health
evaluation: a physiologically based classification system. Crit Care Med, 591-597.
[3]. Tucker, J. (2002). Patient volume, staffing, and workload in relation to risk-
adjusted outcomes in a random stratified sample of UK neonatal intensive care units:
a prospective evaluation. Lancet, 99-107.
[4]. Paul, E., Bailey, M., Van Lint, A., & Pilcher, D. (2012). Performance of
APACHE III over time in Australia and New Zealand: a retrospective cohort study.
Anaesthesia and Intensive Care, 980-994.
[5]. Domingos, P. (2013, May 6). A few useful things to know about machine
learning. Available from Washington University:
http://homes.cs.washington.edu/~pedrod/papers/cacm12.pdf
[6]. R Core Team. (2013, April 25). Available from CRAN: http://www.R-
project.org/.
[7]. Honaker, J., King, G., & Blackwell, M. (2013, April 25). Amelia II: a program
for missing data. Available from Journal of Statistical Software:
http://www.jstatsoft.org/v45/i07/.
[8]. Hall, M., Eibe, F., Holmes, G., Pfahringer, B., & Reutemann, P. (2009, 1). The
WEKA Data Mining Software: An Update. SIGKDD Explorations.
[9]. Weka overview. (2013, April 25). Available from Sourceforge:
http://weka.sourceforge.net/doc/
[10]. Chawla, N. O., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002).
SMOTE: Synthetic Minority Over-sampling Technique. Journal of Ariticial Intelli-
gence Research, 321-357.
[11]. Ling, C. X., & Sheng, V. S. (2008). Cost-sensitive learning and the class im-
balance problem. In C. Sammat; G. Webb, editors. Encyclopaedia of Machine Learn-
ing. Springer.p.231-235.
[12]. Hanley, J., & McNeil, B. (1982). The meaning and use of the area under a re-
ceiver operating characteristic (ROC) curve. Radiology, 29-36.
[13]. Aksoy, S., & Haralick, R. M. (2013, May 20). Feature Normalization and Li-
kelihood-based Similarity Measures for Image Retrieval. Available from
cs.bilkent.edu: http://www.cs.bilkent.edu.tr/~saksoy/papers/prletters01_likelihood.pdf
[14]. Breiman, L. (2001). Random Forests. Machine Learning, 5-32.
[15]. Breiman, L. (1996). Bagging predictors. Machine Learning, 123-140.
Page 20
Joint Proceedings - AIH 2013 / CARE 2013
[16]. Freund, Y., & Schapire, R. E. (1996). Experiments with a new boosting algo-
rithm.In Machine Learning:Proceedings of the Thirteenth International Conference on
Machine Learning, (pp. 148-156). San Francisco.
[17]. Wolpert, D. (1996). The lack of a priori distinctions between learning algo-
rithms. Neural computation, 1341-1390.
[18]. Thornton, C., Hutter, F., Hoos, H., & Leyton-Brown, K. (2013, April 21).
Auto-WEKA: Combined selection and hyperparameter optimisation of classification
algorithms. Available from arxiv.org:
http://arxiv.org/pdf/1208.3719.pdf
[19]. Carauna, R., Nikilescu-Mizil, A., Crew, G., & Ksikes, A. (2013, May
20).[Internet]Ensemble selection from libraries of models. Available from
cs.cornell.edu:
http://www.cs.cornell.edu/~caruana/ctp/ct.papers/caruana.icm
l04.icdm06long.pdf
[20]. Meyer, Z. (2013, April 21). New package for ensembling R models [Internet].
Available from Modern Toolmaking:
http://moderntoolmaking.blogspot.co.nz/2013/03/new-package-
for-ensembling-r-models.html
[21]. Gallivan, S; (2003) How likely is it that a run of poor outcomes is unlike-
ly? European Journal of Operational Research , 150 46 - 52.
[22]. Hong, Y. (2013) On computing the distribution function for the Poisson bi-
nomial distribution. Computational Statistics and Data Analysis 59 41–51
[23]. Sherlaw-Johnson C. 2005 A method for detecting runs of good and bad clini-
cal outcomes on Variable Life-Adjusted Display (VLAD) charts. Health Care Manag
Sci. Feb;8(1):61-5.
Page 21
Joint Proceedings - AIH 2013 / CARE 2013
Towards a Visually Enhanced
Medical Search Engine
Lavish Lalwani1,2, Guido Zuccon1, Mohamed Sharaf2, Anthony Nguyen1
1
The Australian e-Health Research Centre, Brisbane, Queensland, Australia;
2
The University of Queensland, Brisbane, Queensland, Australia.
lavish.lalwani@uqconnect.edu.au, m.sharaf@uq.edu.au,
{guido.zuccon, anthony.nguyen}@csiro.au
Abstract. This paper presents the prototype of an information retrieval system for
medical records that utilises visualisation techniques, namely word clouds and
timelines. The system simplifies and assists information seeking tasks within the
medical domain. Access to patient medical information can be time consuming as
it requires practitioners to review a large number of electronic medical records to
find relevant information. Presenting a summary of the content of a medical
document by means of a word cloud may permit information seekers to decide
upon the relevance of a document to their information need in a simple and time-
effective manner. We extend this intuition, by mapping word clouds of electronic
medical records onto a timeline, to provide temporal information to the user. This
allows exploring word clouds in the context of a patient’s medical history. To
enhance the presentation of word clouds, we also provide the means for calculating
aggregations and differences between patient’s word clouds.
Keywords. Visualisation, Timeline, Word Cloud, Medical Search.
Introduction
Current information systems deployed in clinical settings require practitioners and
information seekers to review all medical records for a patient or enter database-like
queries in order to retrieve patient information. Clinical data is often organised
primarily by data source, without supporting the cognitive information seeking
processes of clinicians and other possible users. For example, “The Viewer”
application deployed by Queensland Health allows clinicians to access all patient
electronic medical records collected by Queensland Health hospitals and facilities1. To
access this information, clinicians need to enter data that allows them to select a patient
(e.g., name, date of birth, Medicare number, etc.); afterwards they are given access to
all information collected for that patient. However, they are unable to search through
the medical records of the selected patient: if clinicians require a patient’s past medical
history, they have to read all medical records for that patient (organised by type of data,
e.g. discharge notes, laboratory reports, etc., and clinical facility). This can be a very
time consuming and tedious way of accessing information, particularly when clinicians
1
Electronic medical record viewer solution, http://www.health.qld.gov.au/ehealth/theviewer.asp
Page 22
Joint Proceedings - AIH 2013 / CARE 2013
want to review a large number of cases for research purposes, e.g. observe the effect a
treatment had on their patient population.
An alternative solution is to deploy an information retrieval system where searches
over patient records can be conducted with keywords, and medical records are ranked
against the user query. We argue that this is a more efficient way for accessing patient
information; previous research has developed systems that are able to search for
relevant information in medical records [1, 2]. This paper considers how these systems
could be improved by enhancing the presentation of results retrieved in answer to
information seekers’ queries. Search results are commonly shown to users as textual
snippets that attempt to capture relevant portions of the medical record. Since these
snippets are small chunks of text extracted from the original document (extractive
summarisation), they often lack important information or can be misleading, especially
if the original document is a medical record [3]. In addition, textual snippets do not
convey an overview of the general clinical picture of a patient. For this reason, it is
difficult to determine whether a medical case matches a search and whether it should be
explored further; this thus requires the information seeker to access and read much of
the document to determine its relevance to the query.
This paper investigates the use of data visualisation as a means for solving this
problem. Data visualisation has the potential to provide a meaningful overview of
medical reports, visits or even a patient’s life and therefore may assist searchers to
determine whether a medical document is relevant and worth further examination. Data
visualisation may provide a simpler approach to augment standard searching methods
for medical data. The remainder of the paper describes a system prototype that
implements two data visualisation techniques: word clouds and timelines.
1. Related Work
Word clouds provide a visual representation of the content of a document by displaying
words considered important in a document. Words are arranged to form a cloud of
words of different sizes. The size of a word within a cloud is used to represent the
importance of that word in the document; often, the importance of a word is computed
as a function of the frequency of that word within a document. Figure 1 shows
examples of word clouds.
In this paper we posit that word clouds have the ability to provide a better
summary of the information contained in a medical record than textual snippets. This is
supported by existing research on employing word clouds within information retrieval
systems. For example, Gottron used a technique akin to word clouds to present news
web pages [4]. In that study it was found that word clouds helped users to decide upon
the relevance of news articles to their search query. Kaptein and Marx used word
clouds to enhance information access to debate transcripts from the Dutch parliament
[5]; they found that word clouds provided an effective first impression of the content of
a debate.
Timelines are an additional data visualisation technique providing a map of events
over time. The visualisation of events on a timeline provides the user with information
related to which events occurred prior (and after) to an event of interest;. In our
scenario, medical records belonging to a patient represents an event. Visualising
medical records over a timeline allows for the possibility of mapping an entire patient’s
medical history within a unique visual representation. Previous research found that
Page 23
Joint Proceedings - AIH 2013 / CARE 2013
employing timelines for displaying patient medical records has the benefit of enabling
clinical audit, reduced clinical errors, and improved patient safety [6]. Bui et al. have
explored the use of timelines to give a problem-centric visualisation of medical reports,
where patient reports are organised around diseases and conditions and mapped to a
timeline [7].
Figure 1. Word clouds computed from a medical record.
2. Word Clouds and Timelines
As supported by the previous research already outlined, this paper posits that word
clouds and timelines can be effective visualisation techniques to provide quick
information access to clinical records. The clinical records used to develop the
prototype system were obtained from the TREC Medical Records Track corpus, a
collection of 100,866 medical record documents taken from U.S. hospitals. Note that
documents belonging to a single patient's admission were grouped together, obtaining a
total of 17,198 groups of records. Next, we present the algorithms used within the
system to generate word clouds and timelines.
2.1. Word Cloud Generation
The generation of a word cloud within our prototype system is a multi-step process.
The first step consists of removing tokens and words from the documents that
convey limited or no information (stop word removal). These may include symbols,
special characters, and words contained in a ‘stoplist’ (e.g. “the”, “a”, “when”, etc.).
This step is used to avoid displaying irrelevant or non-informational words within the
word clouds.
The second step involves stemming the text of the medical reports. Stemming
consists of reducing a word to its base form (stem). Stemming is applied to conflate
syntactical variations of the same word (e.g. plurals, gerund forms, past tense, etc.) into
a single token to represent the fact that they may have the same or similar meaning.
The third step consists of generating a probability distribution over the vocabulary
words w, in a document d, 𝑃(𝑤|𝑑). Since a word cloud cannot display all the words in
a document, this distribution is used to derive the list of words that will form the word
cloud and their final font size (step four). Language models are used to compute such
probability distributions. The probability of a word w in a document d is computed as a
Page 24
Joint Proceedings - AIH 2013 / CARE 2013
function of the occurrence of w in the medical records as the following equation
mathematically explains.
Pλ (w | d ) = (1− λ ) P (w | d ) + λP (w | C ) (1)
In Equation 1, 𝑃(𝑤|𝑑) is calculated as the ratio between the number of
occurrences of 𝑤 in 𝑑 and the total number of words in d (maximum likelihood
estimate). Similarly, 𝑃(𝑤|𝐶) is calculated as the ratio between the number of
occurrences of w in the whole corpus of medical reports C and the total number of
words in C. These probabilities are interpolated according to the parameter 𝜆, which
controls the importance of background information (i.e., P(w|C)) when determining the
importance of word 𝑤 in the context of document d. The use of both the maximum
likelihood estimate and the background language modelling are referred to as Jelinek-
Mercer smoothing; more details on language modelling can be found in [8].
The last step (fourth step) is the generation of the actual word cloud. Words in a
document are ranked in decreasing order of their probability 𝑃(𝑤|𝑑), and only the top
ranked words are selected to be included in the word cloud. The probabilities of the
selected words are mapped into font sizes, and the appropriately sized words are placed
in the word cloud for document d. Figure 1a shows an example of a word cloud
generated from a patient medical report.
2.2. Word Cloud Aggregation
Individual word clouds could be merged to visualise an entire patient hospital visit or
medical history as a unique word cloud. Two word clouds wc1 and wc2 are merged
according to the following equation:
P (w ) = P (w | wc1 ) P (wc1 ) + P (w | wc2 ) P (wc2 ) (2)
where P(w|wci) represents the probability2 of word w in word cloud wci, and P(wci) is
the probability associated to wci. Currently, we consider word clouds to be uniformly
distributed (thus P(wc1) = P(wc2)); however future developments may consider biasing
word clouds according to temporal relations or document types when merging. As
previously stated, Equation 2 can also be used to create a word cloud representing a
complete patient medical history by merging all the word clouds associated to their
medical records. Similarly, Equation 2 can be applied for merging word clouds
associated with reports belonging to different patients.
2.3. Word Cloud Differential
A differential word cloud is designed to highlight the differences between two word
clouds (i.e. between two documents). Since two word clouds are effectively two
probability distributions, their difference can be computed using the Kullback-Leibler
(KL) divergence. Equation 3 provides the means for computing the difference between
word clouds, given the source word clouds wc1 and wc2.
2
P(w|wci) is equivalent to P(w|d) if wci represents the word cloud for document d; however, note that
wci may have also be computed from the merging of other previously computed word clouds.
Page 25
Joint Proceedings - AIH 2013 / CARE 2013
P(wi | wc1 )
DKL (wc1 || wc2 ) = ∑ P(wi | wc1 )log (3)
i P(wi | wc2 )
The magnitude of the KL divergence can be thought of as the degree of difference
between the two word clouds. The value of KL divergence for each word can be used
to generate a word cloud that provides visual information about how the two original
word clouds differ. We refer to this type of word cloud as a differential word cloud
(between wc1 and wc2). In a differential word cloud, the sign of DKL for each word (i.e.
DKL(w,wc1||w,wc2) = P(w|wc1)log[P(w|wc1)/P(w|wc2)]) determines the colour the word
would be painted with. Words with positive DKL values are painted green and words
with a negative DKL values are painted red. In this case, if a word is painted green it
means it has a stronger presence (i.e. higher probability) in wc1. The degree to which
this presence is stronger is signified by the size of the word in the cloud (the bigger the
word, the stronger the difference in presence). The opposite applies for a red colour
word in the differential word cloud. Note that if the calculation was conducted with the
probabilities in reverse order, the colours on the differential word cloud will reverse.
An example of a differential word cloud is shown in Figure 1b.
2.4. Timeline Generation
The generation of timelines involves, for each medical report, extracting the date and
time it was created. This was achieved using metadata information present in the
reports from the TREC Medical Records Track corpus; however, it is acceptable to
assume that similar metadata is present in records from other hospital providers. Since
entire patient admissions were mapped to timelines, after dates and times are extracted
for all records in a patient admission, this metadata, along with the medical record data
are rendered within a timeline created using the Java Script library, Timeline JS3. This
means that when retrieving a particular medical record, it can be displayed within
context of the other reports produced for that patient admission.
3. Integration of Word Clouds and Timelines
The prototype described here is a modular information retrieval system, developed
based on the Apache Lucene 4 framework, specifically for searching archives of
medical records. Its architecture consists of three main modules: the indexer, the
visualiser, and the searcher.
Within the indexer module, medical records are parsed and stored within a
representation appropriate for supporting the retrieval stage (inverted file). The indexer
is built using the Apache Lucene 4.0 incremental indexing capabilities, thus allowing
new documents to be included in the index without re-indexing the previous documents.
The indexer also maintains the relation between medical records and patients.
The searcher module is responsible for retrieving documents from the index that
match a user query. A ranked list of medical admissions is produced as the result of
querying the system.
3
http://timeline.verite.co/
Page 26
Joint Proceedings - AIH 2013 / CARE 2013
The visualiser module has the responsibility of rendering the results of a search
and supporting navigation across search results. The modular architecture of the system
integrates the visualisation methods described in Section 2 within the visualiser module
without modifying the approaches used to index and retrieve documents. Indeed, the
visualiser module is independent of the processes used in the other modules, allowing
for flexibility when devising and testing new visualisation algorithms, as well as
deploying versions of the system tailored to specific scenarios. Figure 2 shows a
screenshot of an implementation of the methods described in Section 2 within the
prototype system visualiser module. The figure illustrated a situation where a user has
submitted a query and is in the process of examining a specific medical record. The
content of the record is rendered as a word cloud allowing the user to quickly
understand the content of the record itself. The text of the recordAnonymized*for*Review*
can be accessed
through the “Reports view” button above the word cloud. The record is also placed
within the timeline of the patient admission to the hospital (bottom of Figure 2).
Figure 2. A screenshot of the visual interface of the system showing the use of word clouds and timelines.
4. Conclusion
In this paper we have presented two techniques, word clouds and timelines, to enhance
search results presentation within medical records search. Word clouds have the
potential to provide a rapid overview of an entire medical report, admission and patient
history. Timelines provide a visual means to represent patient journeys as well as to
place a medical record within the temporal context of other existing records. These
techniques were integrated within the visualiser module of our prototype, a state-of-
the-art medical information retrieval system. Future work will be directed towards a
formal evaluation of the proposed techniques in a real scenario. Possible improvements
will consider n-grams (sequences of n words, e.g. ‘heart attack’) and medical concept
detection and reasoning (e.g. “heart attack” and “myocardial infarction” within a record
should contribute towards the same medical concept) when building and rendering
word clouds.
Page 27
Joint Proceedings - AIH 2013 / CARE 2013
References
[1] Voorhees, E., & Tong, R. Overview of the TREC 2011 Medical Records Track. In Proceedings of TREC
(2011).
[2] Zuccon, G., Koopman, B., Nguyen, A., Vickers, D., & Butt, L. Exploiting Medical Hierarchies for
Concept-Based Information Retrieval. In Proceedings of ADCS (2012), 111-114.
[3] S. Afantenos, V. Karkaletsis & P. Stamatopoulos, Summarization from Medical Documents: a survey,
Artificial Intelligence in Medicine 33 (2005), 157-177.
[4] T. Gottron, Document Word Clouds: Visualising Web Documents as Tag Clouds to Aid Users in
Relevance Decisions, Lecture Notes in Computer Science 5714 (2009), 94-105.
[5] Kaptein, Rianne, and Maarten Marx. Focused Retrieval and Result Aggregation with Political
Data. Information retrieval 13.5 (2010): 412-433.
[6] Gill, J., Chearman, T., Carey, M., Nijjer, S., & Cross, F. Presenting Patient Data in the Electronic Care
Record: the role of timelines. JRSM short reports, 1(4), (2010).
[7] Bui, A. A., Aberle, D. R., & Kangarloo, H. TimeLine: visualizing integrated patient records. Information
Technology in Biomedicine, IEEE Transactions on, 11(4), (2007), 462-473.
[8] Zhai, C. Statistical Language Models for Information Retrieval. Synthesis Lectures on Human Language
Technologies, 1(1), (2008), 1-141.
Page 28
Joint Proceedings - AIH 2013 / CARE 2013
Using Fuzzy Logic for Decision Support in Vital Signs Monitoring
Shohas Dutta, Anthony Maeder, Jim Basilakis
School of Computing, Engineering & Mathematics, Telehealth Research & Innovation Laboratory
University of Western Sydney
Private Bag 1797, Penrith 2751, NSW
shohas6@gmail.com, a.maeder@uws.edu.au, j.basilakis@uws.edu.au
Abstract The overall aim of this research was to utilise information
This research investigated whether a fuzzy logic rule- gathered from personal vital signs monitoring in a
based decision support system could be used to detect laboratory-based smart home environment, and to assist
potentially abnormal health conditions, by processing with clinical care decisions using a fuzzy logic rule-based
physiological data collected from vital signs monitoring clinical decision support system. Fuzzy logic has benefits
devices. An application of the system to predict postural over other algorithmic approaches, as it has the potential
status of a person was demonstrated using real data, to to incorporate values from ordinal, nominal and
mimic the effects of body position changes while doing continuous datasets within its rules, and can capture the
certain normal daily activities. The results gathered in this knowledge associated with these rules in ways that are
experiment achieved accuracies of >85%. Applying this more intuitive to humans.
type of fuzzy logic approach, a decision system could be
constructed to inform necessary actions by caregivers or
for a person themself to make simple care decisions to 2 Vital Signs Monitoring Concepts
manage their health situation. .
Keywords: fuzzy logic, patient monitoring, decision There are numerous examples in literature describing
support, assistive technologies, care management. how monitoring of basic vital signs (i.e. heart rate, blood
pressure, temperature and respiration rate) can play a key
role in health care, e.g. Norris (2006) [39]. This approach
1 Introduction requires software to discover patterns and irregularities as
well as to make predictions. By collecting and analysing
vital signs continuously it can be shown how well the
Current trends in health within our society include the vital organs of the body are working, e.g. heart and lungs
move towards an ageing population profile, and increased (Harries et al. 2009) [40].
needs for complex care management for people with
chronic diseases and multiple co-morbidities. These are Lockwood et al. (2004) [30] provided a review of the
fast growing segments of the population; and so is the clinical usage of vital signs, including monitoring
need for covering their broad ranging and diverse care purpose, limitations, frequency and importance of vital
requirements. External support to manage high-risk (or signs measurements. They suggested that vital signs
unsafe) health situations is often needed for them to monitoring should become a routine procedure in chronic
continue their everyday living routines. This support is disease patients’ care. Bentzen (2009) [43] defined
typically given by both professional and informal chronic diseases as:
caregivers. “diseases which are long in duration, having long term
Due to technological advances in wireless data clinical course with no definite cure, gradually change
communication systems in the last decade, the application over time, and having asynchronous evolution and
of wireless-based vital sign monitoring devices for patient heterogeneity in population susceptibility.”
monitoring has gained increasing attention in the clinical Living with a chronic disease, which increases in severity
arena. Patient health status can be determined based on with age, has a significant impact on a person’s quality of
the acquisition of basic physiological vital signs, life and on their family. Chronic disease patients would
suggesting that a system providing wireless monitoring of be able to play a more active role in managing their own
vital signs has potential benefits for clinical care health by taking vital signs measurements daily and
management of independently living patients as well as participating in meaningful electronic information
their carers. A patient’s physiological state, which exchanges with clinicians.
includes heart rate, blood pressure, body temperature etc., A number of authors have suggested that using smart
can be monitored continuously using wearable medical homes for health monitoring is a promising area for
body sensor devices. The remaining challenge is to gain health care. Chan et al. (2009) [2] in their review paper
sufficient understanding of this data to assist in health described the smart home as a promising and cost-
care needs. effective way to improve home care for elderly people
and people suffering with different chronic diseases.
Vincent et al. (2002) [19] identified three research areas,
which combined to produce the concept of “health smart
home”. These three areas are medicine, information
systems, and home based automatic and remote control
Page 29
Joint Proceedings - AIH 2013 / CARE 2013
devices. A smart home contributes to monitoring of the
patient’s health status continuously, taking into
consideration the patient’s personal needs and wishes in Rule Base
(Inference):
addition to their specific medical requirements. The Fuzzifi- Defuzzifi-
cation cation
information gathered through health status monitoring Aggregate
Activate
systems can feed into an access controlled electronic Accumulate
patient records system for further medical interpretation.
LoPresti et al. (2008) [21] identified different assistive
technologies which can be used in smart homes to reduce
the effect of disabilities and improve quality of life. Figure 1. Elements and structure of fuzzy control.
Wearable and portable devices are used which help to
monitor the vital signs or physiological behaviour of a Y
person living in a smart home. Those devices are worn by
LOW NORMAL HIGH
the user or embedded in the smart home. They are wired
1
or wirelessly connected to a monitoring centre. Recently,
robotic technology has been developed to support basic
activities and mobility for elderly people too.
X1 X2 X3 X4 X5 X6 T
3 Fuzzy Logic Concepts
Figure 2. Fuzzy membership functions of variable T.
Fuzzy logic (Zadeh 1990) [68] is a well established
computational method for implementing rules in 4 Experimental Methodology
imprecise settings, where some adaptability for
prescribing the rules is necessary. A fuzzy system can be
used to match any set of input-output combinations. This section will discuss the design of a laboratory
Fuzzy logic can provide us with a simple way to draw experiment to undertake validation of the approach, using
definite results from vague, ambiguous or imprecise a longitudinal data set of physiological signals which
information. The rule inference system of the fuzzy have been gathered from an experiment involving
model (Jang 1993) [67] consists of a number of monitoring of blood pressure and heart rate signals. It is
conditional IF-THEN rules. For the designer who well known that changes to these vital signs will occur if
understands the system, these rules are easy to write, and the body position is changed from vertical to horizontal.
as many rules as are necessary can be supplied to describe The nature and rapidity of these changes mimics the
the system adequately. changes in vital signs that may occur with onset of some
exacerbated or acute health status in patients.
To improve clinician performance, fuzzy logic-based
expert systems have shown potential for imitating human The laboratory setup used a tilt table to generate changes
thought processes in the complex circumstances of in heart rate and blood pressure measurements that were
clinical decision support (Pandey 2009) [75]. A key correlated with the angle of the tilt table (Figure 3). These
advantage of using fuzzy logic in such situations is that physiological changes would be similar to changes one
the fuzzy rules can be programmed easily, and as a result would expect in circumstances such as changing health
they are easily understood by clinicians. It is different status or other physiological stressors such as an infection
from neural networks and other regression approaches, or blood loss. The result of the fuzzy logic analysis of
where the system behaves more like a black box to such data can be used to detect a change in physiological
clinicians. Schuh (2008) [73] found that fuzzy logic state occurring when the vital signs measures are either
holds great promise for increasing efficiency and increasing or decreasing, compared to a steady state
reliability in health care delivery situations requiring where there are no longitudinal changes in the vital sign
decisions based on vital signs information. This has also measures. This output can be compared against the angle
been observed in specialised situations such as intensive of the tilt table, that will serve as a gold standard for
care (Cicilia et al 2011) [81]. determining whether the system is in a steady state or not.
Fuzzy control is the core computational component of a
fuzzy logic system. It includes the processing of the
measured input values based on the fuzzy rules, and their
conversion into decisions with the help of fuzzy
combination logic. A full description of fuzzy control
principles is beyond the scope of this paper and can be
found in numerous fuzzy logic texts. The functional
elements of fuzzy control can be represented in a block Figure 3. Movement range of tilt table.
diagram in Figure 1, based on fuzzy membership
functions of variables of interest, as shown in Figure 2 for
The tilt table used was a motorized table with a metal
the example of body temperature represented by the
footboard. The subject’s feet were rested on the
variable T.
footboard. Soft Velcro straps were placed across the body
Page 30
Joint Proceedings - AIH 2013 / CARE 2013
for safety reasons, to secure the person when the table 5 Experimental Results
was tilted during the test. When using the tilt table, it was
always tilted upright so that the head of the subject was
above his feet. Small, sticky patches containing The fuzzy logic rules were derived using the blood
electrodes were placed on the subject’s chest. These pressure and heart rate signals from the first of the three
electrodes were connected to an electrocardiograph cycles. These signals were pre-processed to find a
monitor (ECG) to record the electrical activity of the smoothed curve of the recorded raw signals. In this
person heart to be shown as an ECG graph. The ECG smoothing process, the averages of the values of heart
showed the heart rate and rhythm during the test, at a raw rate and blood pressure were calculated for every five
sampling rate of 100Hz and an accuracy of 3%. A blood timestamps using non-overlapping windows. Then these
pressure measuring device was also attached on the average values were used to plot a smooth curve of the
subject’s finger. This was connected to monitors so that systolic blood pressure and peak-to-peak heart rate to
the blood pressure could be observed during the test as establish the trends. Figure 5 shows the training dataset.
well as being recorded.
At the very beginning of the test, the subject was laid flat
on his back on the tilt table. At that time his initial blood
pressure, ECG, and his position angle data were recorded.
After resting for few minutes, the test was started. The
blood pressure and ECG was constantly monitored
throughout the test and instantaneous readings of the data
stream were recorded every second for subsequent
analysis. The following protocol was applied for
changing the positioning of the tilt table:
1. Lying flat at rest for ~60 sec (to gain statistics of resting state)
2. Fast tilt upwards over ~10 sec
3. Very slow tilt downwards over ~30 sec
4. Lying flat resting state ~30 sec Figure 5. Training dataset (top to bottom): blood
5. Medium tilt upwards over ~20 sec pressure, heart rate, tilt angle.
6. Upright resting state ~30 sec
7. Fast tilt downwards over ~10 sec
The fuzzy logic solution has two input variables and one
8. Lying flat resting state ~30 sec
output variable. Using the mean and standard deviation
9. Fast tilt upwards over ~10 sec as a tolerance band for the input variables, three states
10. Upright resting state ~30 sec (Low, Normal, High) are defined. The two input variables
11. Medium tilt downwards over ~20 sec are combined by the AND (i.e. MAX) operator and valid
12. Lying flat resting state ~30 sec states inferred from the values for the tilt angle, as
represented in the decision matrix shown in Table 1.
A sample data set collected recorded using the above
protocol is shown in the graphs in Figure 4. Data sets Table 1. The decision matrix for the training data.
from three repetitions of the protocol were captured using
one of the investigators as the subject, as a pre-ethics Input Variable 1: Systolic Blood
proof-of-concept exercise needed to justify a full human Pressure
research ethics application for extending the work for
recruited subjects in the future. Little variability was Low Normal High
observed in the three data sets, so it was considered
unnecessary to collect further test data. Low Static
Input
Normal Static Static Lowering
Variable
2: Heart
Rate
High Lowering Raising
The following rules based on this table were derived:
RULE 1: IF systolic IS low AND heart_rate IS low THEN
physiological_status IS Unclassified;
RULE 2: IF systolic IS low AND heart_rate IS normal THEN
Figure 4. Data captured from the experiment: (top to physiological_status IS Static;
bottom): angle, footplate force, ECG, blood pressure. RULE 3: IF systolic IS low AND heart_rate IS high THEN
physiological_status IS Unclassified;
Page 31
Joint Proceedings - AIH 2013 / CARE 2013
RULE 4: IF systolic IS normal AND heart_rate IS low THEN Table 2. Matching actual states and predicted states.
physiological_status IS Unclassified;
RULE 5: IF systolic IS normal AND heart_rate IS normal THEN
physiological_status IS Static;
Predicted State (Computed)
RULE 6: IF systolic IS normal AND heart_rate IS high THEN
physiological_status IS Lowering;
Static Raising Lowering Total
RULE 7: IF systolic IS high AND heart_rate IS low THEN
physiological_status IS Static; Actual
State Static 24 1 2 27
RULE 8: IF systolic IS high AND heart_rate IS normal THEN
physiological_status IS Lowering; (Gold
RULE 9: IF systolic IS high AND heart_rate IS high THEN standard) Raising 2 2 2 6
physiological_status IS Raising;
The derived fuzzy rules were applied to the smoothed Lowering 1 0 5 6
data of the test set for the second and third cycles, to
determine the physiological status. By applying fuzzy Total 27 3 9 39
logic to these two cycles of testing data, different regions
in the data were classified into predicted statuses of
Table 3. Classifier positive and negative outcomes.
Static, Raising and Lowering. Figure 6 shows the results
with yellow indicating static status, grey indicating
lowering status and green indicating raising status.
Test Outcome (Static case)
Gold Standard True Positive (24) False Positive (3)
Set (Static case)
False Negative (3) True Negative (9)
Test Outcome (Raising case)
Gold Standard True Positive (2) False Positive (4)
Set (Raising case)
False Negative (1) True Negative (32)
Test Outcome (Lowering case)
Figure 6. Classifying status using the trained rules.
Gold Standard True Positive (5) False Positive (1)
Set (Lowering case)
In order to compare the fuzzy logic output to the gold
False Negative (4) True Negative (29)
standard, statuses needed to be inferred from the angle of
the tilt table. The following protocol was established to
determine three different states categorised as: Static,
Raising and Lowering. Only changes of one or more The resulting indicator values were calculated as follows:
smoothing period timesteps (i.ee >4 sec) were considered.
The protocol used was as follows:
Sensitivity (Static) = 24 / (24+3) = 24 / 27 = 0.89
1. If the change of angle is < 5° and timestamp interval >4 sec, Specificity (Static) = 9 / (9+3) = 9 / 12 = 0.75
then the tilting table is in static state. Sensitivity (Raising) = 2 / (2+1) = 2 / 3 = 0.67
2. If the change of angle (upward) is: 25°< angle<90° and Specificity (Raising) = 32 / (32+4) = 32 / 36 = 0.89
timestamp interval >4 sec, then the tilting table is in Sensitivity (Lowering) = 5 / (5+4) = 5 / 9 = 0.56
abnormal state and in the raising state. Specificity (Lowering) = 29 / (29+1) = 29 / 30 = 0.97
3. If the change of angle (downward) is: 25°< angle<90° and
Accuracy (Static) = (24+9) / 39 = 33 / 39 = 0.85
timestamp interval >4 sec, then the tilting table is in the
Error (Static) = (3+3) / 39 = 6 / 39 = 0.15
lowering state.
Accuracy (Raising) = (2+32) / 39 = 34 / 39 = 0.87
Error (Raising) = (4+1) / 39 = 5 / 39 = 0.13
The results using these steps are summarised in Table 2,
Accuracy (Lowering) = (5+29) / 39 = 34 / 39 = 0.87
and the overall rate of positive and negative outcomes is
Error (Lowering) = (1+4) / 39 = 5 / 39 = 0.13
shown in Table 3. These outcomes were used to analyse
classifier performance using the following indicators:
Across the three states, Sensitivity values ranged from
Sensitivity = TP/(TP+FN) = Prob(+ve test) 0.56 to 0.89, and Specificity values ranged from 0.75 to
Specificity = TN/(TN+FP) = Prob(-ve test) 0.97. The low Sensitivity values are related to the smaller
Accuracy = (TP+TN)/total obs = Prob(correct)
sample sizes for the Raising and Lowering states.
Accuracy rates ranged from 0.85 to 0.87, and Error rates
Error = (FP+FN)/total obs = Prob(wrong)
ranged from 0.13 to 0.15, indicating good performance.
Page 32
Joint Proceedings - AIH 2013 / CARE 2013
In considering the performance of this approach, several provides a robust implementation environment and a
drawbacks affected the achievable accuracy negatively. sufficiently simple rule specification mechanism to allow
The first issue was the time lag in the dropping of the users who are not IT experts to reconfigure the system to
vital sign values when changing the angle of the tilting suit a given vital signs classification problem.
table. While the tilting table was moved rapidly, it took A worthwhile extension of this work would be to improve
several seconds for the physiological status of the human the level of sophistication and automation of the threshold
body to adapt accordingly. As a result, this problem has values for the fuzzy logic classification process. Instead
affected accuracy in determining the physiological status of a simple statistical approach using a set of “normal”
of a person in FastUp or in FastDown status. observations, actual patterns could be captured and stored
Another problem was related to the error rate associated which could be tested with greater severity than smooth
with using the vital signs measurement equipment. When fuzzy functions. The work offers scope to increase the
the position of the tilt table was changed, small amount of ambient intelligence which could be provided
movements of the body affected accurate measuring of in the “smart home” of the future, to help sustain
the physiological data by the monitoring devices. For occupants’ health circumstances.
example, the blood pressure measuring device was
attached with the finger and due to the movement of the
body and fingers it sometimes gave erroneous readings. 7 References
The smoothing function that was applied was intended to
damp out such errors but there is some residual effect.
Chan M, Campo E, Estève D, Fourniols JY., Smart homes
— Current features and future perspectives, Maturitas,
6 Conclusion and Future Work 2009, 64(2), p: 90-97.
Vincent R, Norbert N, Lionel B, Jacques D., Health
"Smart" home: information technology for patients at
We have described an efficient computational approach
home. Telemedicine Journal and eHealth, 2002. 8(4), p:
to the problem of personal monitoring of vital signs, to
395-409.
provide alerts under well defined abnormal health status
conditions which are caused by a known or anticipated LoPresti EF, Bodine C, Lewis C., Assistive technology for
health situation. The purpose of such alerts is to provide cognition Understanding the Needs of Persons with
decision support inputs to carers, to prompt closer Disabilities. Engineering in Medicine and Biology
observations or direct interventions to be performed to Magazine, IEEE, 2008. 27(2), p: 29-39.
help the subjects of care. This could be useful over a wide Lockwood C, Tiffany CH, Page T., Vital Signs. JBI
range of situations such as elderly or disabled living Reports, Wiley Online Library, 2004. 2(6) p: 207-230.
alone, or patients with chronic diseases or multiple co- Norris PR., Toward New Vital Signs: Tools and Methods
morbidities. for Physiological Data Capture, Analysis, and Decision
Fuzzy logic was chosen as an appropriate computational Support in Critical Care, PhD Thesis, Graduate School of
approach due to its simplicity and ease of tuning to suit Vanderbilt University, 2006.
relatively smoothly changing vital signs values. Then the Harries AD, Rony Z, Kapur A, Jahn A, Enarson DA., The
approach was implemented in software, providing a Vital Signs of Chronic Disease Management.
multistage process for classifying the condition of a Transactions of The Royal Society of Tropical Medicine
subject using fuzzy functions for each of several observed and Hygiene, 2009. 103(6), p: 537-540.
vital signs, and then combining these using rules to
determine the overall health status. Bentzen N., WONCA dictionary of general/family
practice, Mental Health in Family Medicine, 2009, 6(1),
Using this approach, a fuzzy logic rule-based decision p: 57–59.
support system could, for example, be used to monitor
daily activities of living and detection of falls for smart Jang JSR., ANFIS: Adaptive-Network based Fuzzy
home residents, in combination with other technologies Inference Systems, IEEE Transactions on Systems, Man,
that have more sensitivity in detecting sudden change of and Cybernetics, 1993, 23(3), p: 665-685.
body posture such as tri-axial accelerometers. Further Zadeh LA., The birth and evolution of fuzzy logic,
research is required to find out the usefulness of such a International Journal of General Systems, 1990, 17(2-3),
fuzzy logic rule-based decision support system when a p: 95-105.
combination of vital signs and acceleration data is used to Schuh CJ., Monitoring the fuzziness of human vital
detect sudden changes in body posture. Parameters, Annual Meeting of the North American
On the basis of this foundation work, fuzzy logic has Fuzzy Information Processing Society, IEEE 2008, p:1-6.
been shown to provide a plausible approach to the general Pandey B, Mishra R., Knowledge and intelligent
problem of classifying health status in situations of computing system in medicine. Computeras in Biology
abnormalities in vital signs patterns. It is anticipated that and Medicine, 2009, 39(3), p: 215-30.
a more extensive system could be built by including Cicília RM L, Glaucia RS, Adrião DDN, Ricardo AMV,
further parameters and more complex rules, using the Ana MG G., A fuzzy model for processing and
same fundamental algorithm. The implementation monitoring vital signs in ICU patients, BioMedical
methodology using an SQL database and fixed form Engineering OnLine, 2011, 10(68).
parameter labelling functions for the fuzzy assignments,
Page 33
Joint Proceedings - AIH 2013 / CARE 2013
A Novel Approach for Improving Chronic Disease
Outcomes using Intelligent Personal Health Records in a
Collaborative Care Frameworkolika
Amol Wagholikar1
1: The Australian e-Health Research Centre, CSIRO Computational Informatics
Abstract.
Background- Effective management of chronic diseases is highly important to
improve health outcomes of chronic disease patients. Emerging initiatives on
online personal health records (PHR) have provided an opportunity to empower
patients living with a chronic disease to take control of their own health data
management. Online PHR solutions also provide data-driven intelligent analyt-
ics capability that can provide an effective view of the patient’s health data to
the patients themselves as well as to their consented clinicians and carers such
as family members fully engaged in their routine care. Research suggests a ten-
dency among chronic disease patients of using self-managed care without as
well as with some support to monitor and manage their chronic disease. The ris-
ing usage of online solutions enable chronic disease patients store their physical
as well as mental health information in self-managed online PHR. There are a
variety of such online PHR mechanisms that are available via desktop comput-
ers, mobile smartphones, smart TVs as well as biometric devices. However, the
main problem of disparate data sources and lack of a universal view of patient’s
health data still exists. These problems needs a novel way of integrating various
types of PHRs in an efficient way and provide effective insights about the pa-
tient’s health to empower and engage the patients in active management of their
chronic condition. Objective- To describe a framework to integrate various
online PHRs for providing effective self-managed and collaborative care.
Methods- Comprehensive research was conducted to analyse current trends of
various PHR mechanisms. A series of discussions were held with the clinical as
well as non-clinical end users of online PHRs to identify the current problems
with accessing PHRs and their expectations about usage of PHR in managing
chronic disease condition. The requirements analysis and emerging technology
trends were utilized to develop a framework that provides intelligent capabili-
ties for a collaborative online platform. Results- The requirements analysis and
discussions with the end-user representatives showed that the proposed frame-
work is considered novel and intuitive by the stakeholders thus confirming our
findings. Conclusion- The results of this investigation specified a novel frame-
work that can enhance the value of PHRs and thus may address usability chal-
lenges identified by the PHR developers as well as the end-users.
Keywords: Personal Health Record (PHR), Collaborative Care, Self-
managed care
Page 34
Joint Proceedings - AIH 2013 / CARE 2013
1 Introduction
A personal health record (PHR) is a record in a tangible document format (e.g. infor-
mation recorded on a piece of paper and/or in an electronic document); in which an
individual patient creates, maintains and controls his/her health related data [1]. The
patient may access, modify as well as control the individual health information before
using it for specific purposes such as self-assessment and sharing it with care provid-
ers through a consent process. The patients may also store a copy of data collected by
their clinicians in their personal health record. The patient’s PHR is a component of
complete set of the patient’s health related information as some information is also
created, stored and managed in hospital and clinic health information systems.
Personal Health Records exist in paper-based (offline) format as well as electronic
format (online). Some percentage of patients may regularly use offline PHR’s to store
and access their chronic disease specific health information. A certain section of pa-
tients may use various online PHR mechanisms (such as website-based tools and
mobile applications provided by private vendors) to manage their health related in-
formation. The online PHR is an electronic record of an individual’s personal health
information stored securely in a central repository that can be accessed by an individ-
ual patient for self-managed care, self-monitoring of health conditions and can be
shared with the clinicians for clinical use. The patients may choose to share the online
electronic health record with clinical information systems used by their care provider
to provide an accurate and a complete set of information required for providing point-
of-care health services at various geographic as well as clinical settings [2, 3, 4].
1.1 Current State-of-the-art
The online personal health record is an emerging discipline of research. The current
research in this discipline makes an attempt to improve value of the personal health
records through application of innovative intelligent data processing methods [5]. The
concept of Personal Health Record (PHR) has been evolved along with the advance-
ments in web-based technologies. The current state-of-the-art suggests that intelligent
PHRs are evolving to add more features such as data exchange, data sharing with
clinicians, and family members. There are certain proprietary PHR solutions as well
as open source PHR solutions. Each of the PHR solution offers common functions
and features that can be accessed by the patients through a web-enabled device with a
web browser. These solutions are evolving to add more features such as data ex-
change, data sharing with clinicians, families and carers. However, there are open
issues in the intelligent PR especially in the area of functions provided by these solu-
tions. Our review of the existing PHR solutions indicates that there is a lack of col-
laborative functions in the PHR [6]. This work attempts to address this gap by propos-
ing a collaborative PHR platform.
Page 35
Joint Proceedings - AIH 2013 / CARE 2013
1.2 Main Problem
There are certainly growing efforts both in private as well as public domains for
adopting online PHR as a data recording and analysis tool for self-managed care, self-
monitoring of disease conditions, as a preventative health intervention and clinical
use. There is growing number and various types of private online PHR solutions that
are available as a web-application and/or mobile application storing patient health
data. The growing number of mobile health applications for tracking and monitoring
exercise is a good example of this evolution (E.g. Nike+ app or the Fitbit app with
optional weight scale and wrist fitness band). The various types of PHR’s can be also
categorised as per the devices that can be used to access the PHR. The types of online
PHR are shown in table 1.
Table 1. Types of Online PHR
No Online PHR Type Access Devices
1 Web-based PHR Desktop Computer, Smart phone, tablet device
2 PHR in Mobile App Smart phone, tablet device
3 PHR in wireless Smart phone + portable device such as blood
monitoring device glucose monitor
The advances in web-based and mobile apps online technologies as well as rising use
of online data recording and analysis solutions has led to the development and launch
of private and public online PHRs. The broad categories of online PHR systems are
illustrated in Figure 1 which links with the relationship illustrated in Figure 1.
Private
Online PHR Systems pro-
vided by private vendors,
mobile health apps
Public
Chronic Disease Patient Online PHR systems pro-
vided by government such
as Australia’s PCEHR
EHR and/or EMR
Fig. 1. Types of Online PHR Systems
Page 36
Joint Proceedings - AIH 2013 / CARE 2013
Due to the widespread use of web-based PHR solutions, the patient’s health data is
stored in various disperse data sources. Thus, despite advances in online solutions, the
core issue of “Health Information SilOs” (HISO) still exists and thus the issue of ac-
curate information for self-managed care as personalised decision support is still
largely unresolved. We propose an approach in the form of a framework that attempts
to address this issue with a new design proposal for an intelligent PHR framework.
2 Methods
This research was undertaken for an initiative that aims to improve journey of chronic
disease survivors. The following main steps were undertaken for our research.
Expert Interviews: A group of clinical experts, patient representatives as well as
technology experts was engaged to understand the real-life issues of chronic dis-
ease survivors. A series of expert interviews were conducted to understand the
main drivers as well as requirements for developing an online technology approach
to leverage advances in PHR, established as well as emerging industry trends in
health information technology. A key challenge of providing a seamless experience
for the patients to manage their own health data was identified during these inter-
views. The challenges of adoption by the chronic disease patients with limited
health and information technology literacy were also identified. The inputs from
the interviews were used to specify requirements of our proposed platform. The
specifications aimed to propose an innovative design of a PHR-based solution us-
ing a vendor-provided PHR platform with customization.
Online Solution Investigation: A comprehensive research was conducted to in-
vestigate global landscape of emerging online PHR solutions including desktop as
well as smart devices (smart phones and tablets) based solutions. The results of the
comprehensive research are summarized in the table below-
Table 2: Online PHR status around the world
Country PHR Solution Roll Out PHR Standards Current Status
Australia PCEHR [7] July 2012 NEHTA Active Adoption
PCEHR in progress
UK NHS 2010 HL7 and others Closed in De-
Healthspace cember 2012
[8,9]
Canada Various Private Since 2009 Proprietary and Active Adoption
And Online open source in progress
US PHRs[10], Big
blue button[11],
blue button+ as a
public PHR in
US
Page 37
Joint Proceedings - AIH 2013 / CARE 2013
Proposed Approach Development: The investigation resulted into recommenda-
tion of our technology platform that can address the issues of HISO in an online
PHR context.
3 Proposed Approach for an Integrated PHR
Our investigation resulted into identifying key challenges and critical needs for
providing a single platform that can provide a holistic view of the patient’s PHR. We
propose a platform that can not only record critical data but also provide intelligent
analysis of patient’s health data to patients as well as their carers. A schematic repre-
sentation of our approach is shown in figure 2.
Fig. 2. Schematic Representation of Our Proposed Approach
One of the central components of our proposed platform is the intelligence engine.
The intelligence engine component will executive algorithms driven by the personal
health data stored in the PHR. The specific details of the algorithms will be reported
Page 38
Joint Proceedings - AIH 2013 / CARE 2013
as the research progresses. Our proposed platform also provides intelligent analytics
of the patient’s health data which improves patient’s own understanding of their
health data. The intelligent analytics also improves monitoring of key health indica-
tors such as weight, mood, and nutrition in a personalized visual dashboard.
Our proposed approach is in the form of a collaborative PHR platform addresses the
issue of disperse health data sources. It allows the patients as well as their care pro-
viders a universal view of patient’s health data governed by the consent process. The
clinicians also have the ability to view as well as add clinical notes to the patient’s
PHR. The patient’s can collaborate with their clinicians as well as similar patients
through the platform which can integrate with patient-driven social network. The
patients can connect with their carers through video, voice as well as text communica-
tion methods enabling different communication mediums. Our platform also im-
proves patient engagement as it enables collection of their physical health data
through wearable health monitoring devices that seamlessly integrate with our plat-
form. This integration with biometric devices improves efficiency in data recording.
3.1 Challenges
Our proposed approach has the potential to support care models that deliver better
health outcomes. However, the successful execution requires understanding of the
challenges involved in online PHR integration. The challenges are -
Adoption: The adoption rate for mobile health applications for health monitoring
and health tracking for self-management among healthy population is increasing
over the last decade. [12]. However, the Australian and worldwide research shows
that the adoption rate for online PHR based solution among adults with chronic
disease conditions is less than expected [13, 14, 15]. Our proposed solution aims to
improve the uptake rate by providing a simple and yet highly effective solution.
The challenges in adoption of online PHR mechanisms are not only applicable to
the patients but clinicians and carers as well.
Table 3. Adoption Challenges by online PHR users
Online PHR User Category Challenges
IT Literacy, Technology device
Chronic Disease Patients preferences, Ability to record and
understand own data (health literacy)
Page 39
Joint Proceedings - AIH 2013 / CARE 2013
Online PHR User Category Challenges
Families IT Literacy, Interpretation of data
Quality of concern Patient recorded
data for treatment decisions, Time to
Primary and Specialist Clinicians
access patient reported data in online
PHR
Data Quality, Time to access pa-
Nursing Staff
tient reported data in online PHR
Quality concern of Patient recorded
Allied Health clinicians
data, Lack of instant physical interac-
(E.g. Physiotherapists)
tion with men
Quality concern of Patient recorded
Care coordinators data for care plan development and
implementation
Data Quality: The data quality in the PHR should be endorsed by the clinicians.
Cost-effectiveness: The evidence suggesting online PHR as a cost-effective tool to
manage personal health information is not clearly established [16, 17, 18, and 19].
Health Outcomes: There is also no clear evidence about better health outcomes
due to online PHR [20].
The above challenges can be addressed through a careful implementation of the pro-
posed platform. The implementation of the proposed platform is under progress and
the evaluation of the platform will be undertaken in a randomised clinical trial set-
tings. The proposed approach has received positive feedback from the patient as well
as clinical community representatives.
4 Conclusion
This research has made an attempt to address the current issues of disperse personal
health data sources. The proposed platform aims to provide a single view of the pa-
tient’s PHR that can empower chronic disease patients and improve collaboration
with the clinicians for self-managed care. The proposed platform will be implemented
and evaluated as this research progresses in future.
Page 40
Joint Proceedings - AIH 2013 / CARE 2013
Limitations
The author acknowledges that the research is a work-in-progress. The proposed plat-
form is not evaluated with a sample data yet. The concepts proposed in this paper are
specified strategic perspective. The research does not provide real-life evaluation of
the proposed PHR platform as well as core design of the intelligent PHR.
Acknowledgement
This research is mainly supported by Movember Foundation as a part of its survivor-
ship initiative. This research is also supported by the Australian e-Health Research
Centre, CSIRO as a project collaborator. The author would like to thank Movember
Foundation and the Australian e-Health Research Centre, CSIRO for their support and
this opportunity.
References
1. Tang, Paul; Ash, Joan; Bates, David; Overhage, J.; Sands, Daniel (2006). "Personal Health
Records: Definitions, Benefits, and Strategies for Overcoming Barriers to Adoption".
JAMIA 13 (2): 121–126. doi:10.1197/jamia.M2025
2. National e-Health Transition Authority (NEHTA), http://www.nehta.gov.au/ehealth-
implementation/what-is-a-pcehr, Viewed 19 February 2013.
3. Markle Foundation Personal Health Working Group, Connecting for Health: A Public-
Private Collaborative. New York, NY Markle Foundation2003;
4. Yamin CK, Emani S, Williams DH, et al. The digital divide in adoption and use of a per-
sonal health record. Arch Intern Med 2011;171:568–74
5. Irini Genitsaridi, Haridimos Kondylakis, Lefteris Koumakis, Kostas Marias, Manolis Tsi-
knakis, Towards Intelligent Personal Health Record Systems: Review, Criteria and Exten-
sions, Procedia Computer Science, Volume 21, 2013, Pages 327-334.
6. Luo G. Open issues in intelligent personal health record - an updated status report for
2012. J Med Syst. 2013 Jun;37(3):9943.
7. National e-Health Transition Authority (NEHTA), http://www.nehta.gov.au/ehealth-
implementation/what-is-a-pcehr, Viewed 19 February 2013.
8. NHS Healthspace,
http://www.connectingforhealth.nhs.uk/systemsandservices/healthspace, Viewed 20 Feb-
ruary 2013.
9. NHS Information Strategy, http://informationstrategy.dh.gov.uk/, Viewed 20 February
2013.
10. US MYPHR, http://www.myphr.com/resources/choose.aspx, Viewed 21 February 2013
11. Big Blue Button, http://www.va.gov/BLUEBUTTON/index.asp, Viewed March 2013,
Viewed 21 February 2013.
12. Big Button Plus, http://bluebuttonplus.org/index.html, Viewed 2 April 2013.
13. NEHTA Compliant Product Registers, https://epipregister.nehta.gov.au/registers, Viewed
25 February 2013.
Page 41
Joint Proceedings - AIH 2013 / CARE 2013
14. Boulos MN, Wheeler S, Tavares C, Jones R. How smartphones are changing the face of
mobile and participatory healthcare: an overview, with example from eCAALYX. Biomed
Eng Online. 2011; 10:24.
15. Australian PCEHR Registration Targets, Pulse IT Magazine Article,
http://www.pulseitmagazine.com.au/index.php?option=com_content&view=article&id=13
17:half-a-million-pcehr-registrations-still-achievable-doha&catid=16:australian-
ehealth&Itemid=328\\, Viewed 22 February 2013
16. A. Sunyaev, D. Chomyi, C. Mauro, and H. Krcmar, "Evaluation Framework for Personal
Health Records: Microsoft HealthVault vs. Google Health," in Proceedings of the Hawaii
International Conference on System Sciences (HICSS 43), Kauai, Hawaii, 2010.
17. Executive Summary Integrated Cancer Care Programme, Report by UnitedHealth Europe.
18. Evaluation of Phase 1 of the One-to-One, Support Implementation Project, Baseline Re-
port Macmillan Cancer Support
19. Imison H, Gorman P, Woods S, et al. Barriers and Drivers of health information technolo-
gy use for the elderly, chronically ill, and underserved. Evid Rep/Technol Assess; 175
(Prepared by the Oregon Evidence-based Practice Center under Contract No. 290–02-
0024). AHRQ Publication No. 09-E004. Rockville, MD: Agency for Healthcare Research
and Quality. Nov 2008.
20. Kim Eung-Hun, Stolyar Anna, Lober William B, Herbaugh Anne L, Shinstrom Sally E,
Zierler Brenda K, Soh Cheong B, Kim Yongmin. Challenges to using an electronic per-
sonal health record by a low-income elderly population. J Med Internet Res. 2009;
11(4):e44.
Page 42
Joint Proceedings - AIH 2013 / CARE 2013
Partially automated literature screening for
systematic reviews by modelling non-relevant
articles
Henry Petersen1 and Josiah Poon1 Simon Poon1 Clement Loy2 Mariska
Leeflang3
1
School of Information Technologies, University of Sydney, Australia
2
School of Public Health, University of Sydney, Australia
3
Academic Medical Center, University of Amsterdam, Netherlands
hpet9515@uni.sydney.edu.au, {josiah.poon,simon.poon}@sydney.edu.au,
clement.loy@sydney.edu.au, m.m.leeflang@amc.uva.nl
Systematic reviews are widely considered as the highest form of medical evi-
dence, since they aim to be a repeatable, comprehensive, and unbiased summary
of the existing literature. Because of the high cost of missing relevant studies,
review authors go to great lengths to ensure all relevant literature is included.
It is not atypical for a single review to be conducted over the course of months
or years, with multiple authors screening thousands of articles in a multi-stage
triage process; first on title, then on title and abstract, and finally on full text.
Figure 1a shows a typical literature screening process for systematic reviews.
In the last decade, the information retrieval (IR) and machine learning (ML)
communities have shown increasing interest in literature searches for systematic
reviews [1–3]. Literature screening for systematic reviews can be characterised
as a classification task with two defining features; a requirement for near perfect
recall on the class of relevant studies (the high cost of missing relevant evidence),
and highly imbalanced training data (review authors are often willing to screen
thousands of citations to find less than 100 relevant articles). Previous attempts
at automating literature screening for systematic reviews have primarily focused
on two questions; how to build a suitably high recall model for the target class
in a given review under the conditions of highly imbalanced training data [1, 3],
and how best to integrate classification into the literature screening process [2].
When screening articles, reviewers exclude studies for a number of reasons
(animal populations, incorrect disease etc.). Additionally, in any given triage
stage a study may not be relevant but still progress to the next stage as the
authors have insufficient information to exclude it (i.e. the title may not indicate
a study was performed with an animal population, however this may become
apparent upon reading the abstract). We meet the requirement for near perfect
recall on relevant studies by inverting the classification task and identifying
subsets of irrelevant studies with near perfect precision. We attempt to identify
such studies by training the classifier using the labels assigned at the previous
triage stage (see Figure 1c). The seamless integration with the existing manual
screening process is an advantage of our approach.
The classifier is built by first selecting terms from the title and abstracts with
the greatest information gain on labels assigned in the first triage stage. Articles
Page 43
Joint Proceedings - AIH 2013 / CARE 2013
Initial
Screening on Exclude Initial
Screening on Exclude
Title Alone
Title Alone
Obtain Title
and Abstract Obtain Title Obtain Title
and Abstract and Abstract
Reviewer 1 Reviewer 2 – ’neutropenia’, but not
Screens on Screens on Reviewer 1
Title and Title and ’infection’ or ’thorax’ Screens on Build and Run
Abstract Abstract Title and Classifier
Abstract
Resolve Exclude
Discrepancies – ’skin’ but not ’thorax’ Resolve Both Exclude Exclude
Obtain Full One or more Include
Text Reviewer 2
– ’immunoglobulin g’ Screens on
Title and
Reviewer 1 Reviewer 2 Abstract
Screens on Screens on
Full Text Full Text – ’animals’ Resolve
Discrepancies
Exclude
Resolve Exclude
Discrepancies
Obtain Full
– ’drug therapy’, but not Text
Include ’risk’ or ’infection’
(a) (b) (c)
Fig. 1: Typical literature screening process for systematic reviews, sample rules
generated by our classifier, and the proposed modified screening process.
are then represented as Boolean statements over these terms, and interpretable
rules are then generated using Boolean minimisation (examples of rules are given
in 1b Review authors can then refine the classifier by selecting only those rules
most likely to describe non-relevant studies, maximising overall precision.
Preliminary experiments simulating the process outlined in Figure 1c on a
previously conducted systematic review indicate that as many as 25% of articles
can be safely eliminated without the need for screening by a second reviewer.
The evaluation does assume that all false positives (studies erroneously excluded
by the generated rules) were included by the first reviewer. Such an assumption
is reasonable; the reason for multiple reviewers is that even human experts make
mistakes. A study comparing the precision of our classifier to human reviewers is
planned. In addition, future work will focus on improving the quality of the gen-
erated rules by trying to better capture reasons for excluding studies matching
those used by human reviewers.
References
1. Aaron M. Cohen, Kyle H. Ambert, and Marian McDonagh. Research paper: Cross-
topic learning for work prioritization in systematic review creation and update.
JAMIA, 16(5):690–704, 2009.
2. Oana Frunza, Diana Inkpen, Stan Matwin, William Klement, and Peter OBlenis.
Exploiting the systematic review protocol for classification of medical abstracts.
Artificial Intelligence in Medicine, 51(1):17 – 25, 2011.
3. Stan Matwin, Alexandre Kouznetsov, Diana Inkpen, Oana Frunza, and Peter
O’Blenis. A new algorithm for reducing the workload of experts in performing
systematic reviews. JAMIA, 17(4):446–453, 2010.
Page 44
Joint Proceedings - AIH 2013 / CARE 2013
Optimizing Shiftable Appliance Schedules across
Residential Neighbourhoods for Lower Energy
Costs and Fair Billing
Salma Bakr and Stephen Cranefield
Department of Information Science, University of Otago, Dunedin, New Zealand
salma.bakr@postgrad.otago.ac.nz, scranefield@infoscience.otago.ac.nz
Abstract. This early stage interdisciplinary research contributes to smart
grid advancements by integrating information and communications tech-
nology and electric power systems. It aims at tackling the drawbacks
of current demand-side energy management schemes by developing an
agent-based energy management system that coordinates and optimizes
neighbourhood-level aggregate power load. In this paper, we report on
the implementation of an energy consumption scheduler for reschedul-
ing “shiftable” household appliances at the household-level; the sched-
uler takes into account the consumer’s time preferences, the total hourly
power consumption across neighbouring households, and a fair electricity
billing mechanism. This scheduler is to be deployed in an autonomous
and distributed residential energy management system to avoid load syn-
chronization, reduce utility energy costs, and improve the load factor of
the aggregate power load.
1 Introduction
Electric utilities tend to meet growing consumer energy demand by expand-
ing their generation capacities, especially capital-intensive peak power plants
(also known as “peakers”), which are much more costly to operate than base
load power plants. As this strategy results in highly inefficient consumption be-
haviours and under-utilized power systems, demand-side energy management
schemes aiming to optimally match power supply and demand have emerged.
Currently deployed demand-side energy management schemes are based on
the interactions between the electric utility and a single household [18], as in
Fig.1(a). As this approach lacks coordination among neighbouring households
sharing the same low-voltage distribution network, it may cause load synchro-
nization problems where new peaks arise in off-peak hours [15]. Thus, it is essen-
tial to develop flexible and scalable energy management systems that coordinate
energy usage between neighbouring households, as in Fig.1(b).
2 Background
The smart grid, or the modernized electric grid, is a complex system comprising
a number of heterogeneous control, communication, computation, and electric
Page 45
Joint Proceedings - AIH 2013 / CARE 2013
(a) (b)
Fig. 1. The interactions between the utility and the consumers in demand-side energy
management schemes are either: (a) individual interactions, or (b) neighbourhood-level
interactions
.
power components. It also integrates humans in decision making. To verify the
states of smart grid components in a simultaneous manner and take human
intervention into account, it is necessary to adopt autonomous distributed system
architectures whose functionality can be modelled and verified using agent-based
modelling and simulation.
Multi-agent systems (MAS) provide the properties required to coordinate
the interactions between smart grid components and solve complex problems in
a flexible approach. In the context of a smart grid, agents represent producers,
consumers, and aggregators at different scales of operation, e.g. wholesale and
retail energy traders [7]. MAS have been deployed in a number of smart grid
applications, with a more recent focus on micro-grid control [6, 17] and energy
management [10, 12] especially due to the emerging trend of integrating dis-
tributed energy resources (DER), storage capacities, and plug-in hybrid electric
vehicles (PHEV) into consumer premises.
In agent-based energy management systems, agents may aim at achieving a
single objective or a multitude of objectives; typical objectives include: balancing
energy supply and demand [4]; reducing peak power demand [13, 16]; reducing
utility energy costs [8, 16] and consumer bills [16]; improving grid efficiency [4];
and increasing the share of renewable energy sources [1, 12] which consequently
reduces the carbon footprint of the power grid. Agent objectives can be achieved
using evolutionary algorithms [8] or a number of optimization techniques such
as integer, quadratic [5, 13], stochastic [4] and dynamic programming [5]. As for
the interactions among agents, game theory provides a conceptual and a formal
analytical framework that enables the study of those complex interactions [19].
Page 46
Joint Proceedings - AIH 2013 / CARE 2013
3 Research Objectives
This research aims at optimizing the energy demand of a group of neighbouring
households, to reduce utility costs by using energy at off-peak periods, avoid
load synchronization that may occur due to rescheduling appliance usage, and
improve the load factor (i.e. the ratio between average and peak power) of the
aggregate load. A number of energy consumption schedulers have been proposed
in the literature [14, 16, 21]; however, those schedulers do not leverage an accu-
rately quantified and fair billing mechanism that charges consumers based on
the shape of their power load profiles and their actual contribution in reducing
energy generation costs for electric utilities [3]. In this paper, we implement and
evaluate an energy consumption scheduler that optimizes the operation times
of three wet home appliances and a PHEV at the household-level based on the
total hourly power consumption across neighbouring households, consumer time
preferences, and a fair electricity billing mechanism.
4 Methodology
We use the findings of Baharlouei et al. [3] to resolve a gap in the findings of
Mohsenian-Rad et al. [16]. Game-theoretic analysis is used by Mohsenian-Rad et
al. [16] to propose an incentive-based energy consumption game that schedules
“shiftable” home appliances (e.g. washing machine, tumble dryer, dish washer,
and PHEV) for residential consumers (players) according to their daily time pref-
erences (strategies); at the Nash equilibrium of the proposed non-cooperative
game, it is shown that the energy costs of the system are minimized. How-
ever, this game charges consumers based on their total daily electric energy
consumption rather than their hourly energy consumption. In other words, two
consumers having the same total daily energy consumption are charged equally
even if their hourly load profiles are different. This unfair billing mechanism may
thus discourage consumer participation as it does not take consumer reschedul-
ing flexibility into consideration. With this in mind, we propose leveraging the
fair billing mechanism recently proposed by Baharlouei et al. [3] to encourage
consumer participation in the energy consumption game.
5 Energy Consumption Scheduler
5.1 Formulation
Assuming a multi-agent system for managing electric energy consumption at the
neighbourhood-level, where agents represent consumers, each agent locally and
optimally schedules his “shiftable” home appliances to minimize his electricity
bill taking into account the following inputs: appliance load profiles, consumer
time preferences, grid limitations (if any), aggregate scheduled hourly energy
consumption of all the other agents in the neighbourhood, and the deployed
Page 47
Joint Proceedings - AIH 2013 / CARE 2013
electricity billing scheme. If the energy cost function is non-linear, knowing the
aggregate scheduled load is required for optimization.
After this optimization, each agent sends out his updated appliance schedule
to an aggregator agent, which then forwards the aggregated load to the other
agents to optimize their schedules accordingly. By starting with random initial
schedules, convergence of the distributed algorithm is guaranteed if household-
level energy schedule updates are asynchronous [16]. The electric utility may
coordinate such updates according to any turn-taking scenario.
We assume electricity distributed to the neighbourhood is generated by a
thermal power generator having a quadratic hourly cost function [23] given by
(1); as this equation is convex, quadratic, and has linear constraints, it can be
solved using mixed integer quadratic programming.
Ch (Lh ) = ah L2h + bh Lh + ch , (1)
where ah > 0, and bh , ch ≥ 0 at each hour h ∈ H = [1, ..., 24]. In (2), Lh and xhm
denote the total hourly load of N consumers and consumer m, respectively [16].
N
X
Lh = xhm , (2)
m=1
To encourage participation in energy management programmes, it is essential
to reward consumers with fair incentives. By rescheduling appliances to off-peak
hours where electricity tariffs are cheaper, we save on utility energy costs and
consequently impose monetary incentives for consumers in the form of savings
on electricity bills. The optimization problem therefore targets the appliance
schedule xhn that results in the minimum bill Bn for consumer (agent) n. The
billing equation proposed by Baharlouei et al. [3], which fairly maps a consumer’s
bill to energy costs (1), is given by (3).
H N
!
X xhn X
h
Bn = PN Ch xm , (3)
h
h=1 m=1 xm m=1
5.2 Set-up
To model the optimization problem such that each agent n individually and
iteratively minimizes (3), we use YALMIP — an open-source modelling lan-
guage that integrates with MATLAB. We consider a system of three households
(agents) and investigate the behaviour of one of those schedulers with respect
to fair billing, lower energy costs, and improved load factor. To model consumer
flexibility in scheduling, we consider two scenarios for the same household where
the consumer’s acceptance of rescheduling flexibility differ. We investigate the
two scenarios for four days in December, March, June and September.
To test our energy consumption scheduler, we choose to schedule a PHEV
and three wet appliances: a clothes washer, a tumble dryer, and a dish washer.
Wet appliance power load profiles are based on survey EUP14-07b [22], which
Page 48
Joint Proceedings - AIH 2013 / CARE 2013
was conducted with around 2500 consumers from 10 European countries. For the
PHEV load, we use the power load profile of a mid-size sedan at 240V–30A [9].
We choose a budget-balanced billing system and calibrate the coefficients
of the hourly energy cost function (1) against a three-level time-of-use pricing
scheme used by London Hydro [11], where the kilowatt-hour is charged at 12.4,
10.4, and 6.7 cents for on-, mid-, and off-peak hours, respectively. Energy con-
sumption of neighbouring households and non-shiftable loads of the household
investigated are taken from a publicly available household electric power con-
sumption data set [2], for the period from December 2006 to September 2007.
5.3 Scenario 1
In this scenario, we assume the consumer is not flexible about appliance schedul-
ing and use common startup times: clothes washing starts at 7 a.m., drying starts
two hours directly after washing starts, dish washing starts at 6 p.m. [22], and
PHEV recharging starts at 6 p.m. [20].
5.4 Scenario 2
The consumer is assumed to be flexible about appliance scheduling in Scenario
2; clothes washing starts any time between 6 a.m. and 9 a.m., drying any time
after washing but before 11 p.m., washing dishes any time after 7 p.m, but before
11 p.m., and PHEV recharging any time after 1 a.m. but before 5 a.m.
6 Results
6.1 Fair Billing
Results indicate that the consumer’s electricity bill for operating household
“shiftable” appliances in Scenario 2 is lower by 70%, 57%, 32%, and 65% com-
pared to that in Scenario 1 for the days chosen in December, March, June, and
September, respectively. This clearly indicates that flexibility is awarded fairly
through the deployed billing mechanism. Figures 2 and 3 depict the appliance
schedules resulting in the minimum bill for the household under investigation
and the aggregate non-shiftable load of neighbouring households, for Scenario 1
and 2 in December, respectively.
6.2 Lower Energy Costs
As we chose a budget-balanced billing system and since appliances are resched-
uled to cheaper off-peak hours, utility energy costs are lower in Scenario 2 by
70%, 57%, 32%, and 65% compared to that in Scenario 1, for the days chosen
across the four seasons, respectively.
Page 49
Joint Proceedings - AIH 2013 / CARE 2013
Fig. 2. Scenario 1: the unscheduled “shiftable” appliance loads of the consumer under
investigation and the aggregate “non-shiftable” neighbourhood-level loads (December)
1
Fig. 3. Scenario 2: the scheduled “shiftable” appliance loads of the consumer under
investigation and the aggregate “non-shiftable” neighbourhood-level loads (December)
6.3 Improved Load Factor
As the “shiftable” appliances of the household under investigation are resched-
uled to operate during off-peak hours instead of peak hours, the load factor of the
Page 50
Joint Proceedings - AIH 2013 / CARE 2013
aggregate load in Scenario 2 is improved by 44%, 13%, 19%, and 28% compared
to that in Scenario 1, for the days chosen across the four seasons, respectively.
This indicates improved resource allocation in the power grid.
7 Conclusion
In this paper, we leverage the fair billing mechanism proposed by Baharlouei
et al. [3] to evaluate the energy consumption scheduling game proposed by
Mohsenian-Rad et al. [16]. We have implemented and evaluated a scheduler
that optimally allocates the operation of “shiftable” appliances for a consumer
based on his time preferences, the aggregate hourly “non-shiftable” load at the
neighbourhood-level, and a fair billing mechanism. As the deployed billing mech-
anism takes advantage of cheaper off-peak electricity prices, we show that it
helps in lowering utility energy costs and electricity bills, and improving the
load factor of the aggregate neighbourhood-level power load. We also conclude
that consumer flexibility in rescheduling appliances is rewarded fairly based on
the shape of his power load profile rather than his total energy consumption.
8 Future Work
Eventually, we intend to investigate an appliance scheduler that coordinates
electric energy consumption among a large number of households (agents).
References
1. Baños, R., Manzano-Agugliaro, F., Montoya, F.G., Gil, C., Alcayde, A., Gómez, J.:
Optimization Methods Applied to Renewable and Sustainable Energy: A Review.
Renewable and Sustainable Energy Reviews 15(4), 1753–1766 (2011)
2. Bache, K., Lichman, M.: UCI machine learning repository (2013),
http://archive.ics.uci.edu/ml
3. Baharlouei, Z., Hashemi, M., Narimani, H., Mohsenian-Rad, A.H.: Achieving Op-
timality and Fairness in Autonomous Demand Response: Benchmarks and Billing
Mechanisms. IEEE Transactions on Smart Grid 4(2), 968–975 (2013)
4. Clement, K., Haesen, E., Driesen, J.: Coordinated Charging of Multi-
ple Plug-in Hybrid Electric Vehicles in Residential Distribution Grids.
In: IEEE PES Power Systems Conference and Exposition (2009),
http://dx.doi.org/10.1109/PSCE.2009.4839973
5. Clement-Nyns, K., Haesen, E., Driesen, J.: The Impact of Charging Plug-In Hybrid
Electric Vehicles on a Residential Distribution Grid. IEEE Transactions on Power
Systems 25(1), 371–380 (2010)
6. Colson, C.M., Nehrir, M.H.: Algorithms for Distributed Decision Making for Multi
Agent Microgrid Power Management. In: IEEE Power and Energy Society General
Meeting (2011), http://dx.doi.org/10.1109/PES.2011.6039764
7. Gammon, R.: CASCADE Overview. In: Sustainable Energy, Complexity Science
and the Smart Grid, Satellite Workshop in the European Conference on Complex
Systems (2012)
Page 51
Joint Proceedings - AIH 2013 / CARE 2013
8. Guo, Y., Li, J., James, G.: Evolutionary Optimisation of Distributed Energy Re-
sources. In: 18th Australian Joint Conference on Advances in Artificial Intelligence.
vol. 3809, pp. 1086–1091. Springer Berlin Heidelberg (2005)
9. Hadley, S.W.: Impact of Plug-in Hybrid Vehicles on the Electric Grid.
Tech. Rep. ORNL/TM-2006/554, Oak Ridge National Laboratory (2006),
http://web.ornl.gov/info/ornlreview/v40 2 07/2007 plug-in paper.pdf
10. Lamparter, S., Becher, S., Fischer, J.: An Agent Based Market Platform for Smart
Grids. In: International Conference on Autonomous Agents and Multiagent Sys-
tems. pp. 1689–1696. IFAAMAS (2010)
11. London Hydro: Regulated Price Plan Time-of-Use Rates (2013),
http://www.londonhydro.com/residential/electricityrates/
12. Mets, K., Strobbe, M., Verschueren, T., Roelens, T., De Turck, F., Develder, C.:
Distributed Multi-agent Algorithm for Residential Energy Management in Smart
Grids. In: IEEE Network Operations and Management Symposium. pp. 435–443
(2012), http://dx.doi.org/10.1109/NOMS.2012.6211928
13. Mets, K., Verschueren, T., Haerick, W., Develder, C., Turck, F.D.: Optimizing
Smart Energy Control Strategies for Plug-in Hybrid Electric Vehicle Charging.
In: IEEE/IFIP Network Operations and Management Symposium Workshops. pp.
293–299 (2010), http://dx.doi.org/10.1109/NOMSW.2010.5486561
14. Mohsenian-Rad, A.H., Wong, V., Jatskevich, J., Schober, R.: Optimal and Au-
tonomous Incentive-based Energy Consumption Scheduling Algorithm for Smart
Grid. In: IEEE PES Conference on Innovative Smart Grid Technologies (2010),
http://dx.doi.org/10.1109/ISGT.2010.5434752
15. Mohsenian-Rad, A.H., Leon-Garcia, A.: Optimal Residential Load Control With
Price Prediction in Real-Time Electricity Pricing Environments. IEEE Transac-
tions on Smart Grid 1(2), 120–133 (2010)
16. Mohsenian-Rad, A.H., Wong, V.W.S., Jatskevich, J., Schober, R., Leon-Garcia, A.:
Autonomous Demand-Side Management Based on Game-Theoretic Energy Con-
sumption Scheduling for the Future Smart Grid. IEEE Transactions on Smart Grid
1(3), 320–331 (2010)
17. Pipattanasomporn, M., Feroze, H., Rahman, S.: Multi-Agent Systems in a Dis-
tributed Smart Grid: Design and Implementation. In: IEEE PES Power Systems
Conference and Exposition (2009), http://dx.doi.org/10.1109/PSCE.2009.4840087
18. Ruiz, N., Cobelo, I., Oyarzabal, J.: A Direct Load Control Model for Virtual Power
Plant Management,. IEEE Transactions on Power Systems 24(2), 959–966 (2009)
19. Saad, W., Han, Z., Poor, H.V., Baar, T.: Game-Theoretic Methods for the Smart
Grid: An Overview of Microgrid Systems, Demand-Side Management, and Smart
Grid Communications. IEEE Signal Processing Magazine 29(5), 86–105 (2012)
20. Shao, S., Pipattanasomporn, M., Rahman, S.: Challenges of PHEV Penetration to
the Residential Distribution Network. In: IEEE Power and Energy Society General
Meeting (2009), http://dx.doi.org/10.1109/PES.2009.5275806
21. Shinwari, M., Youssef, A., Hamouda, W.: A Water-Filling Based Scheduling Algo-
rithm for the Smart Grid. IEEE Transactions on Smart Grid 3(2), 710–719 (2012)
22. Stamminger, R., Friedrich-Wilhelms, R.: Synergy Potential of Smart Ap-
pliances. Deliverable DP2.3 of WP2, EIE Project on Smart Domes-
tic Appliances in Sustainable Energy Systems (2008), http://www.smart-
a.org/D2.3 Synergy Potential of Smart Appliances 4.00.pdf
23. Wood, A., Wollenberg, B.: Power Generation, Operation, and Control. Wiley-
Interscience, 2 edn. (1996)
Page 52
Joint Proceedings - AIH 2013 / CARE 2013
Proposal of information provision to probe
vehicles based on distribution of link travel time
that tends to have two peaks
Keita Mizuno, Ryo Kanamori, and Takayuki Ito
Nagoya Institute of Technology,
Gokiso, Showa, Nagoya 466-8555, JAPAN
mizuno.keita@itolab.nitech.ac.jp,
kanamori.ryo@nitech.ac.jp,
ito.takayuki@nitech.ac.jp
http://www.itolab.nitech.ac.jp/itl2/page_en/
Abstract. In most cities, traffic congestion is a primary problem that
must be tackled. Traffic control/operation systems based on information
gathered from probe vehicles have attracted a lot of attention. In this pa-
per, we examine provision of travel information to eliminate traffic jams.
Although it is conventional to provide the mean of historical accumu-
lated data, we introduce the 25th percentile and 75th percentile values
because a distribution of link travel time tends to have two peaks. As
a result, the proposed method reduced travel time of vehicles compared
with the conventional method.
Keywords: Traffic management, Probe car, Intelligent Transport Sys-
tem
1 Introduction
Automobile traffic jams have become a major problem in many cities of the
world. In Japan, an increase in vehicle emissions and time loss due to traffic
congestion have also become significant problems. As a solution to these prob-
lems, information collected from probe vehicles is attracting attention. In this
research, we assume an environment in which information of the travel time of
a vehicle in the past can be obtained, vehicles can communicate mutually, and
vehicles can share traffic conditions to reduce the travel time of all vehicles.
Thus, we propose a method of providing information to a probe vehicle for re-
ducing travel time of regular vehicles, and show the effectiveness of the proposed
method by simulation experiments.
In this research, we focus on how a distribution of link travel time tends to
have two peaks for historical accumulated data of travel time of the vehicle. In
addition to the mean of historical accumulated data of the link travel time, us-
ing the 25th percentile value and 75th percentile value of historical accumulated
data, we perform path finding and give information to the probe vehicle. Fur-
thermore, to demonstrate that the proposed method of this research is effective,
Page 53
Joint Proceedings - AIH 2013 / CARE 2013
2 Keita Mizuno, Ryo Kanamori, and Takayuki Ito
we implement traffic flow simulation based on the cell transmission model[1][2],
and we perform vehicle movement simulation of the conventional method and
proposed method. We use travel time of the vehicle, which has also been used
in conventional research, for the effect analysis of information provided to the
probe vehicle. In addition, we examine the difference between the time taken to
move in the simulation and travel time to the destination that is expected from
the historical accumulated data of the vehicle.
The remainder of this paper is organized as follows. Background and purpose
of this research are presented in chapter 2, and the distribution of link travel time
having two peaks is discussed in chapter 3. We describe the proposed information
provision method in chapter 4, the vehicle simulation in chapter 5, and the
effectiveness of the proposed method, along with future work in chapter 6.
2 Background and purpose
In this chapter, we describe the background and purpose of this research. Per-
sonal vehicles have become an essential means of transportation for many people.
However, there are many problems we must solve; for example, decline in eco-
nomic efficiency due to traffic congestion, global environmental degradation such
as global warming and air pollution, and many traffic accidents. Transportation
and traffic account for about 20% of carbon dioxide emissions in Japan, and of
that, vehicles account for about 90%[3]. Figure 1 is a diagram showing the rela-
tionship between carbon dioxide emissions and the running speed of a vehicle.
Because we can see that the carbon dioxide emissions from the vehicle decrease
when running speed of the vehicle increases, we must decrease carbon dioxide
emissions by eliminating traffic congestion, and increasing the running speed of
the vehicle. Also, there are approximately 5 billion hours per year in time lost to
congestion in Japan, and the economic loss is 11 trillion yen. Problems caused by
traffic congestion have clearly become serious in Japan, as in many other parts
of the world, and it is necessary to resolve these issues.
In addition to the promotion of next-generation vehicles such as electric cars
as a way to solve these problems, traffic operation and management measures
by Intelligent Transport Systems (ITS), such as providing path information and
road pricing, have attracted attention. The number of vehicles with vehicle per-
ception and navigation systems (probe vehicles) is increasing, and technology
of information collection and provision has also advanced in route search infor-
mation. Further, from the historical accumulated data collected from the probe
vehicle, it is observed that a distribution of link travel time tends to have two
peaks.
About providing information to the probe vehicle, Kanamori et al.[4] sim-
ulated providing information to a probe vehicle using not only the historical
accumulated data collected from the probe vehicle but also predicting the traffic
situation. Morikawa et al.[5] simulated providing information to a probe vehi-
cle using the number of right and left turns in the path to the destination, in
addition to the historical accumulated data collected from the probe vehicle.
Page 54
Joint Proceedings - AIH 2013 / CARE 2013
Proposal of information provision to probe vehicles 3
Fig. 1. Relationship between carbon dioxide emissions and running speed of vehicle
In researches of Kanamori et al. and Morikawa et al., they simulated providing
information that uses the mean of historical accumulated data collected from
probe vehicles, and searches for a route to a destination.
The purpose of this research is to propose a method to use historical accu-
mulated data focusing on the distribution of link travel time, which tends to
have two peaks, and reducing travel link time of vehicles in the simulation.
3 Distribution of link travel time
In this section, we discuss how a distribution of link travel time tends to have
two peaks. Link travel time of the vehicle described in this research is the time
to travel from one intersection to another.
Figure 2 shows example of distribution of link travel time. It is observed that
a distribution of link travel time tends to have two peaks when the vehicles pass
through the intersection, and simulations that reproduce a distribution of link
travel time have been researched[6].
The cause of the link travel time of the vehicle having two peaks is, for
example, a traffic signal. When the vehicle passes through an intersection, a
considerable difference occurs because the vehicle stops at the signal or doesn’t
stop. In previous research, they didn’t consider that a distribution of link travel
time tends to have two peaks; instead, they used the mean value of the link
travel time collected from the probe vehicle.
4 Information provision to probe vehicles
In this chapter, we provide a detailed description of the method of information
provision to the probe vehicle in this research. As usage of the historical accu-
Page 55
Joint Proceedings - AIH 2013 / CARE 2013
4 Keita Mizuno, Ryo Kanamori, and Takayuki Ito
Fig. 2. Distribution of link travel time
mulated data of link travel time for searching the route to the destination, in
addition to a conventional method to provide the mean of historical accumulated
data of the travel time, we introduce provisions of the 25th percentile value and
75th percentile value of historical accumulated data of the travel time in this
research.
Probe vehicle assumed in this paper is sending information of link travel
time and receiving information of path to destination with least travel time.
Information of path to destination with least travel time is predicted by link
travel time collected from probe vehicle.
In this experiment, we use the data of the 25th percentile and 75th percentile
values of the historical accumulated data of link travel time. To decide which
value we will use in this research, we conduct a preliminary experiment. First,
we used only the 25th percentile value of the historical accumulated data in
the information-providing simulation. Second, we used only the 75th percentile
value of the historical accumulated data in the information-providing simulation.
We compared the mean of historical accumulated data of the link travel time
with 25th percentile and 75th percentile values regarding the travel time of the
vehicle. In this research, assuming the differences of factors such as the number
of intersections passed through depending on the travel distance of the vehicle,
Page 56
Joint Proceedings - AIH 2013 / CARE 2013
Proposal of information provision to probe vehicles 5
we compare the mean value, 25th percentile and 75th percentile values by travel
distance of each vehicle.
We set the travel distance of vehicles using the 25th percentile or 75th per-
centile values in the simulation, and conduct information provision simulation
using the 25th percentile and 75th percentile values for searching the route to
the destination.
5 Simulation for evaluation
5.1 Settings of simulation
We use the data of Kichijoji and Mitaka that are provided in the traffic simu-
lation clearing house as a road network used for the evaluation experiment in
this research. The traffic simulation clearing house[7] is an institution providing
various data for validation. The network is composed of 57 nodes and 137 links.
Vehicles in the simulation number about 17,000 units, and approximately 50%
are probe vehicles in this experiment. Further, in order to accumulate link travel
time for the vehicles to be used for route search, the simulation was repeated
about 30 times. Figure 3 is a network diagram from Kichijoji and Mitaka that
is used for the simulation in this research.
Fig. 3. Network of Kichijoji and Mitaka
Page 57
Joint Proceedings - AIH 2013 / CARE 2013
6 Keita Mizuno, Ryo Kanamori, and Takayuki Ito
Table 1 shows survey contents collected by Kichijoji and Mitaka. Investiga-
tion time is set to a high-traffic period.
Since the investigation data contain the times each vehicle entered and exited
the network, we can obtain the travel time to the destination of each vehicle.
Table 1. Details of survey of Kichijoji and Mitaka
investigation time AM 7:00 AM 10:00
target area Mitaka and Musashino, Tokyo
observation points 70
target vehicle four wheel vehicles
survey contents passage time
vehicle number
car model(bus, taxi, and other)
5.2 Traffic flow simulation
In this research, we implemented a traffic flow simulation based on the cell trans-
mission model, in which the repeatability of travel time is high and we can control
the route choice of the vehicle in the simulation. The cell transmission model is
a model that divides the network links into cells and controls the movement of
vehicles by the density of vehicles in a cell.
yi (t) = min{ ni−1 (t), Qi (t), Ni (t) − ni (t)} (1)
– yi (t): number of vehicles moving to the cell of index i at time t
– Qi (t): maximum number of vehicles that can flow into the cell of index i
at time t
– Ni (t): maximum number of vehicles in the cell of index i at time t
– ni (t): number of vehicles in the cell of index i at time t
Equation (1) represents the number of vehicles to move between cells on the
cell transmission model. The number of vehicles that can move to the next cell
is determined by the smallest number of the following: number of vehicles in the
present cell, the amount of empty space in the next cell, or maximum number
of vehicles that can flow into the next cell. Equation (2) represents traffic flow
rate.
q = k ∗v (2)
– q: traffic flow rate in the cell.
– k: vehicle density in the cell.
Page 58
Joint Proceedings - AIH 2013 / CARE 2013
Proposal of information provision to probe vehicles 7
– v: vehicle speed in the cell.
Traffic flow rate can be calculated from the vehicle speed and vehicle density
in the cell. There are many equations that can calculate the vehicle speed from
the density. In this research, we use the formula of Green Shields[8] to calculate
the traffic flow rate.
The traffic flow simulation implemented in this research uses a data set of
network and departure time, departure point, destination point, and whether
the vehicle is a probe vehicle. To verify the reproducibility of the traffic flow
simulation, we compare ours with the traffic flow simulation based on the cellular
automata model[9][10] regarding a coefficient of simple linear regression and root
mean square of the travel time of the vehicle. The cellular automata model is a
discrete model and is easy to implement. In the experiments, root mean square
being close to 0 and a coefficient of simple linear regression being close to 1
represents that the reproducibility of vehicle travel time is high.
Table 2. Comparison of cellular automata model and cell transmission model for
reproducibility of travel time
model root mean square coefficient of simple linear regression
cell transmission 2.029 0.835
cellular automata 3.502 0.339
Table 2 shows the results of a comparison of the coefficient of simple linear
regression and the root mean square regarding the simulation based on the cel-
lular automata model and the cell transmission model. Table 2 shows that the
reproducibility of the travel time in the simulation based on the cell transmis-
sion model is greater than that of the cellular automata model from the values
of both the coefficient of simple linear regression and the root mean square.
Traffic flow simulation that reproduces a distribution of link travel time tend-
ing to have two peaks is required for information provision and shows the effec-
tiveness of proposed method.
Figure 4 shows that the passage number and travel times of the vehicles on
one link in the network when we simulated movement of the vehicles using the
Kichijoji and Mitaka data set on traffic flow simulation. As Figure 4 shows, it
was confirmed that it is possible to reproduce a distribution of link travel time
tending to have two peaks in the traffic flow simulation implemented in this
research.
5.3 Experimental result
Difference of the travel time for each distance of vehicles We show
the comparison results regarding the travel time of vehicles between using the
Page 59
Joint Proceedings - AIH 2013 / CARE 2013
8 Keita Mizuno, Ryo Kanamori, and Takayuki Ito
Fig. 4. Traffic volumes and travel time of the vehicles at a certain link in the simulation
Fig. 5. Difference in travel time of vehicles between using the mean value and 75th
percentile value for route search by travel distance of vehicles
mean value, 25th percentile value and 75th percentile value of the historical
accumulated data of the link travel time.
Figures 5 and 6 show difference of travel time between using the mean, 25th
percentile value, and 75th percentile value for route search by travel distance
of vehicle. The value of the graph subtracts the travel time when using 75th
percentile and 25th percentile values from the travel time in case of using the
mean value. As the value of the graph is large, it represents that the travel time
Page 60
Joint Proceedings - AIH 2013 / CARE 2013
Proposal of information provision to probe vehicles 9
Fig. 6. Difference in travel time of vehicles between using the mean value and 25th
percentile value for route search by travel distance of vehicles
of vehicles using the mean value is more than the travel time of vehicles using
the 25th percentile value and 75th percentile value. In Figure 5, the travel time
of vehicles using the 75th percentile value is less than that using the mean value
regarding vehicles that travel distances of 1,000 meters or more. On the other
hand, in Figure 6, the travel time of vehicles using the 25th percentile value is
less than that using the mean value regarding vehicles that travel distances of
1,000 meters or less.
Proposed method and evaluation In this research, we proposed that vehicles
whose travel distance is 1,000 meters or less perform a route search using the
25th percentile value of historical accumulated data, and vehicles whose travel
distance is 1,000 meters or more perform a route search using the 75th percentile
value of historical accumulated data. The effect analysis is the total travel time
of all vehicles in the simulation.
Figure 7 shows the result of the simulation experiment in each case. Values
in the graph of Figure 7 show the total travel time of all vehicles in each case.
We describe setting of each case. There is no probe vehicle in case 1; that is,
vehicles do not change their routes in repetition. The probe vehicles search for
the route using mean value in case 2, 25th percentile value in case 3, and 75th
percentile value in case 4 as link cost. We use the proposed method in case 5.
As shown in the graph of Figure 7, using both 25th percentile value and
75th percentile value of historical accumulated data reduced the travel time of
all vehicles most.
Page 61
Joint Proceedings - AIH 2013 / CARE 2013
10 Keita Mizuno, Ryo Kanamori, and Takayuki Ito
Fig. 7. Total travel time of all vehicles in each information provision
6 Conclusion and future work
In this research, we presented background information about the problems caused
by the increasing number of vehicles on the road, such as economic losses and
environmental degradation. Also, the number of probe vehicles has increased in
recent years, and the distribution of link travel time tends to have two peaks.
Next, we proposed information provision based on a distribution of link travel
time tending to have two peaks. In the experimental simulation, as the infor-
mation provision to the probe vehicle, we proposed using both the 25th per-
centile and 75th percentile values as a function of travel distance of a vehicle.
We demonstrated that the proposed method reduced the travel time of all vehi-
cles compared with the conventional method.
In future work, we will simulate a large network. In this experiment, since we
used a small network data set, it is necessary to test a larger network to confirm
that the proposed method is effective.
The information method proposed in this research used travel distance of
the vehicles; it is also necessary to use such factors as the departure time of the
vehicles in future research.
References
1. Carlos F. Daganzo: The cell transmission model: A dynamic representation of high-
way traffic consistent with the hydrodynamic theory, Transportation Research B,
28(4):269-287(1994)
2. Carlos F. Daganzo: The cell transmission model, part II: Network traffic, Trans-
portation Research Part B: Methodological, Volume 29, Issue 2, Pages 79-93, (1995)
Page 62
Joint Proceedings - AIH 2013 / CARE 2013
Proposal of information provision to probe vehicles 11
3. Ministry of Land, Infrastructure, Transport and Tourism: amount of carbon dioxide
emissions:
http://www.mlit.go.jp/sogoseisaku/environment/sosei environment tk 000007.html,
Ministry of Land, Infrastructure, Transport and Tourism
4. Ryo Kanamori: Jun Takahashi, Takayuki Ito: A proposal of Route Information Pro-
vision with Anticipatory Stigmergy for Traffic Management, IPSJ SIG Technical
Report. ICS, 2012-ICS-169(1), 1-6, (2012)
5. Takayuki Morikawa: Toshiyuki Yamamoto, Tomio Miwa, Lixiao Wang: Development
and Performance Evaluation of Dynamic Route Guidance System ”PRONAVI” ,
Traffic Engineering, Vol.42, No.3, pp.65-75(2007)
6. Mohsen Ramezani, Nikolas Geroliminis: On the estimation of arterial route travel
time distribution with Markov chains, Elsevier, vol. 46, pages 1576-1590(2012)
7. Traffic Simulation Clearing House: http://www.jste.or.jp/sim/index.html
8. B.D.Greenshields: A Study in Highway Capacity, H. R. B., Proceedings, Vol.14,
p.468(1935)
9. A. Schadschneider, M. Schreckenberg: Cellular automaton models and traffic flow,
J. Phys. A26, L679 (1993)
10. Shinichi Tadaki, Makoto Kikuchi: Stability of congestion phase in a two-
dimensional cellular automata traffic model, Bussei Kenkyu 61(5), 453-457(1994)
11. Hirokazu Akahane, Takashi Ooguchi, Hiroyuki Oneyama: Genealogy of the devel-
opment of traffic simulation model, Traffic Engineering 37(5), 47-55(2002)
Page 63
© AIH 2013 / CARE 2013