=Paper=
{{Paper
|id=None
|storemode=property
|title=First Steps towards a Frequent Pattern Mining with Nephrology Data in the Medical Domain
|pdfUrl=https://ceur-ws.org/Vol-928/0261.pdf
|volume=Vol-928
|dblpUrl=https://dblp.org/rec/conf/csp/NiemannSTH12
}}
==First Steps towards a Frequent Pattern Mining with Nephrology Data in the Medical Domain==
https://ceur-ws.org/Vol-928/0261.pdf
First Steps towards a Frequent Pattern Mining with
Nephrology Data in the Medical Domain
- Extended Abstract -
Matthias Niemann1 , Danilo Schmidt2, Gabriela Lindemann – von Trzebiatowski3,
Carl Hinrichs2,
University Hospital Charité
1
Department of Transfusion Medicine
2
Department of Nephrology
Humboldt University of Berlin
3
Department of Computer Science
matthias.niemann@charite.de
danilo.schmidt@charité.de
gabriela.lindemann@uv.hu-berlin.de
carl.hinrichs@charité.de
Abstract. Constantly increasing, complex and heterogeneous data collections
are characteristic for our time. That occurs in nearly all spheres of our life. One
of the main challenges is the handling of this huge amount of data. Data mining
is a field of computer science which is dealing with this special problem. This
paper describes our project which uses Data mining in the medical domain. Our
medical data come from the Department of Nephrology and include data of
patients which have got kidney transplantations. In this domain a lot of different
kind of data were produced, because patients with organ transplantation must be
in treatment for the rest of their lives. One of the big aims in the transplantation
science is a long-term graft survival. In this project we use Data mining to
identify new pattern in this complex data, which have an influence on a better
outcome after kidney transplantation.
Keywords: data mining, electronic patient record, medical data
Introduction
The advances in transplant medicine have grown rapidly over the last 50 years.
Despite a variety of new transplantation methods, surgical techniques, immuno-
suppressive regimens and innovative rejection treatment it is not yet managed
successfully to increase long-term survival of transplanted kidneys significantly.
Because of the complex interplay between the existing physical damage caused by a
long period of renal insufficiency, especially in the area of blood vessels, and
complex immunological reactions, it is not trivial to make specific statements without
an adequate epidemiological background.
The Charité – Universitätsmedizin Berlin has one of the most assured and valid
collection of records of hundreds of kidney transplants. To analyze the wealth of data
to identify new patterns that allow a useful patient outcome statement is a major
challenge to identify important factors retrospectively.
Statistic analysis is the regularly way of finding correlations in complex data sets.
Preliminary the researcher must have an intuition of interesting parameters. But this
also means that he can only find what he a-priori searches for. A suitable way to avoid
this disadvantage is data mining. Data mining is working explorative and the chance
to get a statistical bias is not very high, because an artificial system has no intuitive
preconditions. Data mining allows processing of a huge amount of data with the
possibility of getting details of potential interests in short time [HAN et. al. 2011]. In
the Charité several data sources exit with a huge amount of different highly complex
data. The number of data sets is high enough to find significant rules in this data.
The aim of the project has been the implementation of an explorative data search
engine including Data mining features. One of these features is the Frequent Pattern
Mining which we use in our project.
Data source
Our plan is a joint project between the department of Nephrology and the Laboratory
of Tissue Typing. This implies that the data are coming from both departments. The
bulk of the data comes from the nephrology, which deals with illnesses of the kidney
including the organ replacement (dialysis, transplantation). Illnesses of the kidney
cause a considerable cost in the German health service and worldwide. Considering
Germany, annually more than 2.5 billion Euros are necessary for the financiering of
dialysis treatments, whereby the incidence of the dialysis-requiring kidney
insufficiency raises annually around 5-8%. The transplantation of a kidney as
alternative to dialysis is reduced by the small availability of donor organs.
The Charité - Universitätsmedizin Berlin has three departments for the treatment of
nephrological diseases due to the historical development of the union of the clinical
centers of the University Hospitals of the Humboldt University of Berlin and Freie
Universität Berlin.
Figure 1: TBase© data structure
Each one of the three mentioned nephrology departments treats 3000 - 4000 cases per
year. The availability of all cases of Berlin’s university nephrology departments, the
sum of about 12.000 cases, would open completely new possibilities of detection of
rare disease patterns and quality management [SCHMIDT et. al. 2010].
All of the three are using the web-based Electronic Patient Record database TBase©.
It is based on a “normal” relational database structure, which means, that it has fixed
patterns of entities and their relations (see figure 1 only to get an impression). We
used the web-based Electronic Patient Record TBase© which was implemented in a
German kidney transplantation program as cooperation between the Nephrology of
Charité Campus Mitte and Campus Virchow and the AI Lab of the Institute of
Computer Sciences of Humboldt University of Berlin [SCHRÖTER et. al. 2000],
[LINDEMANN et. al. 2000]. Currently TBase© automatically integrates essential data
from the laboratory, clinical pharmacology, nuclear medicine, findings from
radiology and administrative data from the SAP©-system of the Charité.
The Laboratory of Tissue Typing is the second partner of our project. For the
patients which are coming from the patient record TBase© the laboratory identifies the
tissue type, which is very important for transplantation. Currently we have more than
10 million data sets of laboratory values, examination and treatment data which were
compiled over the last 13 years. We will use these data sets for our Data mining
project, supporting the retrieval of information about the complex mechanism of an
optimized outcome for the transplantation and patient survival hidden in the huge
amount of data. Questions of this kind are of great interest because these complex
mechanisms are not completely comprehended at present. We are sure that our Data
mining project will give some hints and support in this special research area.
Pre-processing of the medical data
In our Frequent Pattern Mining two dimensions are interesting: the time and the
value. For the first dimension “time” the user can define three categories of time
intervals by our GUI.
1.) Static time interval with a start date and end date
2.) Event time interval with start event and an end event
3.) mixed time interval with event or date as start or end point
In our tool it is possible to generate as many time intervals as needed with respect to
these categories (see figure 2).
Figure 2: time slices
For the second dimension “value” the user can preprocess the data in so called norm
ranges. In a first step different labeled values which are meaning the same contend
can be combined under a common “norm family name”.
This is important because in retrospective data sets it is normal that parameters are not
consistent labeled over the time.
In a next step the user has the possibility to define the value range e.g. for low,
medium and high values or to categorize this data e.g. in positive or negative values.
Here again our tool has the possibility to create as many norm definitions as needed
(see figure 3).
Figure 3: norms
Following these definitions, the tool creates a CQL (Charité Query Language) - file
where all these information / rules are saved. The CQL file has a user identifier. So,
later it can be provided to a client computer with respect to the user who defined the
query (see figure 4 and 5).
Figure 4: CQL file
Additionally it creates another automated file. This file regulates the next steps in the
mining process and allows a parallel processing of the data. The files include the
schedule of the necessary tasks (see figure 5 and 6).
Figure 5: CQL file is user based
Figure 6: Automated file (XML view)
Figure 7: Automated file (structured view)
The preprocessing algorithm generates the mining item via the CQL rules out from
the data source. The first preliminary result of our project is a table with all created
items.
Outlook
These structured items are suitable for the Frequent Pattern Mining. Frequent
Pattern Mining is a well-known Data mining technique. At the beginning of the 90’s
years Agrawal published the first efficient algorithm for the generation of association
rules. The name of this algorithm is “Apriori”. [HAN et. al. 2011] A further approach
is the “FP-Growth” algorithm published by Han et al.. The “FP-Growth” algorithm is
based on a tree structure. In contrast to the “Apriori” algorithm this allows a better
performance in huge amount of patterns or very long patterns. We except to find such
patterns in the medical data and therefore we think that the FP-growth algorithm is the
most useful one. In a next step we will implement such an algorithm.
Further we intend to develop an adjustment of confidence and support values,
because we are sure it is important and comfortable for the users. Moreover we will
create a feature to filter found items. All in all we want to reduce the complexity of
control of the tool for the convenience for users.
At the same time we will implement an exporter so that the user has the possibility to
investigate the items with other data mining /statistic tools e.g. WEKA or SPSS.
At the end of our development we will split the medical data in a training bulk and a
control bulk. After the Frequent Pattern Mining processing we will evaluate the
results with statistic methods.
Conclusion
We made a data integration of transplantation data, laboratory values and clinical
reports which come from the electronic patient record TBase and laboratory values,
which comes from the Laboratory of Tissue Typing. We have built a tool for
preprocessing for the Frequent Pattern Mining. Via a GUI the users have the
possibility to define time scales with respect to defined events. Furthermore the users
can configure values with special characteristics. These defined rules are stored in an
especial file structure named CQL (Charité Query Language). These files can be
referred to the user. The file will sent to a server were the items are generated
according to the defined rules. The final items set can be used for the Frequent Pattern
Mining. The plan is to use a FP-Growth Algorithm. Furthermore an exporter should
be implemented for other mining or statistic tools.
References:
[HAN ET. AL. 2011] J. Han, M. Kamber, J. Pei: Data Mining: Concepts and Techniques, Morgan
Kaufmann Series in Data Management Systems, 3rd revised edition, ISBN: 978-
0123814791, New York, 2011.
[LINDEMANN ET. AL. 2000] G. Lindemann, L. Fritsche: Web-Based Patient Records - The
Design of TBase2. In Bruch, Köckerling, Bouchard, Schug-Paß (Eds.) New Aspects of High
Technology in Medicine; Seiten 409-414; 2000.
[SCHMIDT ET. AL. 2010] D. Schmidt , G. Lindemann,: Precision and Recall of the Intelligent
Catalogue in the OpEN.SC - Extended Abstract -. In: Proceedings of “Concurrency,
Specification & Programming” – CS&P2010; Vol. I, II, III. Informatik Berichte, Institute of
Computer Science, Humboldt University Berlin, ISSN: 0863-95X, Germany, 2010.
[SCHRÖTER ET.AL. 2000] K. Schröter, G. Lindemann, L. Fritsche: TBase2 – A web-based
Electronic Patient Record. Fundamenta Informaticae 43, 343-353, IOS Press, Amsterdam,
2000