=Paper= {{Paper |id=Vol-1359/paper6 |storemode=property |title=A RESTful Approach for Developing Medical Decision Support Systems |pdfUrl=https://ceur-ws.org/Vol-1359/paper6.pdf |volume=Vol-1359 |dblpUrl=https://dblp.org/rec/conf/esws/WellerMMM15 }} ==A RESTful Approach for Developing Medical Decision Support Systems== https://ceur-ws.org/Vol-1359/paper6.pdf

Services and Applications over Linked APIs and Data – SALAD2015 37

A RESTful Approach for Developing
Medical Decision Support Systems

Tobias Weller, Maria Maleshkova, Keno März, and Lena Maier-Hein

AIFB, Karlsruhe Institute of Technology (KIT)
tobias.weller@student.kit.edu, maria.maleshkova@kit.edu
German Cancer Research Center
k.maerz@dkfz.de, l.maier-hein@dkfz.de

Abstract. Current developments in the medical sector are witnessing
the growing digitalization of data in terms of patient tests, records and
trials, use of sensors for monitoring and recording procedures, and employ-
ing digital imagery. Besides the increasing number of published guidelines
and studies, it has been shown that clinicians are often unable to ob-
serve them correctly during the actual care process. These trends provide
the foundation for the development of medical assistance systems which
process the gathered data and assist the physicians in making decisions,
preparing treatment plans, and can even guide surgeons during invasive
procedures. In this paper we demonstrate how a RESTful architecture
combined with applying Linked Data principles for data storage and
exchange can effectively be used for developing medical decision support
systems. We propose different autonomous subsystems that automatically
process data relevant to their purpose. These so-called ”Cognitive Apps”
provide RESTful interfaces and perform tasks such as generating sam-
ple patient data, converting and uploading data, and deducing medical
knowledge by using inference rules. The result is an adaptive decision
support system, based on distributed decoupled Cognitive Apps, which
can preprocess data in advance but also support real-time scenarios. We
demonstrate the practical applicability of our approach by providing an
implementation of a system for processing patients with liver tumors.
Finally, we evaluate the system in terms of knowledge deduction and
performance.

1 Introduction
The growing use of sensors in the medical domain, designated devices for recording
patient data, and the digitalization of medical knowledge in terms of recording
trials or medical guidelines result in large data volumes, which are hard to process
and manage by individual physicians. Nowadays, most of the patient data is
stored in semi-structured document formats such as spreadsheets, while the
results of clinical trials are published directly as text in papers. At the same time
more and more sensors are being used to observe patients, resulting in large data
volumes. As a consequence, not only is it difficult to benefit from all available
data in order to solve a particular medical case, it also becomes unfeasible for

a physician to mentally process all the patient data according to current clinical
studies and keep track of new studies at the same time. To alleviate this situation,
we have presented a concept to support clinical decision making using holistic
data analysis [3]. This paper shows the realization of this vision using a RESTful
architecture for medical decision support systems, which supports physicians at a
decision. In particular, we advocate a solution based on formally modeling patient
data with RDF and applying Linked Data principles to publish and interlink
individual records [3]. Furthermore, we incorporate studies by describing them
as formalized rules in RDF.
These rules are interpreted and executed by multiple Cognitive Apps, which
are accessible via a RESTful interface and consume and produce Linked Data.
This provides flexible and adaptive composition of the system. In summary, this
paper makes the following contributions:

1. We describe a rule-based decision system built up from individual Cognitive
Apps.

2. We introduce an exemplary Cognitive App for processing medical guidelines,
with a RESTful interface and described in Linked Data.

3. We provide a specific implementation for a use case scenario and demon-
strate the added value in terms of automatically deduced additional patient
knowledge.

4. We show the suitability of the rule-based decision support system for real-time
scenarios, while dealing with large data volumes.

This paper is structured as follows: the following section introduces our med-
ical scenario. Section 3 describes our approach toward designing a decoupled
REST-based decision support system, while Section 4 provides the specific im-
plementation details. We demonstrate the practical applicability of our solution
by realizing a specific medical use case and evaluating the added value to the
decision support. We evaluate the system in terms of its suitability of supporting
real-time scenarios, in cases where physicians need data usable within intraoper-
ative situations (Section 5). In addition, we will show how many new facts were
generated by applying the inference rules to the patient data. Finally, related
work is described in Section 6, and we summarize our contributions and provide
some conclusions in Section 7.

2 Motivation Scenario
Despite the abundance of medical data, currently, the choice of treatment is
usually not obvious, as it depends on a wide range of factors. In a previous
publication, we defined three concepts [3]:
i) Patient data represents all data that can be acquired for a patient for
whom the treatment plan is prepared. This information can be extracted from

images, laboratory reports or other sources of information (e.g. clinical reports,
hospital databases etc). It can be related to the disease, the organ anatomy and
function or general information (e.g. age, habits etc). ii) Factual knowledge is
written down in quotable sources (e.g. clinical guidelines, studies). This allows
the physician to make predictions about the morbidity and mortality of the
disease and the possible interventions. Guidelines give more specific directions on
treatment options for a specific patient. However, they typically merely give rough
directions, taking into account only a fraction of the patient individual parameters
(e.g. size and number of tumors), while detailed treatment decisions remain to
the surgeon iii) Practical knowledge results from experience. It comprises case
knowledge that encompasses the ability to interpret patient data, form a prognosis
and deduce implications for the treatment, as well as expert knowledge about
treatment options and their respective strengths and weaknesses.
The challenges of diagnosing and providing patient individual treatment
plans, can be summarized as follows: i) collect and integrate patient data so
that it can be processed and interpreted in a unified manner; ii) capture factual
knowledge given in the form of medical guidelines by formalizing it in terms of
rules; iii) compensate for the lack of observing the factual knowledge. We address
all challenges, focusing especially on the last one, by developing an approach
for automatically deducing further patient knowledge in the form of mortality
probabilities, procedure level of suitability, etc. by applying the rules from ii)
on the patient knowledge base i), thus compensating for the lack of observing
factual knowledge iii).

3 Developing a Decision Support System

In the following, we describe in detail our design for realizing a decision support
system capable of supporting this scenario, as well as being flexible enough for
enabling further medical scenarios.
Figure 1 shows a high-level overview of a decision support system for deduc-
ing additional patient knowledge. We adopt a classical three-tier architecture
including - Data tier, Business-logic tier, and Client tier.
Data tier – semantic knowledge base consisting of distributed interlinked
repositories: 1) a central file storage (XNAT), which is used to store data
generated by users and other systems, and make this data accessible within the
knowledge base. 2) a semantic Wiki (Surgipedia), which is the data hub in
the system and allows modelling metadata and linking it to all knowledge base
relevant data instances. For example, Surgipedia contains links to files stored
in XNAT or other external data sources. Furthermore, it provides support for
annotating the medical guidelines with metadata. 3) a repository for storing
Patient data where all patient test results are saved, based on a formally
specified patient model. 4) a repository for storing factual knowledge in the
form of medical guidelines and studies (Rules). Rules are formally defined in N3
format, so that they can directly be applied to the patient data. 5) a repository for
commonly used data models (Onto 1). Data in the knowledge base is published

Fig. 1. High-Level overview of the Decision Support System for Deducing Patient
Knowledge

using the Linked Data principles. In this way, the interoperability between the
components of the system, the integration of new data into the knowledge base
and the ability to interpret this data is realized and guaranteed.

Business-logic tier – implemented via distributed reusable RESTful pro-
cessing blocks in the form of Cognitive Apps. The communication and interaction
between the knowledge base and the Cognitive Apps is based on RESTful in-
terfaces. Cognitive Apps are Web APIs, which also have a semantic description
based on the Linked Data principles. The RESTful architecture and the se-
mantic description allows interpreters like Data-Fu [4] to automatically execute
Cognitive Apps. This particular scenario includes the following Cognitive Apps
- the PatientWrapper automatically converts the patient test results into a
shared patient model and stores them in the knowledge base. The Webformular
provides a semi-automated support for converting the medical guidelines into
formalised rules in RDF. It takes as input the published studies and the manually
composed guidelines via an input GUI and generates corresponding rules. Finally,
MedDeduction uses the patient data and the rules in order to automatically
generate new patient knowledge via deduction and thus provide more data for
supporting a better decision.

Client tier – implemented via individual Use Case Apps supporting specific
physician decision tasks. For this particular scenario, we have implemented a
Use Case App for inputting medical guidelines, in order to assist the process of
extracting formalized rules. Another Use Case App provides a user interface for
displaying the deduced patient knowledge, including predictions for the morbidity
and mortality of a certain intervention (Hepatectomy).

Fig. 2. Overview of the steps of each component in the business tier

4 Patient Knowledge Deductive System

In this section we describe the implementation of the decision support system,
which consists of four Cognitive Apps (PatientWrapper, MedDeduction, Patient-
Generator and WebFormular) and the Knowledge Base. These components of
the system are implemented in Java. The results of the Cognitive Apps are
fed back into the knowledge base. The Apps run on an Apache Tomcat 7 Web
Server. The processing steps are executed by performing a HTTP POST method
and submitting the corresponding parameters as parameter queries. Only the
WebFormular is not RESTful due to the higher number of parameters that are
transferred during processing of the inference rules. An overview of the steps of
each component in the business tier is given in Figure 2.
The WebFormular1 is a HTML page for entering the medical guidelines in
the form of rules. It is divided in two parts - the condition part of the inference
rule and the conclusions. An example rule is {hasHepatectomy ∧ hasHCC ∧
Albumin < 40} → {Death due to progression after 1 Year is 0.1336}
(Rule ID: 423). The preprocessing of the entered conditions and implications
will be done within a .jsp file. A servlet converts the preprocessed data into the
inference rule in N3 format and provides it for downloading. Thereby the user
receives the inference rule in RDF and proceeds to upload it into Surgipedia and
link it to a wiki page. In addition, the user can enter metadata about the rule.
In this way medical guidelines from publications can be stored and described in
the knowledge base.
Due to privacy constraints real patient data cannot be used for testing our
system implementation. Evaluation has been performed in [3] with a smaller
1
https://github.com/TobiasWeller/Webformular

amount of real patient data. This paper validates the technical concept on a large
scale by using a large number of patients. Therefore, the PatientGenerator2 ,
which is a Cognitive App that generates patient data for testing the correctness
and performance of the implementation, as well as for evaluating the knowledge
deduction process helps by generating a large number of patients. For example,
we generated the Patient 4.899960.7 with an Albumin value of 32.1483 g/dL. The
parameters for this patient fit one of the rules. So later on, it will be deduced
that this patient has Hepatectomy and HCC.
The PatientWrapper3 performs a HTTP POST request and takes as input
the patient data (in a spreadsheet format) and the predefined patient model. The
patient data is then transformed into RDF and uploaded to the knowledge base.
Once the rules and the patient data are in the system, we can deduce additional
medical knowledge. In order to do this, the MedDeduction4 retrieves the rules
from the knowledge base, checks if patients matches the conditions of the rules,
and inserts the corresponding deduced new triples. In total it provides four
functions. The first function is for deducing all inference rules for all patients.
The second deduces one inference rule for all patients. The third tries to deduce
all inference rules for one patient. The fourth is for deducing one inference rule
for one patient. The Cognitive App can be executed by performing a HTTP POST
request and transmitting the corresponding parameters. The HTTP POST request
for executing the deducing process is the following:
curl POST http://aifb-ls3-vm2.aifb.kit.edu:8080/MedDeduction/
Executer/AllRuleAllPatient
Our approach towards implementing the decision support system provides a very
flexible distributed solution, where new processing components, i.e. Cognitive
Apps, can be integrated on demand, while alternatives providing the same
functionality can be used to optimize the results.

5 Evaluation
We evaluate the implemented decision support system based on two criteria -
knowledge deduction and performance evaluation. For the knowledge deduction
we show that there was new knowledge deduced by applying the formalized rules.
For the performance evaluation we show that the system supports real-time
scenarios and we compare the results to a local execution of the system. In
order to conduct the experiments, we generated 1,000 patients with the help of
the Patient Generator. The values for the corresponding factors were randomly
generated according to a given range. We used 60 rules, which were taken from
studies, and converted with the help of the Webformular.
Knowledge Deduction In total there were 18,444 new facts generated or
rather 129,108 new triples (one fact consists of 7 triples). There were no new
patients generated. There was no new knowledge deduced for 4 out of the 1,000
2
https://github.com/TobiasWeller/Patientgenerator
3
https://github.com/TobiasWeller/PatientWrapper
4
https://github.com/TobiasWeller/MedDeducter

Experiment Total Runtime Min Time p. Rule Max Time p. Rule Avg. Time
p. Rule
Local 20.575 min 0.337 sec (0 Facts) 80.31 sec (855 Facts) 0.067 sec
Remote 3.843 min 0.09 sec (0 Facts) 13.066 sec (0 Facts) 0.013 sec

Table 1. Evaluation Local vs. Remote Deployment

patients, because no rule fit to these patients. The rule that generated most new
facts was: {MilanCriteriaFulfilled = True} → {hasHepatectomy = True}
In total this rule took 9.141 seconds on the server and was performed for 861
Patients. The rule that took the shortest time was:
{ColorectalCarcinoma = True ∧ hasHepatectomy = True} →
{Probability of 2 Year Disease Free Survival = 0.37}
In total, this rule took 0.09 seconds on the server and did not result in new facts
or triples. We looked up among others observations created by these two rules to
comprehend the results. The deduction was performed correctly and did not lead
to flawed observations.
Performance Evaluation We compared the performance between execution
on a local and on a remote system. The first experiment runs the deduction
process locally on a workstation and communicates with the knowledge base
over the Web (Local). The second experiment runs the deduction process on a
remote virtual machine (Remote). However, the virtual machine is on the same
system as the knowledge base. The local machine has 4 Cores at 2.93 Ghz with 8
GB memory. The virtual machine has 4 Cores at 2.6Ghz with 16 GB memory.
We compared for each rule the number of new facts that were generated.
Both, the local and the remote experiment used the same inference rules and
patient data, and produced the same number of facts. All results were valid and
the number of facts, generated by both experiments, were the same. Naturally,
the runtimes were different. Table 1 contains the measured results. We measured
the total runtime for deducing all 60 rules, the shortest time for deducing a rule
and the longest time, as well as the average time for deducing one rule.
It can be clearly seen that the remote system has and advantage against the
local solution. However, this is based on the fact that the deductive system runs
on the same machine as the knowledge base. Therefore, no long transmission time
is needed to transfer the data. In three cases, the experiment on the remote server
took longer than the local experiment. However, these rules had not generated
new observations.
In total 18,444 new observations were created during a total runtime of 3.843
min for the remote experiment. This leads to 0.013 facts and rule per second.
Since in average the deduction of a fact for one rule takes 0.013 seconds, we can
assume that for 1,000 rules the deduction for one patient sums up to 13 seconds.
This makes the decision support system suitable for real-time use cases, where
the physician can immediately take the additional knowledge into consideration
for making a decision and planning an appropriate treatment.

6 Related Work

IBM Research developed in collaboration with the Cleveland Clinic Lerner College
of Medicine of Case Western Reserve University a medical domain expert system.
This cognitive computing technology, called WatsonPaths, supports clinical
reasoning by exploring a complex scenario and drawing conclusions. It pulls its
knowledge from reference materials, clinical guidelines and medical journals in
real-time. On the base of this knowledge, it disproves a set of hypotheses to
generate new factors in order to support diagnosis and treatment options.5
Another development is the FM-Ultranet [6, 5]. This is a decision support
system using case-based reasoning in the ultrasonography sector. This decision
support system exploits image analysis and pattern recognition techniques to
improve and train ultrasound scans interpretation and diagnosis of foetus malfor-
mations and abnormalities. Thereby it uses past similar ultrasound scans, stored
in the database, for interpreting and diagnosing the actual scans.
CARE-PARTNER is a computerised medical knowledge-support assistance
that offers its functionality on the web [2]. It proposes case-based and rules-
based reasoning and information retrieval methods to provide useful knowledge
to physicians. The system is implemented on the concept of evidence-based
medical practice.

7 Conclusions

The increasing digitalization of medical data calls for new solutions that support
physicians in planning a treatment strategy. To this end, we introduced a rule-
based medical decision support system based on a decoupled distributed RESTful
architecture. We combined REST with Linked Data to present a novel approach
that has not been previously tested in the medical domain. The system consists
of multiple Cognitive Apps that process the medical data. The so implemented
system successfully derives additional knowledge about patients, thus assisting
the physicians in making decisions. As shown in the valuation section, the system
is also suitable for real-time scenarios.
The future work includes the integration of further information sources in order
to enlarge the number of inference rules. A significant contribution, therefore,
would be the automated extraction of inference rules from studies and guidelines.
Acknowledgments This work was carried out with the support of the
German Research Foundation (DFG) as part of project A02, I01, and S01,
SFB/TRR 125 Cognition-Guided Surgery. We would particularly like to thank
Patrick Philipp, Mohammadreza Hafezi, Arianeb Mehrabi and Marco Nolden.
All of the authors state no conflict of interests. All studies have been approved
and performed in accordance with ethical standards. Patient data were gathered
and evaluated under informed consent only.
5
http://www.research.ibm.com/cognitive-computing/watson/watsonpaths.
shtml

References
1. De Clercq, P. and Kaiser, K. and Hasman, A. Computer-interpretable guideline
formalisms. Studies in Health Technology and Informatics, 139: pages 2243, 2008.
2. Bichindaritz, I. and Kansu, E. and Sullivan, K. M. Case-based reasoning in care-
partner: Gathering evidence for evidence-based medical practice. In German Work-
shop on Experience Management. GWEM 03, pages 334-345, June 2006.
3. März, K. and Hafezi, M. and Weller, T. and Saffari, A. and Nolden, M. and Fard, N.
and Majlesara, A. and Zelzer, S.and Maleshkova M. and Volovyk M. and Gharabaghi
N. and Wagner M. and Emami G. and Engelhardt S. and Fetzer A. and Kenngott H.
and Rezai N. and Rettinger A. and Studer R. and Mehrabi A. and Maier-Hein L.
Toward knowledge-based liver surgery: holistic information processing for surgical
decision support. International Journal of Computer Assisted Radiology and Surgery,
10: April, 2015.
4. Stadtmüller, S. and Speiser, S. and Harth, A. and Studer, R. Data-Fu: A Language
and an Interpreter for Interaction with Read/Write Linked Data. International World
Wide Web Conferences Steering Committee: pages 12251236, 2013.
5. Balaa, Z. E. and Strauss, A. and Maximini, K. Fm-ultranet: a decision support
system using case-based reasoning, applied to ultrasonography. In Workshop on CBR
in the Health Sciences, pages 0-3, 2003.
6. Balaa, Z. E. and Traphoener, R. Case-based decision support and experience man-
agement for ultrasonography. In German Workshop on Experience Management.
GWEM 03, pages 277-278, 2003.