=Paper=
{{Paper
|id=Vol-2164/paper5
|storemode=property
|title=The FairGRecs Dataset: A Dataset for Producing Health-related Recommendations
|pdfUrl=https://ceur-ws.org/Vol-2164/paper5.pdf
|volume=Vol-2164
|authors=Maria Stratigi,Haridimos Kondylakis,Kostas Stefanidis
|dblpUrl=https://dblp.org/rec/conf/semweb/StratigiKS18
}}
==The FairGRecs Dataset: A Dataset for Producing Health-related Recommendations==
https://ceur-ws.org/Vol-2164/paper5.pdf
The FairGRecs Dataset: A Dataset for Producing
Health-related Recommendations
Maria Stratigi1 , Haridimos Kondylakis2 , and Kostas Stefanidis3
1
University of Tampere, Tampere, Finland, Maria.Stratigi@uta.fi
2
ICS-FORTH, Heraklion, Greece, kondylak@ics.forth.gr
3
University of Tampere, Tampere, Finland, kostas.stefanidis@uta.fi
Abstract. Nowadays, the number of people who search for information related
to health online has significantly increased, while the time of health profession-
als for recommending useful sources online has been reduced to a great extend.
As such, providing valuable information to users for health related issues, based
on their personal health profiles, in the form of suggestions, approved by their
caregivers, can significantly improve the opportunities that users have to inform
themselves online about health problems and possible treatments. However, due
to several legal and ethical constraints, personal health profiles usually are not ac-
cessible. In this paper, we present FairGRecs, a synthetic dataset that can be used
for evaluating and benchmarking methods that produce recommendations related
to health documents based on individual health records. Specifically, FairGRecs
can create, via a fully parametrized API, synthetic patients profiles, containing
the same characteristics that exist in a real medical database, including both in-
formation about health problems and also relevant documents.
1 Introduction
Medicine is undergoing a revolution that is transforming the nature of health-care from
reactive to preventive. One of these challenges in this revolution is the problem of the
quality and the amount of information that can be found online [1], since health infor-
mation is one of the most frequently searched topics on the Web. Especially during the
last decade, the number of users who look online for health and medical information
has dramatically increased. However despite the increase in those numbers and the vast
amount of information currently available online, it is very hard for a patient to accu-
rately judge the relevance of some information to his/her own case and to identify the
quality of the provided information.
Furthermore, the optimal solution for patients is to be guided by healthcare providers
to appropriate sources of information [1], [13]. Delivering accurate information sources
to a patient, increases his/her knowledge and changes the way of thinking, which is
usually referred as patient empowerment [7], [8]. As a result, the patient’s dependency
for information from the doctor is reduced. Also, patients feel autonomous and more
confident about the management of their disease [16]. To this direction, health providers
have the history of their patient’s and their interests, in order to make an informed de-
cision about the information that would likely be beneficial for them. However, health
providers have less and less time to devote to their patients. As such, guiding each
individual patient appropriately is a really difficult task.
On the other hand, the use of group-dynamics-based principles of behavior change
have been shown to be highly effective in enhancing social support through promoting
group cohesion in physical activity [2], in reducing smoking relapse [3] and in pro-
moting healthy dietary habbits [10]. In small groups, therapy sessions enjoy a social
component as participants can share experiences and discussion. In those therapy ses-
sions, a caregiver can guide patients to more optimal resources over the Web. However,
if identifying online information content for a single patient is a difficult task, identify-
ing information for a group of participants is a really challenging one.
To this direction, in our work [14, 15], we focused on recommending interesting
health documents, to groups of users. Our motivation was to offer a list of recommen-
dations to a caregiver who is responsible for a group of patients. The recommended
documents need to be relevant, based on the patients current profile, namely by exploit-
ing the patients personal healthcare record (PHR) data. However, although it is really
common for patients to look for health information and sometimes to rate related doc-
uments on the Web, their profiles are usually not accessible, neither linked to those
documents. Among others, legal and ethical constraints prohibit the collection and the
exploitation of such a dataset.
This way, in this paper, we present a synthetic dataset, FairGRecs, that can be used
for evaluating and benchmarking methods that produce recommendations related to
health documents. More specifically, we rely on the EMRBots dataset4 , which contains
synthetic patients profiles, containing the same characteristics that exist in a real medi-
cal database, such as patients admission details, demographics, socioeconomic details,
labs and medications, extending it with a document corpus and a rating dataset. By ex-
ploiting the FairGRecs dataset, interested users can create patients that have provided
rankings for health documents. To link document contents with patients, we use the
ICD105 ontology, namely the International Statistical Classification of Diseases and
Related Health Problems, which is a standard medical classification list maintained by
the World Health Organization. FairGRecs is fully parametrized and is offered via an
API.6
This dataset has been used already for optimizing the Personal Health Informa-
tion Recommender (PHIR) [5], [8], [9], developed within the EU project iManage-
Cancer [6]. PHIR is a recommendation engine for recommending high quality cancer
documents selected by health providers to patients.
The rest of this paper is structured as follows. Section 2 introduces our synthetic
dataset, consisting of patients data, a document corpus and a ratings dataset. We also
include a dataset creation example to showcase how to build these datasets. Section 3
describes the developed application programming interface, while Section 4 concludes
with a discussion on the usefulness of the produced dataset.
4
http://www.emrbots.org
5
http://www.icd10data.com/
6
https://bitbucket.org/MariaStratigi/fairgrecs-dataset/overview
2 The FairGRecs Dataset
In general, a recommender system requires access to a certain amount of information
– set of medicinal documents, a dataset containing ratings that the users have given to
those documents and the personal health information of the users, to make suggestions
to users. The absence of real data, motivated us to develop a fully parameterized tool
that automatically creates the necessary datasets, that we do not have access to.
The first major obstacle is the acquisition of the dataset containing the personal
health information of the users. Among others, legal and ethical constraints prohibit the
collection and the exploitation of such a dataset. A further constraint is that the health
information dataset has to be linked to a document corpus via a users’ ratings dataset.
So the problem we face is the need of three interlinked datasets, that are void of any
legal and ethical constrains. The first step to overcome this obstacle, is the adoption of
the 10.000 chimeric patient profiles provided by EMRBots.
2.1 Patient Profiles Dataset
Unlike other methods that typically obscure or shift real patients’ data elements, the
EMRBots proposed methodology is invulnerable in terms of security, because it does
not rely on real data elements pulled from an existing Electronic Medical Record (EMR);
therefore, it is not associated with privacy concerns in regard to individuals’ sensitive
data.
The data is generated according to pre-defined criteria and is not based on any hu-
man data. These criteria are divided into Population-level and Patient-level character-
istics. The first group offers an array of values in order to define demographics, such
as gender, marital status, major language, ethnicity, date of birth and income level. Af-
ter the completion of the Population-level configuration, there are n patients generated.
For each such patient, the Patient-level configuration associates with them additional
details. These details include randomly generated length of stay (in days) and start and
end dates. Furthermore, each admission is associated with laboratory measurements and
a chief complaint randomly selected from a list of International Classification of Dis-
eases, Tenth Revision, Clinical Modification (ICD-10-CM) codes. In the last step of the
patient-level configuration, laboratory values are added. Laboratory values are based on
35 common types - for instance, sodium levels, creatinine levels, or platelet count.
2.2 Document Corpus & User Preferences Dataset
One of the most important information a recommender system requires is a dataset
containing the users preferences towards a set of items [12]. These preferences can
take many formats, such as ratings, check-ins or textual reviews [11]. For this work,
our chosen format is that of a ratings dataset. Most specifically, we produce a dataset
containing ratings associated with users for particular health related documents.
Based on the patients’ health profiles dataset that we have already acquired, we
generate two new datasets. The first is a document corpus that consist of documents’
id codes and their corresponding keywords, and the second is the ratings dataset, that
incorporates the ratings given by the patients to the documents.
The generation process for both of these datasets, is fully parameterized. It is worth
mentioning that, because our main focus is the development of the ratings dataset, heavy
emphasis is given to the health profile of the patients. This is mostly presented as the
health problems of each patient. As mentioned in the previous section this information
is documented using ICD-10 codes.
Document Corpus. The generation of the document corpus, includes a document id
and the corresponding keywords for each created document. As such we are not gener-
ating full text documents. To achieve that, we take advantage of the ICD-10 ontology
tree. We generated a numDocs number of documents, for each second level category
of the ICD-10 ontology (i.e., for each node that belongs in the second level of the on-
tology tree). We will call these primary nodes. The ICD-10 ontology tree has 295 such
nodes. Here, we make the assumption that a document cannot be related to more than
one category. For example, a document cannot impart information for pregnancy and
the perinatal period, cause these areas are represented by two different primary nodes
in the tree.
For the documents’ corresponding keywords, we randomly selected numKeyW ords
words from the description text of the nodes in each subsequent subtree of the primary
node that the document belongs. From those descriptions, we removed the most com-
mon words – often called stopwords, such as “the”, “and”,“it” and made our selection
from the rest. A visual description is provided in Figure 1.
Fig. 1. The created documents will be anchored to the primary nodes, while their keywords will
be selected from the subtree.
When ranking items based on human preferences, the most common observed dis-
tribution of these ratings is the power law distribution [4]. In order to make our dataset as
plausible as possible, we depict this distribution by randomly selecting a popularDocs
number of documents that will be the most popular. These documents span through all
the categories and are selected arbitrarily.
Ratings dataset. In order to create the ratings dataset, first we assume that all the
patients have given at least a minimum non zero number of ratings. This assumption was
made in order to make our datasets easier to work within the domain of recommender
systems, and more specifically the systems that operate under a collaborative filtering
design. In general these systems operate by finding similar users to a given user, and
extrapolating an item’s rating based on the scores given by those users [14, 15]. If a
user has not given any ratings, we will stumble upon the cold start problem, where we
cannot find any similar users to him/her and subsequently, we will not be able to provide
him/her with any recommendations.
In order to avoid this problem, we have surmised that all patients have given a
numRatings number of ratings. Specifically, we have divided the patients into three
groups – occasional, regular and dedicated. The users in each group have given low,
average and high number of ratings, respectively. The number of ratings are randomly
selected from a numerical range, based on which of the three groups the patient belongs.
Based on that number, we will create a corresponding number of rating nodes for
each user. A rating node links a user to a document, i.e., the user has rated that docu-
ment. A user cannot have more than one rating node for the same document, but can
have many nodes for different documents that belong in the same category.
We have divided the user’s rating nodes into two groups; healthRelevant and
nonRelevant. Using the health problems data (noted in ICD-10 nodes) of each user,
the first group of ratings will go to documents belonging to the same subtree as one of
their health problems, while the second group will be randomly assigned to the rest of
the documents. Our assumption here is that the patients will be interested not only in
documents regarding their health problems, but also to some extent in others as well.
Finally, in the last step, we assign rating values for each rating node that we gener-
ated previously. We choose the nodes randomly, and assign to them a value in the range
of 1 to 5. The user is able to define the total number of ratings with a specific value that
will be present in the rating dataset. This is accomplished as shown before with the use
of percentages. For example, the administrator can define that the 25% of all ratings
will have the value of 1.
2.3 Datasets Creation Example
In Section 2.2, we analyzed the proposed method of creating two datasets - documents
and ratings - in lieu of real data. In Tables 1 and 2, we present the parameters needed for
creating an example dataset and we briefly explain the values given to them. Further-
more, we selected to use the 10.000 patients chimeric dataset provided by the EMR-
Bots. After all the necessary steps were completed the number of items in the document
corpus was 79.650 and the total number of ratings generated was 1.576.872.
Figure 2 depicts the distribution of ratings in the documents. We partition the ratings
in groups of 50. Most of the documents (71%) have received ratings in the range of [50-
100]. In the second place (21%), we have the documents that have been rated from 0 to
50 times, while if we accumulate all the documents which have been rated more than
200 times, they merely make up of the 1.12% of the corpus. As expected, these results
simulate a power law, where the prominent items are few, and the plethora of documents
have very low popularity.
Table 1. The parameters needed to creating the document corpus.
Parameter Name Explanation Value
numDocs The number of documents created for each different cat- 270
egory of health problems, based on the ICD10 ontology
tree.
numKeyWords The number of randomly selected keywords, attached 10
to each document.
popularDocs The number of documents, that will be most popular in 70
each category, in order to simulate a power law distri-
bution.
Table 2. The parameters needed to create the ratings dataset.
Partitions Parameter Name Explanation Value
Group occasional Users give [20,100] ratings 50% of all patients
Groups
Group regular Users give [100,250] ratings 30% of all patients
Group dedicated Users give [250,500] ratings 20% of all patients
One Ratings valued as 1 20% of all ratings
Two Ratings valued as 2 10% of all ratings
Scores
Three Ratings valued as 3 30% of all ratings
Four Ratings valued as 4 20% of all ratings
Five Ratings valued as 5 20% of all ratings
Ratings
healthRelevant Ratings relevant to health problems 20% of user’s ratings
nonRelevant Ratings not relevant to health problems. 80% of user’s ratings
Fig. 2. The distribution of the ratings in the document corpus, where we partition the number of
ratings in groups of 500.
3 Application & Programming Interface
In order to streamline the creation process of the datasets and make it more user friendly,
we developed an application and the corresponding Application Programming Interface
(API). Those incorporates all the functions described in the previous section. They were
both developed using Java, and they are available online 7 . The application interface,
shown in Figure 3, is divided into four main tabs, each dedicated to a different part of
the datasets creation process.
The first tab called ’File Paths’, is a form for loading the two EMRbots dataset files,
that contain the patients basic information and their health problems. In addition, the
ICD-10 ontology is needed in a xml format, as well as a stopwords file containing the
most common words in the English language.
The second tab is about the document corpus. The user needs to enter a numerical
value, regarding the number of documents that will be created per category (i.e., first
level nodes), and the number of keywords each document will have. The final input
concerns the number of documents that will be popular per category, in order to simulate
a power law distribution.
In the third tab, we have accumulated all the parameterized variables regarding the
patients. These predominantly concern the patients partitioning into groups. Specifi-
cally, the user can divide the patients into the three groups by setting the percentage of
the patients that belong in each group. The user is also able to define the minimum and
maximum number of ratings per different group. Finally, the distribution of the user’s
7
https://bitbucket.org/MariaStratigi/fairgrecs-dataset/overview
Fig. 3. Visualization of the four different tabs of the API. As input, we have entered the default
numbers.
ratings to documents that are relevant to his/her health problems is again accomplished
by the use of percentages.
The last tab concerns the distribution of values to the ratings nodes. As before, the
user defines the percentage of ratings with a specific value. The selection of a document
to assign a value is done randomly.
4 Discussion & Conclusion
To the best of our knowledge, the FairGRecs dataset, is currently the only dataset avail-
able, combining personal health record, documents and ratings, offering a unique op-
portunity for experimenting with recomender systems in the health domain. As such
it fills an important gap in the area of health recomender systems and it reuses and
significantly extends state of the art datasets. Furthermore, is of particular interest to
the Semantic Web Community, as in its core a widely used taxonomy is used, opening
many possibilities for subsequent exploitation through reasoning techniques.
In addition, there is significant evidence of usage by the community of health rec-
ommender systems, as currently more and more platforms emerge enabling patients
to access high quality health related information. Nevertheless, we do not only offer
a specific dataset but also an application and the corresponding API, enabling experi-
mentation with endless possibilities, as well as its wider adoption and extensibility. The
source code is also available allowing further extensions by the community.
References
1. G. M. Berg, A. M. Hervey, D. Atterbury, R. Cook, M. Mosley, R. Grundmeyer, and D. Acuna.
Evaluating the quality of online information about concussions. Journal of the American
Academy of PAs, 27:1547–1896, 2014.
2. I. Brandon, K. Daniel, C. Patrice, and T. Nicholas. Testing the efficacy of ourspace, a brief,
group dynamics-based physical activity intervention: A randomized controlled trial. J Med
Internet Res, 18(4):e87, May 2016.
3. D. Y. T. Cheung, H. C. H. Chan, J. C.-K. Lai, V. W. F. Chan, P. M. Wang, W. H. C. Li, C. S. S.
Chan, and T.-H. Lam. Using whatsapp and facebook online social groups for smoking relapse
prevention for recent quitters: A pilot pragmatic cluster randomized controlled trial. J Med
Internet Res, 17(10):e238, Oct 2015.
4. H. Halpin, V. Robu, and H. Shepherd. The complex dynamics of collaborative tagging. In
Proceedings of the 16th International Conference on World Wide Web, WWW ’07, pages
211–220, New York, NY, USA, 2007. ACM.
5. G. Iatraki, H. Kondylakis, L. Koumakis, M. Chatzimina, K. Marias, and M. Tsiknakis. Per-
sonal health information recommender: implementing a tool for the empowerment of cancer
patients. ecancer 12 851, 2018.
6. H. Kondylakis, A. I. D. Bucur, F. Dong, C. Renzi, A. Manfrinati, N. M. Graf, S. Hoffman,
L. Koumakis, G. Pravettoni, K. Marias, M. Tsiknakis, and S. Kiefer. imanagecancer: De-
veloping a platform for empowering patients and strengthening self-management in cancer
diseases. In IEEE CBMS, pages 755–760, 2017.
7. H. Kondylakis, L. Koumakis, E. Genitsaridi, M. Tsiknakis, K. Marias, G. Pravettoni,
A. Gorini, and K. Mazzocco. Iems: A collaborative environment for patient empowerment.
In 12th IEEE International Conference on Bioinformatics & Bioengineering, BIBE, pages
535–540, 2012.
8. H. Kondylakis, L. Koumakis, E. Kazantzaki, M. Chatzimina, M. Psaraki, K. Marias, and
M. Tsiknakis. Patient empowerment through personal medical recommendations. In MED-
INFO, page 1117, 2015.
9. H. Kondylakis, L. Koumakis, M. Psaraki, G. Troullinou, M. Chatzimina, E. Kazantzaki,
K. Marias, and M. Tsiknakis. Semantically-enabled personal medical information recom-
mender. In ISWC, 2015.
10. J. Meng, W. Peng, Y. S. Shin, and M. Chung. Online self-tracking groups to increase fruit and
vegetable intake: A small-scale study on mechanisms of group effect on behavior change. J
Med Internet Res, 19(3):e63, Mar 2017.
11. E. Ntoutsi and K. Stefanidis. Recommendations beyond the ratings matrix. In Proceedings
of the Workshop on Data-Driven Innovation on the Web, DDI@WebSci 2016, Hannover,
Germany, May 22-25, 2016, pages 2:1–2:5, 2016.
12. E. Ntoutsi, K. Stefanidis, K. Rausch, and H. Kriegel. Strength lies in differences: Diversi-
fying friends for recommendations through subspace clustering. In Proceedings of the 23rd
ACM International Conference on Conference on Information and Knowledge Management,
CIKM 2014, Shanghai, China, November 3-7, 2014, pages 729–738, 2014.
13. S. Schaller, V. Marinova-Schmidt, M. Setzer, H. Kondylakis, L. Griebel, M. Sedlmayr,
E. Graessel, M. J. Maler, S. Kirn, and L. P. Kolominsky-Rabas. Usefulness of a tailored
ehealth service for informal caregivers and professionals in the dementia treatment and care
setting: The ehealthmonitor dementia portal. JMIR Res Protoc, 5(2):e47, Apr 2016.
14. M. Stratigi, H. Kondylakis, and K. Stefanidis. Fairness in group recommendations in the
health domain. In ICDE, 2017.
15. M. Stratigi, H. Kondylakis, and K. Stefanidis. FairGRecs: Fair group recommendations by
exploiting personal health information. In DEXA, 2018.
16. M. Wiesner and D. Pfeifer. Adapting recommender systems to the requirements of personal
health record systems, 2010.