=Paper=
{{Paper
|id=Vol-1840/paper_1
|storemode=property
|title=Recommending Doctors and Health Facilities in the HealthNet Social Network
|pdfUrl=https://ceur-ws.org/Vol-1840/paper_1.pdf
|volume=Vol-1840
|authors=Fedelucio Narducci,Cataldo Musto,Marco Polignano,Marco de Gemmis,Pasquale Lops,Giovanni Semeraro
|dblpUrl=https://dblp.org/rec/conf/ecir/NarducciMPGLS17
}}
==Recommending Doctors and Health Facilities in the HealthNet Social Network==
https://ceur-ws.org/Vol-1840/paper_1.pdf
Recommending doctors and health facilities
in the HealthNet Social Network
Fedelucio Narducci, Cataldo Musto, Marco Polignano
Marco de Gemmis, Pasquale Lops, Giovanni Semeraro
Department of Computer Science
University of Bari Aldo Moro, Italy
name.surname@uniba.it
Abstract. In this paper we present HealthNet (HN), a social network
that helps patients to meet the best doctor for her health condition.
The core component of HN is a recommender system that suggests to
the user patients similar to her, and generates suggestions about doctors
and hospitals that best match her patient profile. Currently an alpha
version of HN is available only for Italian users, but in the next future
we want to extend the platform to other languages. We organized three
focus groups with patients, practitioners, and health organizations in
order to obtain comments and suggestions. All were very enthusiastic by
using the prototype version of HN1 .
Keywords: e-health, social network, recommender system, smart health
1 Introduction and Motivations
The e-health, defined as the healthcare practice supported by electronic process
and communication [3], is changing the landscape of clinical practice and health
care. A recent survey demonstrated that 72% of U.S. Internet users looked online
for health information within the past years2 . Similarly, in Italy, 84% of young
people between 18 and 35 years old use the Web for looking for health infor-
mation3 . This new trend can be defined as an evolution of the word of mouth
that generally characterizes the process that firstly conducts a patient to find a
solution to her condition. Indeed, in the same above mentioned survey 60% of
U.S. adults got information or support from friends and family when they have
a health problem and 24% of adults got information or support from other who
have the same health condition. Furthermore, the Associated Press-NORC Cen-
ter for Public Affairs Research in a 2014 survey founds 4 in 10 American people
saw information on a ratings website such as HealthGrades.com, Yelp.com, or
Angies List as a decisive factor in deciding on a particular doctor4 . To share
1
An English demo of HN is available at http://193.204.187.192:8080/HealthNetVideo/
2
http://www.pewinternet.org/2013/01/15/health-online-2013/
3
http://it.ejo.ch/tag/eikon-strategic-consulting
4
http://www.apnorc.org/projects/Pages/HTML%20Reports/finding-quality-
doctors.aspx
2 F. Narducci et al.
health information generates a more informed and empowered patient by recon-
figuring the patient/care team relationship towards a patient-centered medicine.
One of the most relevant initiative in that direction is the U.S. social network
PatientsLikeMe (PLM)5 . This social network enables patients to share, com-
pare and contrast different diagnoses and treatments with people who have the
same conditions who are anywhere in the world [1]. PLM counts 300,000 pa-
tients sharing 2,300 different conditions. In addition to PLM, there are a lot of
forums, blogs, and more generally web sites, that deal with health problems.
However, the information available in these websites is often confused, difficult
to understand and can lead to easy self-diagnosis often wrong [4].
In this paper we present HealthNet (HN), a social network whose main goal
is to help a user in finding a solution for her health conditions. The main idea
behind HN is the same as PLM: sharing knowledge, finding similar patients,
comparing their experiences. However, the main difference with respect to PLM
is the embedded recommender system that is able not only to discover simi-
larities between patients, but also to exploit the data coming from the patient
community for suggesting practitioners and hospitals that best fit a patient pro-
file. In this way, HN deters the self-diagnosis since it just helps the patient to
find a doctor or a health facility. Differently from a classical recommender system
which generally builds a user profiles in order to suggest items potentially inter-
esting for a given user [2, 7, 8], in this work the user profile is the patient health
status, and the doctors and health facilities are the goal of the recommendation.
Other Health Related Recommender Systems (HRRS) are presented in the
literature [12]. In [13] the authors compare content-based and collaborative rec-
ommendation techniques for developing a web-based recommender system for
suggesting relevant websites related to prostate cancer. Khan et al. [5] devel-
oped a tag-based recommendation engine for suggesting users having similar
health profiles, relevant information resources such as articles or blogs on health
promotion, and community resources such as local health facilities. In [10] a
health recommender system using rough sets, survival analysis approaches and
rule-based expert systems is proposed. The goal is to suggest clinical exami-
nations in the case, for example, the patient can afford only a limited number
of tests which have to be ranked according to their priority. Roitman et al. in
[11] defined a HRRS that combines Personal Health Records (PHRs) and drugs
information for avoiding adverse drug reactions. In [6] Morrell and Kerschberg
described a system which uses an agent based framework to retrieve content
from web resources related to an individual’s PHR entries. Even though seman-
tic techniques are already used in the research community for retrieval [9] and
recommendation tasks, to the best of our knowledge the HRRS implemented
in HN is the first one able to suggest doctors and hospitals by performing a
semantic matching between patient profiles.
The rest of this paper is organized as follows. Section 2 describes the platform
and its general architecture, and Section 3 draws conclusions and future work.
5
http://www.patientslikeme.com
Recommending doctors and health facilities in the HealthNet Social Network 3
Fig. 1: General Architecture of HealthNet
2 The HealthNet social network
HN is implemented as a standard social network where users are patients. The
first step of the interaction with the system is the user registration. After that,
the patient can insert her medical data. More specifically, the patient can insert
information about: conditions, treatments (e.g., drugs, dosages, side effects, surg-
eries), health indicators (e.g., blood pressure, body weight, laboratory analysis,
etc.), consulted doctors, hospitalizations. In this way, HN centralizes individual’s
health data allowing a simple and organized access to them. Furthermore, the
user can take advantage from sharing her data by obtaining suggestions in terms
of doctors, health facilities, and other resources useful for her conditions. In or-
der to receive recommendations the user should insert at least one condition she
is affected by. For each condition, the user can click on the ”How can I cure it?”
button and receives suggestions. It is worth noting that the HN user can decide
to be anonymous, by indicating only a nickname during the registration step.
Accordingly, the health data inserted in HN are not linkable to a real identity
thus preserving the user privacy.
In Figure 1 a general architecture of the platform is depicted. There are three
main components: the Profile Manager, the Social Manager, the Recommender
System. The interaction with the system occurs through a Web GUI.
4 F. Narducci et al.
2.1 Profile Manager
This component manages all information related to the patient state of health.
The Profile Manager allows a user to decide which data she wants to share
with the community and which data wants to maintain private. The component
manages also information about consulted doctors, hospitalizations, success or
failure of therapies/treatments, monitoring. The Profile Manager stores these
information in a patient profile which is exploited by the Recommender System
for generating suggestions. In the actual version of the system the patient profile
is composed of two distinct dimensions: conditions and treatments. These are
the only information exploited by the system for computing patients similarities.
2.2 Social Manager
This component manages the activities related to the social network interactions.
It allows one to establish relations between patients by means of friendship con-
nections. Friends share their updates, information about drugs, most common
conditions or symptoms they are used for, side effects, dosages. The Social Man-
ager also manages the Health Point of Interests (HPOIs). A HPOI is a point of
interest which offers services related to the health domain useful for the HN com-
munity. A HPOI can be an association which offers home care, or an organization
which supports patient relatives, for example. The association or organization
which desires to have a page on HN for adrvetising its activities, can sign up on
the platform and indicate the conditions for which offers services or facilities.
The HPOIs are then suggested by performing a simple matching between the
relevant conditions defined by the HPOI owner and the patient condition.
2.3 Recommender System
This is the core component of the system. The Recommender System exploits
the profile of the patients to suggest other similar patients as well as the best
doctors or hospitals according to a given patient profile. The similarity between
patients is computed in terms of shared conditions and shared treatments. The
component computes a semantic matching between the conditions by exploring
the hierarchy of diseases 6 . The idea is that a patient with prostate cancer
and another with testicular cancer, for example, should have a high similarity
score since both conditions affects organs belonging to the male reproductive
system (in the disease hierarchy). Hence, even though the two patients do not
have the same condition can share useful experiences. Similarly, the matching
between treatments takes into account not only their names but also the active
ingredients (for drugs), and organ it affects (for surgeries). The recommender
system combines the similarity score deriving from the patient conditions and the
score deriving from the patient treatments to compute a similarity score between
two patients. Suggestions of doctors and hospitals are thus ranked by analyzing
6
http://apps.who.int/classifications/icd10/browse/2010/en
Recommending doctors and health facilities in the HealthNet Social Network 5
doctors and hospitals consulted by the most similar patients. Furthermore, the
ranking takes also into account official quality indicators generated by the Italian
Ministry of Health7 .
More formally, the similarity score between two patients is computed as fol-
lows:
Pk Pn 0
0 i=1 j=1 sc (pci , pcj )
s(p, p ) = α +
k+n (1)
Pz Pr 0
i=1 j=1 st (pti , ptj )
+(1 − α) ,
z+r
where k (respectively n) is the number of conditions p (respectively p0 ) is affected
by, pc is a condition of the patient p, z (respectively r) is the number of treat-
ments for p (respectively p0 ), pt is a treatment for the patient p, sc (pci , p0cj ) is
the condition similarity between ci , and cj , st (pti , p0tj ) is the treatment similarity
between ti , and tj , computed as follows:
#C
(
1 ∗ log #P , if ci = cj
sc (pci , p0cj ) = 1
c i (2)
sp(ci ,cj ) , otherwise
(
1, if ti = tj
st (pti , p0tj ) = (3)
0, otherwise
If the two conditions are the same, the similarity score sc is equal to a weight
computed as the ratio between the number of conditions in the database (#C)
and the number of patients affected by that condition (#P ci ). The goal of this
additional weight is to give higher similarity to patients which share rare diseases.
If the two conditions are different the score sc is computed as the reciprocal
of the length (number of edges) of the shortest path sp which connects the
two conditions in the disease hierarchy. More simply, the treatment similarity
is 1 when the treatments are the same or (for drugs) have the same active
ingredient, 0 otherwise. Treatment similarity and condition similarity score can
differently contribute to the patient similarity score by changing the α value.
The patient similarity is thus used for ranking the list of suggested doctors
and hospitals. Doctors and hospitals are ranked according to the scoreDoc and
scoreH. The scoreDoc for the doctor dz and the patient pi is computed by
multiplying for each patient pj in the database the similarity score with pi and
the rating assigned by pj to the doctor dz . Similarly, the scoreH, takes into
account the similarity patient, the user rating rj for a given hospital hm and
a quality indicator produced by the Italian Health Ministry for each Italian
hospital8 . The community indicator and the ministry indicator can be differently
weighted by changing the β value.
7
For each hospital the number of admissions and the number of deaths are reported
for a given treatment. http://95.110.213.190/PNEed13/
8
http://95.110.213.190/PNEed13/
6 F. Narducci et al.
p
X
scoreDoc(dz , pi ) = s(pi , pj ) ∗ rj (dz ) (4)
j=1
p
X
scoreH(hm , pi ) = β s(pi , pj ) ∗ rj (hm ) (1 − β)qi(hm ) (5)
j=1
3 Conclusion and Future Work
In this paper we presented HealthNet, a health social network available in al-
pha version for Italian users. HN suggests doctors or health facilities for a given
patient condition by using experiences shared from the patient community. The
recommender system implements a semantic matching able to compute a simi-
larity also between patients which do not share the exact same condition. We are
building a dataset in order to test our algorithm trough an in-vitro experimen-
tal evaluation. Subsequently we intend to perform a case study with real users.
In the future work we want to extend the system to other languages, evaluate
different similarity measures and allow user to export and share (e.g, with her
practitioners) all her health data.
References
1. M. Al-Ubaydli. How social networks enable patients to be more involved in their
healthcare. The Guardian, April 2012.
2. M. De Gemmis, L. Iaquinta, P. Lops, C. Musto, F. Narducci, and G. Semeraro.
Learning preference models in recommender systems. In Preference Learning, pages
387–407. Springer, 2010.
3. V. Della Mea. What is e-health (2): the death of telemedicine? Journal of Medical
Internet Research, 3(2):e22, 2001.
4. B. W. Hesse, D. E. Nelson, G. L. Kreps, R. T. Croyle, N. K. Arora, B. K. Rimer,
and K. Viswanath. Trust and sources of health information. Archives of Internal
Medicine, 165(22):2618–2624, 2005.
5. S. Khan, A. Cohall, and R. Kukafka. A tag based recommendation engine to
suggest information resources in an online community for health promotion. In
Annual Symposium proceedings. AMIA Symposium, pages 1002–1002, 2007.
6. T. G. Morrell and L. Kerschberg. Personal health explorer: A semantic health
recommendation system. In Data Eng. Workshops (ICDEW), 2012, pages 55–59.
IEEE, 2012.
7. C. Musto, F. Narducci, P. Lops, and M. de Gemmis. Combining collaborative and
content-based techniques for tag recommendation. In International Conference on
Electronic Commerce and Web Technologies, pages 13–23. Springer, 2010.
8. C. Musto, G. Semeraro, P. Lops, M. De Gemmis, and F. Narducci. Leveraging so-
cial media sources to generate personalized music playlists. In International Con-
ference on Electronic Commerce and Web Technologies, pages 112–123. Springer,
2012.
Recommending doctors and health facilities in the HealthNet Social Network 7
9. F. Narducci, M. Palmonari, and G. Semeraro. Cross-language semantic retrieval
and linking of e-gov services. In International Semantic Web Conference, pages
130–145. Springer, 2013.
10. P. Pattaraintakorn, G. M. Zaverucha, and N. Cercone. Web based health recom-
mender system using rough sets, survival analysis and rule-based expert systems.
In Rough Sets, Fuzzy Sets, Data Mining and Granular Computing, pages 491–499.
Springer, 2007.
11. H. Roitman, Y. Messika, Y. Tsimerman, and Y. Maman. Increasing patient safety
using explanation-driven personalized content recommendation. In Proc. of the 1st
ACM Int. Health Informatics Symposium, pages 430–434. ACM, 2010.
12. M. Wiesner and D. Pfeifer. Health recommender systems: Concepts, requirements,
technical basics and challenges. Int. J. of Environmental Research and Public
Health, 11(3):2580–2607, 2014.
13. H. Witteman, M. Chignell, and M. Krahn. A recommender system for prostate can-
cer websites. In Annual Symposium proceedings. AMIA Symposium, pages 1177–
1177, 2007.