SPECIAL SECTION ON BIG DATA TECHNOLOGY

AND APPLICATIONS IN INTELLIGENT TRANSPORTATION

Received May 7, 2020, accepted May 16, 2020, date of publication June 1, 2020, date of current version July 1, 2020.
Digital Object Identifier 10.1109/ACCESS.2020.2995763

Envisioning Insight-Driven Learning Based on

Thick Data Analytics With Focus on Healthcare
JINAN FIAIDHI , (Senior Member, IEEE)
Department of Computer Science, Lakehead University, Thunder Bay, ON P7B 5E1, Canada
e-mail: [email protected]
This work was supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC) under
Grant DDG-2020-00037, and in part by the Lakehead University.

ABSTRACT Detecting and analyzing patient insights from social media enables healthcare givers to better
understand what patients want and also to identify their pain points. Healthcare institutions cannot neglect
the need to monitor and analyze popular social media outlets such as Twitter and Facebook. To have a study
success, a healthcare giver needs to be able to engage with their patients and adapt to their preferences
effectively. However, data-driven decision-making is no longer enough, as the best-in-class organizations
struggle to realize tangible benefits from their data-driven analytics investments. Relying on simplistic textual
analytics that use big data technologies to learn consumer/patient insights is no longer sufficient as most of
these analytics utilize sort of bag-of-words counting algorithms. The majority of projects utilizing big data
analytics have failed due to the obsession with metrics at the expense of capturing the customer’s perspective
data, as well as the failure in turning consumer insights into actions. Most of the consumer insights
can be captured with qualitative research methods that work with small, even statistically insignificant,
sample sizes. Employing qualitative analytics provide some kind of actionable intelligence which acquires
understanding to broad questions about the consumer needs in tandem with analytical power. Generating
insight, on one hand, requires sound techniques to measure consumers’ engagement more precisely and
offers depth analytics to the consumer data story. On the other hand, turning relevant insights into actions
requires incorporating actionable intelligence across the business by verify hypotheses based on qualitative
findings by using web analytics to see if these axioms apply to a large number of customers. The first
component of our visionary approach is dedicated to identifying the relationships between constituents of
the healthcare pain points as echoed by the social media conversation in terms of sociographic network
where the elements composing these conversations are described as nodes and their interactions as links.
In this part, conversation groups of nodes that are heavily connected will be identified representing what
we call conversation communities. By identifying these conversation communities several consumer hidden
insights can be inferred from using techniques such as visualizing conversation graphs relevant to given
pain point, conversation learning from question answering, conversations summaries, conversation timelines,
conversation anomalies and other conversation pattern learning techniques. These techniques will identify
and learn the patient insights without forgetting from the context of conversation communities, are tagged
as ‘‘thick data analytics’’. Additionally machine learning methods can be used as assistive techniques to
learn from the identified thick data and build models around identified thick data. With the use of transfer
learning we also can fine tune these models with the arrival of new conversations. The author is currently
experimenting with these seven insights driven learning methods described in this paper with massive
geo-located Twitter data to infer the quality of care related to the current COVID-19 outbreak.

INDEX TERMS Patient insights, thick data, learning patient preferences, graph based algorithms,
graph-based machine learning, graph-based transfer learning, Neo4j, twitter conversation communities.

I. INTRODUCTION to collect patient’s feedback from notable blogs (e.g. from

The traditional solution to understanding patient experience PatientsLikeMe.org, Drugs.com) and eliciting patient com-
and insights from social media is the use of web scrapers plaints from governmental systems like the FDA FAERS [1].
However, a more trendy approach is to use Big Data ana-
The associate editor coordinating the review of this manuscript and lytics for learning patients’ insights especially from social
approving it for publication was Dalin Zhang. media [2]. Both techniques have been intensively used in the

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
114998 VOLUME 8, 2020
J. Fiaidhi: Envisioning IDL Based on Thick Data Analytics With Focus on Healthcare

past decade with limited success to capture patient experi- recruiting and retaining participants, and there were many no
ence and explain why patients do what they do [3]. Such validation feedback from the public about the issues raised
approaches prove to be daunting if it is built on prediction by these participatory methods. However, the Delphi method
models that are based on quantitative datasets and methods shows more promising support across the health sector as
which do not incorporate important consumer values and it has been recognized as an optimal method for consensus
insights. According to Pink et al. [4] ‘‘ Big data by nature building, with use of anonymous feedback from an expert
strips away context and cannot bring to light the qualitative panel and statistical analysis techniques to interpret the data.
and ethnographic nature embedded in the consumer narrative The iterative nature of the Delphi process avoids some of
feedback to uncover people’s emotions, stories, and models the pitfalls of other IDL methods, such as the effects of
of their world.’’ Without it, companies are basing important dominant persons or the tendency to conform to a particular
business decisions off incomplete data — like what led to viewpoint [7]. Despite this promising support, Delphi process
Nokia failure in 2009 [5]. To bring context and then consumer suffers from the following weaknesses [8]:
insights is the use of qualitative research based on small sam- • Does not take into account widely differing opinions or
ples and corpuses (‘‘small data’’). It is argued that the unique large changes in public opinions (paradigm shifts [9])
value of such qualitative research lies in data thickness. This • The initiator’s point of view may dominate in the analy-
is achieved through a process that is called thickening or thick sis
data analytics [6]. However, the thick data research did not • Time-consuming
provide enough guides into how to mine healthcare thick data • Requires high participant motivation
patterns from social media. In this research presenting three • The quality of the participants affects the outcomes
strategies to thicken health related narrative data extracted There are more problems and concerns associated with the
from social media like twitter by developing a qualitative conceptualization and meaningful assessment and measure-
exploration methodology that utilizes conversations’ struc- ment of all the current IDL methods, have also been identi-
tures, identification of related components in conversations fied [10]. Each of these concerns and weaknesses points to
and learning from conversations. The qualitative exploratory different directions for methodological research on IDL that
method seeks to define and interpret unclear phenomena involve patient and public involvement with high engaging
through non-numerical methods of measurement that focus degree and empowering learning to have an impact [11].
on meaning and insight. Many healthcare institutions tried to benefit from other busi-
nesses by incorporating added value capabilities on top of
II. HEALTHCARE-BASED INSIGHT-DRIVEN LEARNING the classical IDL methods to provide better predictively and
Insight-driven learning (IDL) for healthcare is all about find- higher acceptability [12]. Figure 2 illustrate these added value
ing methods that caregiver use for patient-centered outcomes techniques on top of the available IDL methods.
prediction. Usually such methods incorporate patient, public
and caregiver views into the process of forming predictions.
However, the current methods used in healthcare involve
direct patient involvement as an integral and transformative
part of the methods processes. Figure 1 lists some of the
notable IDL methods.

FIGURE 2. Added value techniques to enhance IDL methods.

The added value techniques adds an extra layer on top

of the analytics of the IDL methods for identifying con-
sumer insights, however, it proved to be daunting to add
more analytics on top of unsuccessful analytics. Central to
FIGURE 1. Current patient insights learning methods.
having a successful IDL method is to understand the patient
‘pain points.’ According to a recent survey by Medici [13]
Based on these methods, patient involvement was not reveals widespread patient frustration with the America’s
impactful because these methods was not successful in healthcare system, illustrating why we need to reimagine the

VOLUME 8, 2020 114999

 J. Fiaidhi: Envisioning IDL Based on Thick Data Analytics With Focus on Healthcare

way we design our IDL methods if are heading to a more be used for communicating via email, performing Internet
value-based healthcare systems [14]. IDL methods need to searches, interacting with repositories including Electronic
take the stress out of healthcare, for both caregivers and their Health Records and using specific healthcare applications.
patients. IDL methods need to incorporate the connectiv- The smartphone is one of the fastest growing sectors in the
ity ecosystem available at and around the healthcare with technology industry, and its impact in medicine has already
patients detailing their issues and pain points over popular been significant [16]. Social media channels such as Twitter
blogs as well as utilizing venues of the mobile virtual care are gaining increasing acceptance as mechanisms for instan-
to chat with their doctors, therapists and healthcare profes- taneous scientific dialogue with other caregivers as well as
sionals from their smartphones or tablets. However, learning part of the physician’s medical practice [17], [18].
patient’s pain points covers wide range of diverse issues Our proposed exploratory IDL methods described in this
according to the different prospective of patients. This the section is based on the notion of Thick Data and it meant
reason that justifies conducting qualitative analytics (which to be used by physicians and researchers for a specific or a
focuses on detailed, individualized responses to open-ended narrow area of inquiry to explain everyday lives of patients
questions) as opposed to quantitative analytics (which favors and explores their pain points from social media like Twitter
standardized questions and representative, statistically sig- in order to understand patients set of preferences, attitudes,
nificant sample sizes). Figure 3 lists some of the popular timelines, experiences, opinions, emotions, behavior, context,
patient pain points related to their journey in healthcare. social dynamics and sensory information. We consider our
It is important to notice that these patients’ pain points are proposed method as narrative inquiry that uses field texts
highly subjective. Even if two patients have exactly the same such as tweets, conversations based on replies, retweets and
problem, the underlying causes of that problem could differ physician generated hashtags, and conversational communi-
greatly from one patient to another. ties as the units of analysis to research and understand the
way patients create meaning in their lives. Making sense of
a life involves exploring the complex interactions between
the caregiver and the conversational communities of patients
and mediated by the particular topic, hashtag or features in
patient lives over time. Thus meaning of a conversation is
continually constructed and reconstructed. This construction
and reconstruction occurs within, and is made visible through
patient stories as expressed by their tweets. Stories reflect
experience and can be crawled through the narratives. The
conversations crawler is the first method in our proposed IDL
set of methods (see Figure 4). The crawler starts by reading
the tweets stream relevant to the physician geolocation or the
hashtag or hashtags of this physician and convert the word
sequence into a term matrix in which the rows are tweets and
the columns are words. The values matching a document with
FIGURE 3. Popular patient pain points. a word in the matrix is counted using the tf-idf algorithm.
This matrix is then used by a community graph construction
Too often to ease the impact of these pain points, it falls algorithm like the TextRank [19] to produced highly focused
to physicians to provide solutions for their patients. While graph of tweets that are related to certain search topic and
challenging, balancing evidence with patient perspectives context through adjusting the weights and context of the
is essential to protecting patient engagement and promot- conversation lenses. Notable algorithms will be considered
ing more patient-appropriate medical decisions [15]. It will for the purpose of producing the primitive conversation graph
require physicians to improve their skills in compassion and including Edge Betweenness, Walktrap, Label Propagations,
expand their view of treatment beyond literature and guide- SpinGlass and Louvain [20]. To measure the robustness of
lines to understand patient-specific, cultural, and societal the community detection algorithms to give insight in the
contexts as well as to involve new skills to mine for further strength of the divisions of a network into communities,
knowledge to provide prognostic views to their patients. we proposing to use the LFR measures and benchmarks [21]
as well as a large dataset of health related tweets from reliable
III. DEVELOPING IDL METHODS FOR RESPONDING TO sources like the one in [22]. Based on the LFR measures
PATIENT PAIN POINTS OVER THE INTERNET on can select the most effective algorithm for detecting con-
Advancements in technology including mobile and web versation communities. The conversation crawler produces
applications have always had major impacts in medicine summaries for each of the community conversations as well
with tools and services. The smartphone and handheld as visualizing the overall community conversation graph. The
devices are one of the most ubiquitous and dynamic trends summarization algorithm that can be used by the crawler is a
in communication, in which one’s mobile phone can also simple extraction-based summarization [23].

115000 VOLUME 8, 2020

J. Fiaidhi: Envisioning IDL Based on Thick Data Analytics With Focus on Healthcare

anomalies can be expressed using clear Cypher3 queries. For

example, we know that fake accounts botnets over Twitter
tend to share content in short bursts, using recently regis-
tered accounts with few connections [25]. To discover such
anomaly we can run a simple Cypher query like this: that:
MATCH (account:Account)–(ip:IP)–(post:Post)
–(article:Tweet)
WHERE account.friends < 10 AND article.url =
‘wwww.GoDaddy.com’
AND post.timestamp > 1588291200
AND post.timestamp < 1590969599
RETURN account
Which will return all Twitter accounts that have
FIGURE 4. The twitter conversation crawler. less than 10 follower’s connections and shared a link
to wwww.GoDaddy.com between Friday, May 1, 2020
12:00:00 AM and Sunday, May 31, 2020 11:59:59 PM?
The first addition to the IDL methods is subdividing the However, results of the Cypher queries can be further visu-
conversational community networks based on the linkage alized as a graph in Neo4j Keylines4 API. Moreover, it quite
style (e.g. retweet, mention and hashtag). Each style tweets common to use Cypher to query the conversation networks
conversation may be captured as a visualized story using the about routine thick data questions that have a good impact on
Neo4j.1 The timeline of these conversations about a pain the caregiver practice like:
point (e.g. like using #tag) can be displayed with Neo4j Common Query 1: Graph some of the caregiver men-
TimeTree API.2 Figure 5 illustrates the way such timeline tions?
visualization will be. MATCH
(u:Me:User)-[p:POSTS]->(t:Tweet)-[:MENTIONS]
->(m:User)
WITH
u,p,t,m, COUNT(m.screen_name) AS count
ORDER BY
count DESC
RETURN
u,p,t,m
LIMIT 10
Common Query 2: List the most influential followers?
MATCH
(follower:User)-[:FOLLOWS]->(u:User:Me)
RETURN
follower.screen_name AS user, follower.followers
AS followers
FIGURE 5. Conversation community timeline.
ORDER BY
followers DESC
The following method will provide explorations for
Common Query 3: The hashtags caregiver have used most
anomalies that can be detected from conversational commu-
often?
nities. To tackle anomaly or abnormality detection problem,
MATCH
we are proposing to use many techniques have been devel-
(h:Hashtag)<-[:TAGS]-(t:Tweet)
oped in the past decades, especially for spotting outliers and
<-[:POSTS]-(u:User:Me)
anomalies in unstructured collections of multi-dimensional
WITH
data points like those provided by the conversational commu-
h, COUNT(h) AS Hashtags
nities. Anomaly detection in conversation networks refers to
ORDER BY
detecting users’ abnormal opinions and sentiment patterns as
Hashtags DESC
well as special temporal aspects of such patterns [24]. Since
LIMIT 10
we have stored our conversational community’s networks
RETURN
in Neo4j graph database which offer huge advantages over
h.name, Hashtags
traditional SQL or relational databases. Queries to detect
1 https://neo4j.com/ 3 https://neo4j.com/developer/cypher-query-language/
2 https://github.com/graphaware/neo4j-timetree 4 https://cambridge-intelligence.com/keylines/

VOLUME 8, 2020 115001

 J. Fiaidhi: Envisioning IDL Based on Thick Data Analytics With Focus on Healthcare

Common Query 4: What is my Follow back rate? editor will convert it into Cypher code. The next IDL method
MATCH in our thick data analytics focuses on providing more context
(me:User:Me)-[:FOLLOWS]->(f) oriented graphs from the conversational communities like
WITH adherence to medical instructions graphs and compliance of
me, f, size((f)-[:FOLLOWS]->(me)) as providing a medical service. For example by in response to
doesFollowBack Common Query 6 we may end with a conversation network
RETURN (Figure 6) that reveals medication adherence problem.
SUM(doesFollowBack) / toFloat(COUNT(f)) AS
followBackRate
Common Query 5: What are my Follower Recommenda-
tions - tweeting about me, but I don’t follow?
MATCH
(ou:User)-[:POSTS]->(t:Tweet)-[mt:MENTIONS]
->(me:User:Me)
WITH
DISTINCT ou, me WHERE
(ou)-[:FOLLOWS]->(me)
AND NOT
(me)-[:FOLLOWS]->(ou)
FIGURE 6. Conversation network revealing medication adherence
RETURN problem.
ou.screen_name
Common Query 6: Provide links from interesting The last IDL method of this thick data framework is to use
retweets? machine learning to infer the prognosis status of the patient
MATCH cases from the conversation networks. This is a challenging
(:User:Me)-[:POSTS]-> area of research as machine learning is rarely been used
(t:Tweet)-[:RETWEETS]->(rt)-[:CONTAINS] on graph-based data like our conversation networks [26].
->(link:Link) Machine learning on conversation networks can help us to
RETURN do something like the followings:
t.id_str AS tweet, link.url • Predicting if a relationship exists between two nodes
AS url, rt.favorites AS favorites • Scoring and classification of nodes, edges and whole
ORDER BY graphs
favorites DESC Our machine learning IDL method will use node embed-
LIMIT 10 ding (implemented using random walks) for link prediction.
Common Query 7: Who is tweeting with my top hash- Embedding’s are often generated such that nearby nodes in
tags? the graph have similar embedding tensors [27]. Therefore
MATCH a comparison between (e.g. Euclidean distances) provides
(me:User:Me)-[:POSTS]->(tweet:Tweet)-[:TAGS] a likelihood of linkage. Some methods like Node2Vec [28]
->(ht) actually directly train the embedding models on the pres-
MATCH ence/absence of links.
(ht)<-[:TAGS]-(tweet2:Tweet)<-[:POSTS] Our final step is to fine tune the conversation modeling pro-
-(sugg:User) cess captured in the machine learning IDL method when new
WHERE wave of tweets arrives with using transfer learning. Transfer
sugg <> me learning has gained increasing attention due to the inferior
AND NOT performance of machine learning algorithms with insufficient
(tweet2)-[:RETWEETS]->(tweet) training data. Most of the previous transfer learning methods
WITH and APIs (e.g. ULMFiT, OpenAI GPT, ELMo, Stanford GloVe
sugg, collect(distinct(ht)) as tags and Google AI’s BERT), aim to learn a mapping function
RETURN between feature spaces based on the inherent correspondence
sugg.screen_name as friend, size(tags) as common across the source and target domains or labeled instances.
ORDER BY However, in many real world applications like our conversa-
common DESC tional networks, existing methods may not be robust when the
LIMIT 20 correspondence across domains is noisy or labeled instances
If a physician wanted to create other type of queries then are not representative. We are experimenting recently with
s/he can use the Arrows5 graphics editor to draw it and the adopting a recent transfer learning via ‘‘Feature Isomorphism
Discovery (TLFid)’’ [29] to discover common substructures
5 http://www.apcjones.com/arrows/# across feature spaces and learning a feature mapping function

115002 VOLUME 8, 2020

J. Fiaidhi: Envisioning IDL Based on Thick Data Analytics With Focus on Healthcare

from the target domain to the source domain and form and their conversations. These methods start with identifying the
restructure the conversational networks on what currently conversation networks based on popular community detec-
called community embedding transfer learning [30]. To learn tion algorithms and uses the graph algorithms provided by
such embedding, our insight depends on the community the Neo4j API to infer and visualize the relevant and focused
detection algorithms and the node embedding techniques on conversations providing thick data observations to the physi-
the new arriving conversations. Node embedding will be used cians/healthcare givers who presented their questions on
to identify new conversation communities to be used for the values of their care as seen by the Twitter community
community embedding [31]. Figure 7 illustrates our overall around them. Moreover, this paper envisioned other two IDL
vision framework for learning patient insights from health- methods to build machine learning models (IDL 6) for the
care Twitter conversational networks. inferred thick data observations based on node2vector tech-
niques as well as to update these machine learning mod-
els whenever new conversation was detected using graph
based transfer learning (IDL 7) using community embedding
technique. The author is currently experimenting with these
IDL methods using geo-located Twitter data relevant to the
North Western Ontario region about patient tweets related to
COVID-19 care. The source of the COVID-19 Twitter Data
is from IEEE recent repository.6 More on these results are in
the publication pipeline that is coming soon.
ACKNOWLEDGMENT
The author would like to express his appreciation to Dr. Sabah
Mohammed of Lakehead University for his first reading edits
on this article.
REFERENCES
[1] T. Sakaeda, A. Tamon, K. Kadoyama, and Y. Okuno, ‘‘Data mining of the
FIGURE 7. Overall thick data insight driven learning methods. public version of the FDA adverse event reporting system,’’ Int. J. Med.
Sci., vol. 10, no. 7, p. 796, 2013.
[2] B. Ristevski and M. Chen, ‘‘Big data analytics in medicine and healthcare,’’
J. Integrative Bioinf., vol. 15, no. 3, Sep. 2018.
IV. CONCLUSION [3] Y. Wang, L. Kung, and T. A. Byrd, ‘‘Big data analytics: Understanding its
Many healthcare organizations today are changing their sys- capabilities and potential benefits for healthcare organizations,’’ Technol.
Forecasting Social Change, vol. 126, pp. 3–13, Jan. 2018.
tems to the new Value-Based Healthcare (VBHC) which is [4] S. Pink, H. Horst, J. Postill, L. Hjorth, T. Lewis, and J. Tacchi, Digital
becoming a leading approach to improving patient and health Ethnography Principles and Practice. Newbury Park, CA, USA: Sage,
2019.
system outcomes around the world [32]. It has been consid- [5] H. Bouwman, C. Carlsson, J. Carlsson, S. Nikou, A. Sell, and P. Walden,
ered as the only way of organizing healthcare to transform ‘‘How Nokia failed to nail the Smartphone market,’’ in Proc. 25th Eur.
health outcomes. However, the current healthcare systems Regional Conf. Int. Telecommun. Soc. (ITS), Brussels, Belgium, Jun. 2014,
pp. 1–18.
does not allow to see the full interconnected parts contribut- [6] G. Latzko-Toth, C. Bonneau, and M. Millette, ‘‘Small data, thick data:
ing to the health outcomes. Few are taking a long-term Thickening strategies for trace-based social media research,’’ in The
holistic approach (e.g. focus groups, surveys, and scrap- SAGE Handbook of Social Media Research Methods, A. Quan-Haase and
L. Sloan, Eds. London, U.K.: SAGE, Jan. 2017, ch. 13.
ping data from healthcare websites) to gain insights and [7] L. Gelinas, J. S. Weissman, H. F. Lynch, A. Gupta, R. Rozenblum,
enhance outcomes. Much of these approached have high costs E. A. Largent, and I. G. Cohen, ‘‘Oversight of patient-centered outcomes
are driven by the lack of connected information and poor research: Recommendations from a Delphi panel,’’ Ann. Internal Med.,
vol. 169, no. 8, pp. 559–563, 2018.
patient’s engagement. This article focuses on social media, [8] D. Fink-Hafner, T. Dagen, M. Doušak, M. Novak, and M. Hafner-Fink,
in particular Twitter, as a source for patient insights due ‘‘Delphi method: Strengths and weaknesses,’’ Adv. Methodol. Statis-
to its omnipresence, dynamism, engaging and geolocation tics/Metodoloski Zvezki, vol. 16, no. 2, pp. 1–19, 2019.
[9] J. Brett, S. Staniszewska, I. Simera, K. Seers, C. Mockford, S. Goodlad,
focus. Twitter has enabled patient’s discussions to spread D. Altman, D. Moher, R. Barber, S. Denegri, A. R. Entwistle, P. Littlejohns,
and evolve, adding a new channel of valuable insights for C. Morris, R. Suleman, V. Thomas, and C. Tysall, ‘‘Reaching consensus on
healthcare. With millions of discussions happening across reporting patient and public involvement (PPI) in research: Methods and
lessons learned from the development of reporting guidelines,’’ Brit. Med.
Twitter networks in real time, the potential of listening to J. Open, vol. 7, no. 10, Oct. 2017, Art. no. e016948.
your patients on Twitter cannot be understated. This paper [10] S. Staniszewska, A. Adebajo, R. Barber, P. Beresford, L.-M. Brady, J. Brett,
described a holistic vision to identify useful patient insights J. Elliott, D. Evans, K. L. Haywood, D. Jones, C. Mockford, M. Nettle,
D. Rose, and T. Williamson, ‘‘Developing the evidence base of patient
over Twitter relevant to the patient pain points using qualita- and public involvement in health and social care research: The case for
tive insights driven methods (IDLs). This visionary approach measuring impact,’’ Int. J. Consum. Stud., vol. 35, no. 6, pp. 628–632,
described five notable IDLs methods (IDL 1 to IDL 5) that Nov. 2011.
can provide physicians and healthcare providers with mean- 6 https://ieee-dataport.org/open-access/corona-virus-covid-19-
ingful insights about their local patient’s communities and geolocation-based-sentiment-data

VOLUME 8, 2020 115003

 J. Fiaidhi: Envisioning IDL Based on Thick Data Analytics With Focus on Healthcare

[11] L. J. Brighton, S. Pask, H. Benalia, S. Bailey, M. Sumerfield, J. Witt, [24] X. Sun, C. Zhang, and L. Li, ‘‘Dynamic emotion modelling and anomaly
S. de Wolf-Linder, S. N. Etkind, F. E. M. Murtagh, J. Koffman, and detection in conversation based on emotional transition tensor,’’ Inf.
C. J. Evans, ‘‘Taking patient and public involvement online: Qualita- Fusion, vol. 46, pp. 11–22, Mar. 2019.
tive evaluation of an online forum for palliative care and rehabilitation [25] M. Kelsey. (Dec. 4, 2013). Officials Worry Fake Health Exchange
research,’’ Res. Involvement Engagement, vol. 4, no. 1, p. 14, Dec. 2018. Websites Could Defraud, Confuse Consumers. Huffington Post.
[12] J. Fiaidhi, S. Mohammed, and J. Fiaidhi, ‘‘Thick data: A new qualita- Sep. 13, 2013. [Online]. Available: https://www.huffpost.com/entry/fake-
tive analytics for identifying customer insights,’’ IT Prof., vol. 21, no. 3, health-exchange-websites_n_3916399
pp. 4–13, May 2019. [26] J. Zhou, G. Cui, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, and M. Sun,
[13] N. Survey. New Survey Report Reveals Top Pain Points for Healthcare ‘‘Graph neural networks: A review of methods and applications,’’ 2018,
Consumers. Medici Blog, Nov. 12, 2019. [Online]. Available: arXiv:1812.08434. [Online]. Available: http://arxiv.org/abs/1812.08434
https://blog.medici.md/new-survey-report-reveals-top-pain-points- [27] M. Heimann and D. Koutra, ‘‘On generalizing neural node embedding
for-healthcare-consumers methods to multi-network problems,’’ in Proc. KDD MLG Workshop, 2017,
[14] J. F. Baumhauer and K. J. Bozic, ‘‘Value-based healthcare: Patient-reported pp. 123–127.
outcomes in clinical decision making,’’ Clin. Orthopaedics Rel. Res., [28] A. Grover and J. Leskovec, ‘‘node2vec: Scalable feature learning for
vol. 474, no. 6, pp. 1375–1378, Jun. 2016. networks,’’ in Proc. 22nd ACM SIGKDD Int. Conf. Knowl. Discovery Data
[15] C. Tan and L. Loh. Bridging the Information Gap Between Physician and Mining, Aug. 2016, pp. 855–864.
Patient Perspectives on Health. CMAG Blog, Apr. 21, 2017. [Online]. [29] S. Di, J. Peng, Y. Shen, and L. Chen, ‘‘Transfer learning via feature
Available: https://cmajblogs.com/bridging-patient-physician-gap/ isomorphism discovery,’’ in Proc. 24th ACM SIGKDD Int. Conf. Knowl.
[16] E. Ozdalga, A. Ozdalga, and N. Ahuja, ‘‘The smartphone in medicine: Discovery Data Mining, Jul. 2018, pp. 1301–1309.
A review of current and potential use among physicians and students,’’ [30] V. W. Zheng, S. Cavallari, H. Cai, K. C.-C. Chang, and E. Cambria, ‘‘From
J. Med. Internet Res., vol. 14, no. 5, p. e128, Sep. 2012. node embedding to community embedding,’’ 2016, arXiv:1610.09950.
[17] D. Dunlop, ‘‘Integrating Twitter into your physician communication [Online]. Available: http://arxiv.org/abs/1610.09950
strategy,’’ Manage. Healthcare, vol. 2, no. 3, pp. 245–253, [31] S. Cavallari, V. W. Zheng, H. Cai, K. C.-C. Chang, and E. Cambria, ‘‘Learn-
2018. ing community embedding with community detection and node embed-
[18] J. M. Alpert and F. E. Womble, ‘‘Just what the doctor tweeted: Physicians’ ding on graphs,’’ in Proc. ACM Conf. Inf. Knowl. Manage., Nov. 2017,
challenges and rewards of using Twitter,’’ Health Commun., vol. 31, no. 7, pp. 377–386.
pp. 824–832, Jul. 2016. [32] C. Colldén and A. Hellström, ‘‘Value-based healthcare translated: A com-
[19] Z. Wang, Y. Feng, and F. Li, ‘‘The improvements of text rank for domain- plementary view of implementation,’’ BMC Health Services Res., vol. 18,
specific key phrase extraction,’’ Int. J. Simul. Syst., Sci. Technol., vol. 17, no. 1, p. 681, Dec. 2018.
no. 20, pp. 1–11, 2016.
[20] P. K. Gopalan and D. M. Blei, ‘‘Efficient discovery of overlapping commu-
nities in massive networks,’’ Proc. Nat. Acad. Sci. USA, vol. 110, no. 36, JINAN FIAIDHI (Senior Member, IEEE) is cur-
pp. 14534–14539, Sep. 2013. rently a Full Professor of computer science and the
[21] A. Lancichinetti, S. Fortunato, and F. Radicchi, ‘‘Benchmark graphs Graduate Coordinator of the Biotechnology Ph.D.
for testing community detection algorithms,’’ Phys. Rev. E, Stat. Phys. Program with Lakehead University. She is also an
Plasmas Fluids Relat. Interdiscip. Top., vol. 78, no. 4, Oct. 2008,
Adjunct Research Professor with the University of
Art. no. 046110.
[22] D. Darmon, E. Omodei, and J. Garland, ‘‘Followers are not enough: A Western Ontario. She is also the Chair of Big Data
multifaceted approach to community detection in online social networks,’’ for eHealth with the IEEE ComSoc. She is also
PLoS ONE, vol. 10, no. 8, Aug. 2015, Art. no. e0134860. the Editor-in-Chief of the International Journal of
[23] N. Chatterjee and S. Mohan, ‘‘Extraction-based single-document summa- Extreme Automation and Connectivity in Health-
rization using random indexing,’’ in Proc. 19th IEEE Int. Conf. Tools Artif. care (IGI Global).
Intell. (ICTAI), Oct. 2007, pp. 448–455.

115004 VOLUME 8, 2020