Journal of Management Information Systems

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/mmis20

Unveiling the Hidden Truth of Drug Addiction: A

Social Media Approach Using Similarity Network-
Based Deep Learning

Jiaheng Xie, Zhu Zhang, Xiao Liu & Daniel Zeng

To cite this article: Jiaheng Xie, Zhu Zhang, Xiao Liu & Daniel Zeng (2021) Unveiling
the Hidden Truth of Drug Addiction: A Social Media Approach Using Similarity Network-
Based Deep Learning, Journal of Management Information Systems, 38:1, 166-195, DOI:
10.1080/07421222.2021.1870388

To link to this article: https://doi.org/10.1080/07421222.2021.1870388

View supplementary material Published online: 02 Apr 2021.

Submit your article to this journal Article views: 1102

View related articles View Crossmark data

Citing articles: 3 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=mmis20
JOURNAL OF MANAGEMENT INFORMATION SYSTEMS
2021, VOL. 38, NO. 1, 166–195
https://doi.org/10.1080/07421222.2021.1870388

Unveiling the Hidden Truth of Drug Addiction: A Social Media

Approach Using Similarity Network-Based Deep Learning
Jiaheng Xiea, Zhu Zhangb,d, Xiao Liue, and Daniel Zengb,c,d
a
Department of Accounting and MIS, Lerner College of Business & Economics, University of Delaware, Newark,
DE, USA; bState Key Laboratory of Management and Control for Complex Systems, Institute of Automation,
Chinese Academy of Sciences; cUniversity of Chinese Academy of Sciences; dShenzhen Artificial Intelligence and
Data Science Research Institute (Longhua); eArizona State University, Tempe, AZ, USA

ABSTRACT KEYWORDS
Opioid use disorder (OUD) is an epidemic that costs the U.S. healthcare Computational design
systems $504 billion annually and poses grave mortality risks. Existing science; deep learning; social
studies investigated OUD treatment barriers via surveys as a means to media analytics; health IT;
HealthTech; opioid
mitigate this opioid crisis. However, the response rate of these surveys
addiction; addiction
is low due to social stigma around opioids. We explore user-generated treatment
content in social media as a new data source to study OUD. We design
a novel IT system, SImilarity Network-based DEep Learning (SINDEL),
to discover OUD treatment barriers from patient narratives and
address the challenge of morphs. SINDEL significantly outperforms
state-of-the-art NLP models, reaching an F1 score of 76.79 percent.
Thirteen types of treatment barriers were identified and verified by
domain experts. This work contributes to information systems with
a novel deep-learning-based approach for text analytics and general­
ized design principles for social media analytics methods. We also
unveil the hurdles patients endure during the opioid epidemic.

Introduction
Social media have been fundamentally changing the approach businesses communicate, colla­
borate, consume, and create [4]. According to a McKinsey Quarterly report, 50 percent of the
more than 1,700 surveyed businesses use social networking, 41 percent use blogs, 25 percent use
wikis, and 23 percent use microblogs [11]. These social media platforms, coupled with analytics
methods, offer powerful solutions to support business-related functions [58, 61], such as
inferring consumer preferences, peer-to-peer and targeted marketing, and stock prediction,
among others [7, 59]. In the medical domain, social media analytics is also a forerunner of
reorganizing medical knowledge creation and medical work in the digital age, and a complement
to established health research methods. Kallinikos and Tempini [29] analyzed data from a web-
based, medical research network with patient self-reports, and assessed the value of social media
for extracting medical knowledge and understanding patient experience. Witnessing the great
potential of social media in health analytics research, we aim to explore an online social network
with patient self-reports to uncover patient experience to treat opioid use disorder (OUD).

CONTACT Jiaheng Xie [email protected]

Alfred Lerner College of Business & Economics 303 Alfred Lerner Hall Newark, DE 19716, USA.
The two first authors have contributes equally.
Supplemental data for this article can be accessed on the publisher’s website
© 2021 Taylor & Francis Group, LLC
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 167

Opioids are a group of pain-relievers that interact with opioid receptors in cells [33]. In
2016, 11.8 million Americans misused prescription opioids or used illicit opioids [60].
Among them, 2.1 million suffered from opioid addiction. The cost of OUD exceeds
$504 billion in 2015, or 2.8 percent of the GDP [15]. These costs are reflected in healthcare,
lost productivity, treatment for substance use disorder, and the criminal justice system. To
address the nation’s opioid epidemic, the Support for Patients and Communities Act
expands access to treatments of OUD. Patients who remain on treatments for a longer
time tend to have better outcomes. The risk of relapse greatly escalates if patients stop these
treatments. Nevertheless, retaining patients in treatment remains a significant challenge.
Only 17.5 percent of patients with OUD continue to receive treatments [43]. Such a low
adoption rate is due in part to significant access barriers to treatments, such as the fear of job
loss, limited insurance coverage, social stigma, and treatment programs located in limited
areas with long waitlists. Understanding these barriers is the premise to reduce overdose
mortality, mitigate the transmission of infectious diseases, and lower healthcare costs.
Surveys are extensively used in prior work to understand the barriers to OUD treatments
[21, 41]. These studies are challenged by the narrow patient population, as individuals with
OUD are difficult to reach if they are not actively under treatment. User-generated content
in social media can bridge this gap. Studies show that 59 percent of adults participate in
health social media platforms, which offer a vast amount of patient narratives about their
medication-taking behavior [45]. Social media has also been used as an information
technology (IT) application to provide innovative insights on various urgent health issues
[12, 19]. Due to the anonymous nature of online forums, patients are willing to share their
experience of taking prescription and illicit opioids and elaborate on their real decision-
making on OUD treatments (e.g., Figure 1). This large-scale patient-reported information
creates an unprecedented potential to study OUD treatment barriers from the patient
standpoint and facilitate analyses on heterogeneous patients timely [14]. Nevertheless,
social media analytics approach is still lacking in OUD research.
Significant challenges still linger to understand patient perspectives from drug forums
despite their enormous potential. First, OUD treatment barriers are contingent on patients’
life events over time. Effective surveillance for OUD treatments necessitates fine-grained
automated models that could identify barriers timely. Second, patients use vocabulary
different from that of health professionals. Only 43 percent of symptom terms in social
media are presented as exact or synonymous with terms in the professional language [55].
Such a language gap stems from the morphs (fake alternative names) that patients use to
describe drugs and treatments in order to avoid censorship and surveillance, entertain
readers, or use personal writing styles. For instance, morphs of Oxycodone include Oxy, O.
C., Oxycet, Oxycontin, and more. Morphs of OUD treatments, such as meth, MMT, and
chocolate chip cookies, are extensively used in patient discussions. The literal meanings of

Figure 1. Examples of patient narratives about opioid use disorder (OUD) treatment.
168 XIE ET AL.

the morphs are distant from their contextual meanings. This gap between the literal and
contextual meanings of morphs poses a significant challenge to understand patient discus­
sions. Fortunately, the patient-generated morphs, like other community-specific languages,
can be learned through representation learning and knowledge construction.
Motivated by the critical need for fine-grained and automated techniques to understand
OUD treatment barriers in drug forums, we propose a novel computational method:
SImilarity Network-based DEep Learning (SINDEL). We define the barriers to OUD
treatment as patient self-described factors leading to the denial of OUD treatments.
SINDEL extends the state-of-the-art text mining models with a similarity network-based
component and multi-view deep learning architecture. The similarity network-based com­
ponent bridges the literal and contextual semantics of morphs. The multi-view deep
learning architecture enhances the learning performance on sparse OUD-related narratives
through a recurrent and parallel hierarchical structure. SINDEL collects timely social media
data about opioid use and identifies existing and emerging treatment barriers that could
assist proactive interventions.
Our study makes the following contributions to information systems (IS) literature, data
analytics methodology, and healthcare practice. First, we develop the SINDEL method to extract
OUD treatment barriers from drug forums. The proposed method significantly outperforms the
baseline models. The performance enhancement is attributed to SINDEL’s capability of inter­
preting the semantic meaning of morphs accurately. This finding contributes to the literature of
text analytics and linguistics that incorporating a similarity network in language models could
disentangle the complex semantics in specialized knowledge forums. SINDEL can also be
generalized to extract information from many other text genres containing specialized morphs,
such as hacker forums, health social media, and product reviews.
Second, our study falls under the realm of computational design science research that
aims to design analytics solutions to problems with societal impact [48]. We provide an
automated framework to understand patient decision-making. In line with the design
science guidelines in Hevner et al. [24], this framework serves as the IT artifact. To select
the best artifact design, we conducted rigorous experiments to test the hyperparameters in
the input and hidden layers and the parameters in the similarity network, and to benchmark
with other plausible systems, among others. The proposed framework contributes to the IS
knowledge base, as the design principles (model selection and hyperparameter tuning) and
the new method offer generalized guidelines for design science research in social media
analytics.
Third, our findings involve the heterogeneity among patients and open a window to
understanding illicit drug users’ behavior. The drug forum exhibits rich knowledge and
patient experience generated by patients. Our findings can complement the behavioral
science research on OUD with comprehensive patient experience data. We discover 13
types of OUD treatment barriers, including lack of medical literacy, social stigma, concerns
about job loss, withdrawal reactions and side effects, and more. Many of these treatment
barriers have not been noted by prior survey studies, such as side effects of treatment,
concerns about treatment addiction, poor patient-physician relationships, and depressed
mental status. We provide valuable implications for medical professionals and policymakers
to understand individual opioid-taking behavior and the real treatment barriers faced by
patients. Tailored intervention measures can be taken accordingly to prevent medical and
financial ramifications, improve OUD management, and reverse the opioid crisis.
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 169

Literature Review
Our study is related to four research areas. First, we survey the literature on OUD treat­
ments for the current understanding of OUD. Second, we review online community and
design science to explore IT systems to assist opioid research. Third, we summarize opinion
mining in social media to bring in design principles. Fourth, we review morphology studies
to propose a method that can address the technical challenges to identify OUD treatment
barriers in drug forums.

Barriers to OUD Treatments

In the past decade, opioid use disorder (OUD) has drawn attention from clinical practi­
tioners, policymakers, and researchers, who attempt to understand and remove the treat­
ment barriers to OUD. Table 1 shows the recent studies on OUD treatment barriers.
Surveys are commonly used to investigate the barriers to OUD treatment, with subject
sample sizes ranging from 20 to over 4,000 patients [41]. The barriers to OUD treatment
identified in these studies can be summarized as three categories: 1) System-related: the
factors related to healthcare systems and regulations, such as government and insurance
policies and funding barriers [32]; 2) Provider-related: the factors related to health provi­
ders, such as lack of DEA waiver, lack of institutional support, lack of resources, and
geographic constraints [53]; 3) Patient-related: patient-specific factors, such as the fear of
pain and lack of information on treatments [21].
Understanding the barriers to OUD treatment forms the basis for developing effective
interventions. Current studies investigated the barriers to OUD treatment via surveys,

Table 1. Recent Studies on opioid use disorder (OUD) Treatment Barriers.

Author Year Method Participants OUD Treatment Barriers Intervention
Livingston et al. [37] 2018 Interview 20 Access to methadone Enhance the uptake and availability of
expertise, evidence-based treatment
professional support
Hassamal et al. [21] 2017 Review 5 Financial, Engage high-risk patients in substance
administrative, policy, abuse treatment, decrease
lack of information recidivism, close monitoring and
supervision
McKenna [41] 2017 Survey 4,100 Financial National implementation of the
Affordable Care Act (ACA)
Stumbo et al. [57] 2017 Interview 283 Fear of uncontrolled Use separate addiction outpatient
pain, stigma of therapy groups
addiction
Andrilla et al. [6] 2017 Survey 1,124 Lack of physician’s Drug Encourage physicians to add or
Enforcement maintain this service, including
Administration (DEA) targeting former prescribers
waiver
Sharma et al. [53] 2017 Review NA Financial, regulatory, Expand Medicaid, cover all medications
geographic, for OUD, discourage managed care
attitudinal, logistic organizations to erect further
administrative barriers
Bojko et al. [10] 2016 Interview 86 Unclear treatment goal, Quality improvement, patient
provider and societal education and medical professional
negative attitude, development
legislation
Hutchinson et al. [26] 2014 Interview 92 Physician’s lack of Increase the number of physicians who
institutional support offer buprenorphine, technical
assistance training
170 XIE ET AL.

which offer only cursory descriptions of some well-known barriers, lacking depth and
comprehensiveness. These surveys only capture a snapshot of barriers. In reality, many
patient-level barriers are complicated by patient characteristics and policy changes, causing
them to vary over time. Hence, surveys are unlikely to offer a comprehensive view of
treatment barriers. Moreover, survey studies are provider-centric, lacking understanding
from the patients’ perspectives. This dominant provider perspective poses a significant
challenge to translating these treatment barriers into real-world settings as providers have
minimal control over patient decision-making on OUD treatments. Furthermore, patients
are reluctant to disclose their issues with OUD treatments, especially illicit drug users. Some
patients intentionally avoid being connected to OUD treatments due to the social stigma
around opioids.

Online Social Community and Design Science

Innovative approaches are needed to understand OUD treatment barriers, patient beha­
viors, and potential measures that can deliver care to patients in need. Health big data from
social media make innovative projects possible and open opportunities for investigations
that can yield insights into and understanding of issues such as patient decision-making,
human motivation, and social phenomena [8, 14]. Because of the anonymous nature of
social media, many patients, including illicit drug users, actively share their drug-taking
experience with their peers.
Since the early 2000s, many social media applications have efficiently gathered a large
volume of self-reported experience from a diverse patient population [14, 34]. This patient
self-reported experience not only provides dynamic information but also covers an unpre­
cedented scale of the patient population with heterogeneous characteristics. Yet, no social
media analytics approach has been taken in OUD treatment studies.
Computational methods can be integrated into IT systems to yield more insights from the
big health data with patients’ vantage point. The IS literature also stressed the value of health
social media analytics, as online patient communities provide insights into patients’ illnesses
[13, 65]. Since data being generated are increasingly unstructured and coming from networks
of patients, it is essential to design more powerful algorithms and better knowledge repre­
sentation schemes for making sense of this heterogeneous health information [2, 5]. Hevner
et al. [24], Gregor and Hevner [20], and Chen et al. [14] have emphasized the necessity for
such a design and computational category of IS research and outlined a guideline for design
science studies so that cutting-edge algorithms can be leveraged to tackle significant problems
with societal impact. This computational design science paradigm has contributed indispen­
sable design principles and analytical methods to various IS research areas, such as healthcare
IT, cybersecurity, e-commerce, and more [1, 3, 17, 40, 51, 56].

Opinion Mining and Social Media Analytics

Extensive literature in IS utilizes opinion mining techniques to understand the factors of
consumer behaviors, such as purchase decision, technology acceptance, and more [64].
Opinion mining usually includes the following procedures: data collection, opinion identi­
fication, aspect extraction, aspect clustering, and production summary [23]. The barriers to
OUD treatment are aspects related to treatments, patients, and insurance, among others,
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 171

which can be extracted with aspect extraction techniques in opinion mining. Most research
in aspect extraction focuses on social media data. There are four common categories of
techniques in aspect extraction: extraction based on high-frequency noun phrases, extrac­
tion based on exploiting opinion and aspect relations, extraction based on sequence learn­
ing, and extraction based on topic modeling.

Extraction Based on High-Frequency Noun Phrases

When users express their opinions about various aspects of a product, they use similar noun
phrases frequently. These noun phrases and nouns can be determined by syntactic parsing
and POS tagging. Li et al. [35] proposed a method for improving feature extraction
performance in online reviews. Their method, based on frequent nouns and noun phrases,
is consisted of three components: frequency-based mining and pruning, order-based filter­
ing, and similarity-based filtering. Jeyapriya and Selvi [27] extracted nouns and noun
phrases from each review sentence and used a minimum support threshold to find frequent
aspects in review sentences. Extraction based on high-frequency noun phrases is effective
when aspects are expressed in consistent and common terms. When a new aspect or
a variation of existing aspect emerges, this approach is not able to capture it.

Extraction Based on Relations Between Opinion Words and Aspects

The opinion words can be used to describe different aspects. Qiu et al. [46] focused on two
fundamental and important issues, opinion lexicon expansion and target extraction, and
suggested the double propagation approach. They performed extraction based on syntactic
relations that draw links between review words and aspects of products. This approach is
limited because syntactic relations among words in reviews may not be accurately extracted
and consumers do not always follow grammar rules when writing the reviews.

Extraction Based on Topic Modeling

Topic models are unsupervised methods for aspect extraction in which they assume that any
document contains k hidden topics. Some of these topics can be aspects of products.
Statistical methods, such as latent semantic analysis (LSA) and latent Dirichlet allocation
(LDA) with bag-of-words representation of documents, can be used in document-level
aspect extraction. Ma et al. [39] proposed a probabilistic topic model approach based on
LDA in order to search over citizens’ opinions about city issues on online platforms. Luo
et al. [38] devised a Quad-tuple LSA for detection and rating aspects in product reviews.
Sentences in drug forums are very long, because of the noisy nature of social media data and
the misuse of punctuations. If we apply topic modeling at the sentence level, each sentence
contains many topics. The majority of these topics are irrelevant to OUD. Because of the
extensive existence of irrelevant topics, few to none words would be generated by OUD-
related topics. Additionally, as each word has a probability to be generated by any topic,
OUD-related topics may still generate non-treatment barrier words. The unsupervised topic
models cannot pinpoint the words about OUD treatment barriers, since there are no
training labels. For these reasons, topics models are not feasible in this study.

Extraction Based on Sequence Learning

The sequence learning approach is a promising technique for aspect extraction, which are
usually developed for named entity recognition in natural language processing. This
172 XIE ET AL.

approach learns the patterns with labeled reviews and determines whether a word belongs
to an entity of interest, in this case, an aspect expression. Support vector machine (SVM),
conditional random fields (CRF), and hidden Markov model (HMM) are supervised
learning methods that can be used to extract aspects [36]. Aspects were extracted from
a collection of blog posts using machine learning methods, and the results were used as
statistical patterns for aspect extraction. CRF achieves satisfying performance in formal text
genres, but it falls short in processing patient narratives in drug forums. This is because CRF
treats words as discrete atomic symbols, which require accurate input for training and
prediction.
Unlike the data used in existing opinion mining studies, drug forums have unique
features. When describing OUD treatments, patients use many morphs to represent
drugs and treatment options to avoid censorship and surveillance. We refer to morphs as
any derived terms other than the generic medical term. For instance, heroin can be referred
to as H, hero, and China white by different users. None of the aforementioned extraction
approaches are suited for this issue. Traditional text mining methods are not capable of
interpreting the underlying semantics of morphs because they use discrete atomic symbols
to represent the words.

Morphology and Deep Learning

In order to address the challenges in opinion mining in drug forums, we draw insights from
morphology and deep learning studies. Morphology is the study of words, how they are
formed, and their relationship to other words in a language. In drug forums, users invent
new slangs and idiomatic expressions to describe their experience. Related work in mor­
phology aims to address the challenges of understanding slangs, synonyms, and other types
of word variations in unstructured text and hence can help us understand OUD-related
expressions in drug forums. Table 2 shows the recent literature in morphology. Similar to
the drug morphs and patient idiomatic expressions in drug forums, morphology studies
tackle internet slangs, synonyms, and semantically similar terms [62, 66]. The main body of
literature in morphology utilizes distributed representation and deep learning methods,
such as word embedding and recurrent neural network (RNNs), to interpret the semantics
of morphs.
Word embedding is a vector-based representation of words commonly used in deep
learning models for natural language processing. According to the Distributional
Hypothesis, words with similar meanings occur with similar neighbors [50]. Word embed­
ding represents each word in a vector of its surrounding words instead of by the symbolic

Table 2. Recent studies on morphology in online texts.

Author Year Data Method Morphs of Interest
Simpson et al. [54] 2018 Tweets Word embedding Emerging drug terms
Yao et al. [62] 2018 News articles Word embedding Word changes over time
Qu et al. [47] 2017 Wikipedia Word embedding Entity synonyms in Wikipedia
Huang et al. [25] 2017 WSJ articles BLSTM Synonyms in news articles
Sha et al. [52] 2017 Tweets Character-level embedding Internet morphs in Twitter
He et al. [22] 2016 Search query Optimization Attribute synonyms in search queries
Nguyen et al. [42] 2016 Web pages Word embedding Antonym and synonym in text
Zhang et al. [66] 2016 NYT articles Word embedding Semantically similar terms in articles
Notes: WSJ = Wall Street Journal; BLSTM = bidirectional long short-term memory; NYT = New York Times.
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 173

word itself. Word embedding can effectively represent sparse entities, entities with typos,
and entities with variations in social media data for machine learning models.
The RNN is a class of artificial neural network where connections between units form
a loop. This feature creates internal states in the network and enables information to persist
during the learning process. RNN with word embedding input achieves good performance
in many sequence learning tasks, such as parts-of-speech tagging, named entity recognition,
and machine translation [9, 44]. Long Short-Term Memory (LSTM), an improvement to
standard RNN, addresses the long-term dependency issue. LSTM can add or remove
information in each internal state through the internal gates in the LSTM units.
A Bidirectional Long Short-Term Memory (BLSTM) network connects two hidden layers
of opposite directions to the same output, thus making future input data available for the
current state. BLSTM has achieved leading performance for named entity recognition on
noisy user-generated text, because of its ability to consider interconnected information in
a sentence.
The challenge to identify OUD treatment barriers is exacerbated by the morphs in online
OUD discussions. The morphs of drugs and treatment options are widely used in OUD
discussions. The literal meaning of these morphs is distant from their contextual meaning.
Morphs for drugs in the same class are semantically closer than the morphs across drug
classes. Each opioid class has a unique effect, regimen, and instruction, leading to varying
semantic contexts in OUD discussions. For instance, Oxycodone is a semi-synthetic opioid,
the morphs of which include Oxy, O.C., Oxycet, Oxycontin, and more. The morphs of
heroin include H, China white, and more. Oxy and O.C. are not only morphs of opioid
drugs, but they represent the same drug class (Oxycodone) as well. Therefore, Oxy and O.
C. are more closely related than Oxy and China white. The vector representation of words
models the semantic context of words using the neighboring words of the focal word.
Interconnected relationships within the same opioid class are neglected by this method.
A similarity network of words can help address this issue by explicitly examining the
contextual similarity of words. In addition to the semantic relatedness to local neighboring
words, key entities of interest, such as drug names and treatment effects, are related due to
the medical context. Deep learning models for sequence learning are well-suited for
extracting this interdependency across different terms in a sentence. Therefore, we are
motivated to propose a new deep learning architecture that incorporates the vector repre­
sentation of words and semantic similarity in a network to extract treatment barriers from
drug forums.

Research Method
OUD Treatment Barrier Mining Problem Formulation
Let P denote a set of n users P ¼ fp1 ; . . . ; pn g in the drug forum. Let O denote a set of
opioids O ¼ fo1 ; . . . ; om g reported by users in P. A review set vij is a set of all reviews posted
by user pi about opioid oj regarding his or her drug-taking experience in drug forums. vij
contains m sentences sK , where K 2 f1; 2; . . . ; mg. Sentence sk with n words can be
ð1Þ ð2Þ ð nÞ ðt Þ
denoted as ½wk ; wk , . . ., wk �, where wk is the t-th word in sk . If a user pi with opioid
oj decides not to seek treatment, he or she would be likely to describe the barriers to OUD
treatment in the review vij . The barriers to OUD treatment of user pi reporting opioid oj are
174 XIE ET AL.

h i
q ðt Þ
denoted as b1ij ; b2ij ; . . . ; bij (q is the number of barriers). A barrier bkij can be a word wk or
ð1Þ ð2Þ ðmÞ
a phrase [wk ; wk ; . . . ;wk ].
The OUD treatment barrier mining involves two tasks. The first task is a sequence
ðt Þ
learning task, aimed at predicting whether each word wk in sentence sk belongs to the
beginning of a barrier, inside of a barrier, or outside a barrier. The input is a sentence. The
q
output is a set of barrier phrases b1ij ; b2ij ; . . . ; bij . The second task is a clustering task. The
clustering task groups similar barriers together and identifies the general types of barriers.
The input is the barriers identified from the first task. The output is several clusters of
barriers.

Similarity Network-Based Deep Learning Approach

The OUD treatment barrier mining has two objectives: barrier extraction and barrier
clustering, following the established procedure of aspect-based opinion mining (aspect
extraction and aspect clustering) [64]. Barrier extraction identifies patient self-described
q
OUD treatment barriers b1ij ; b2ij ; . . . ; bij for each patient pi taking opioid oj . Since patients use
different expressions to describe the same type of barrier, barrier clustering groups the
identified OUD treatment barriers based on their semantic meaning. The process of the
OUD treatment barrier mining is shown in Figure 2. The novelty of our approach is
highlighted in red. The proposed OUD treatment barrier mining approach receives
a sentence from drug forums as the input. Two parallel representation models represent
the sentence with two vectors. Branch 1 utilizes word embedding to generate semantic
vectors for each word. Branch 2 creates a similarity network of words and generates
a network representation for each word. The two representations are fused to recognize
the OUD treatment barriers in the sentence. A clustering model is utilized to cluster the
extracted barriers into meaningful categories of general barriers. We elaborate on the details
of our deep learning approach in the following subsections.

Similarity Network-Based Representation

The proposed similarity network-based representation contains two parallel representa­
tions. The first representation is a word embedding to capture the semantic meaning of
words, so that morphs can be interpreted as their intended meaning. Given a word, the
objective of word embedding is to learn the vector representation of this word based on its
neighboring words in the same sentence. This vector-based representation could be
obtained by pre-trained embedding from large corpora or locally-trained embedding.
Empirical studies suggest that pre-trained embedding, such as word2vec, yields superior
performance than locally-trained embedding, due to the large training corpora (Google
News). We build our word embedding model leveraging word2vec. To incorporate the
semantic and syntactic characteristics of the OUD context into this representation, we set
the embedding vector to be trainable. The embedding vector will learn to adjust based on
the learning objective of our deep learning model (OUD treatment barrier mining). The
resulting model obtains an array of semantic vectors with 300 dimensions.
The second representation aims to construct a network of words in order to capture the
interconnected relationships. In this network G ¼ ðV; EÞ, each node V is a word, and the
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 175

Figure 2. The opioid use disorder (OUD) treatment barrier mining framework.

edge E is the semantic similarity between words. As such, each word is linked to a set of
words that are closely related. For instance, Oxy is linked to O.C. and Oxycet, because they
are the most similar morphs. Oxy is not linked to China white, because they belong to
different drug classes. This word similarity network addresses the limitation of word
embedding by considering the semantic relationships among entities of interests. Instead
of using the representation of the focal word, we use similar words that are connected to the
focal word as the second representation for the focal word. Figure 3 shows an example.
In the example in Figure 3, Oxycodone (wð1Þ ) is linked to Oxy, O.C., Oxycet, and
Oxycontin, because they belong to the same drug class. We use wð2Þ , wð3Þ , wð4Þ , and wð5Þ
to represent wð1Þ . Likewise, China white, H, and hero are linked to heroin. We use wð7Þ , wð8Þ ,
and wð9Þ to represent wð6Þ . In our corpus, we construct a similarity network for all words and
compute the similarity between each pair of words. We select a set of most similar words for
each word and link them together. The network size (number of similar words) is deter­
mined in the empirical analyses with the highest performance. Word similarity is computed
using the cosine similarity of word embedding as shown in Equation 1.
ðiÞ ð jÞ
simij ¼ kxðxiÞ kkx
x
ð jÞ k (1)

where variables xðiÞ and xð jÞ are the word embedding of word wðiÞ and wð jÞ . Given word w, let
wð1Þ ; wð2Þ ; . . . ; wð10Þ be the top ten words that are the most similar to word w. Let
simð1Þ ; simð2Þ ; . . . ; simð10Þ be the similarity between word w and the other ten words. Let
xð1Þ ; xð2Þ ; . . . ; xð10Þ be the word embedding for wð1Þ ; wð2Þ ; . . . ; wð10Þ . The similarity network
representation of word w is defined in Equation 2.
X
10
xw ¼ simðiÞ xðiÞ : ð2Þ
t¼1
176 XIE ET AL.

Figure 3. An example of a word similarity network.

Multi-view Deep Learning Architecture

To effectively extract the barriers to OUD treatment, we utilize BLSTM architecture with
a multi-view learning framework. Since a sentence follows semantic and syntactic rules,
words in different locations in a sentence may exhibit semantic dependencies regardless of
the word order. This bidirectional structure could capture such dependencies from both
directions. We devise a multi-view BLSTM model that processes the word embedding
representation and the similarity network representation in parallel. The multi-view
BLSTM model contains two branches. Each branch has an independent BLSTM layer that
contains bidirectional LSTM units. Our model is called SImilarity Network-based DEep
Learning (SINDEL).
The LSTM units in branch one take the word embedding as the input, and the LSTM
units in branch two take the similarity network representation as the input. The computa­
tional process for branch two is shown in Equations 3-8. The computational process in the
first branch is the same, except that the input at each time step is word embedding xðtÞ
ðt Þ
instead of similarity network representation xs .

Similarity network-based input gate:

� �
ðt Þ ðt Þ ðt 1Þ
is ¼ σ Wsi xs þ Usi hs bsi ; (3)

Similarity network-based forget gate:

� �
ðt Þ ðt Þ ðt 1Þ
fs ¼ σ Wsf xs þ Usf hs þ bsf ; (4)

Similarity network-based output gate:

� �
ðt Þ ðt Þ ðt 1Þ
os ¼ σ Wso xs þ Uso hs þ bso ; (5)

Similarity network-based cell state:

� �
ðt Þ ðt Þ ðt 1Þ
us ¼ σ Wsu xs þ Usu hs þ bsu ; (6)

Similarity network-based memory cell:

 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 177

ðt Þ ðt Þ ðt Þ ðt Þ ðt 1Þ
cs ¼ is � us þ fs � cs ; (7)

Similarity network-based hidden state:

� �
ðt Þ ðt Þ ðt Þ
hs ¼ os � tanh cs : (8)
ðt Þ ðt 1Þ
where xs is the current input, and hs is the previous hidden state. W, U, and b are
ðt Þ ðt Þ ðt Þ
weight parameters with values between 0 and 1. The gates (is , fs , and os ) and
ðt Þ ðt Þ
memory cell cs take the similarity network representation xs at time step t and
information from previous time steps in the last hidden state hðt 1Þ . Each forward or
backward hidden state has 128 dimensions. We condense useful information from the
300-dimensional xðtÞ to 128 dimensions in the LSTM cell. The learning rate in gradient
descent is 0.1. The dropout rate is 0.2. The above computation (BLSTM) is processed
independently for the word embedding branch and the similarity network branch.
These two branches are then fused together using element-wise sum as a multi-view
learning framework. Figure 4 shows a graphic illustration of the model architecture.
The red part indicates the innovation of this study.
The inputs to SINDEL are sentences from the research corpus. Each word in
a sentence is represented with word embedding and the similarity network repre­
sentation. At each timestep, two BLSTM layers process two representations of
a sentence in parallel and generate two vectors of the same length as the input
representation. These two vectors are merged using element-wise sum to maintain
sequential information. Finally, this merged vector is passed to a Softmax layer to
predict the word type (beginning of a barrier, inside of a barrier, or outside of
a barrier).

Figure 4. Similarity network-based deep learning architecture.

178 XIE ET AL.

OUD Treatment Barrier Clustering

Following the second step of aspect-based opinion mining, aspect clustering, we design
a barrier clustering module. From drug reviews, SINDEL extracts OUD treatment barrier
expressions, many of which may describe the same type of barriers. For instance, “expen­
sive” and “cannot afford the treatment” both refer to the cost of treatments. Grouping
similar aspect expressions, which are domain synonyms, is critical for effective opinion
mining. We utilize a clustering method to identify the general types of OUD treatment
barriers. As word embedding represents the semantic meaning of words in drug forums, we
use word embedding as the features for the clustering model. Given a treatment barrier
containing t words: [wð1Þ ; wð2Þ ; . . . ; wðtÞ ], the word embedding of each word is
[xð1Þ ; xð2Þ ; . . . ; xðtÞ ]. The representation of this treatment barrier is the element-wise average
over each dimension of the word embedding of those words (Equation 9). A salient number
of studies utilized k-means for clustering. We, therefore, use the k-means clustering method
to group OUD treatment barriers.

1X t
xb ¼ xðiÞ : ð9Þ
t i¼1

The methodological novelty of this study is three-folded. First, we design a new similarity
network-based representation that can capture the relationships among morphs. This
representation addresses the gap that existing semantic representations neglect word net­
works or graphs. Second, to consider two independent representations while incorporating
meaningful information from both representations, we devise a multi-view learning frame­
work. Third, we propose a barrier mining approach, including a classification module and
a clustering module. This approach takes an innovative lens at social media data to unveil
hidden patient behavior at a large scale. This automated approach offers an opportunity for
health IT researchers to study patient behavior with scalable analytical methods.

Empirical Analyses
Data Preparation
The research testbed comes from a popular drug discussion platform, Drugs-Forum.com.
This platform enables OUD patients to interact with peers without being judged, because of
the anonymity and the specificity of the platform. Patients on this platform elaborate on
their drug use, addiction, and treatment experience, including illicit drug use, such as
heroin, cocaine, and fentanyl. The research data contains narratives from the OUD patients’
perspective about illicit drugs and creates an unprecedented opportunity to extend the
investigation on opioid addiction to an underrepresented population.
We collected the posts from Drugs-Forum related to drug use1 (i.e., Buprenorphine,
Methadone, and Opiate & Opioid addiction) from the start of Drugs-Forum to
September 1, 2018. The raw dataset encompasses 27,154 posts, involving all prescription
and illicit drugs in the forum, including amphetamine, cocaine, anti-depressants, tranqui­
lizers, sleeping pills, 3,4-Methylenedioxymethamphetamine (MDMA), ethnobotanicals,
gamma hydroxybutyrate (GHB), lysergic acid diethylamide (LSD), psilocybin mushroom,
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 179

buprenorphine, hydrocodone, morphine, oxymorphone, codeine, hydromorphone, opium,

heroin, oxycodone, methadone, and tramadol.
We randomly sampled 3,000 posts by repeatedly drawing a non-duplicate number
between 1 and 27,154 using a random number generator. Four expert annotators reviewed
these sampled posts and annotated the OUD treatment barriers for model training pur­
poses. The inside, out, beginning (IOB) labeling scheme is adapted to assign tags for each
word in a sentence. Each word has a label suggesting if it is inside (I), outside (O), or the
beginning (B) of an expression of OUD treatment barriers. Figure 5 shows an example of
the annotation. In model evaluations, if the predicted word label (I, O, or B) matches the
true label, the prediction of this word is correct.
To test inter-annotator reliability, we leverage Cohen’s Kappa. The Kappa value for the
OUD treatment barrier annotation is 0.92, indicating excellent reliability. A fifth expert
annotator reviewed the disagreements and made the final judgment. We further segmented
the posts into sentences with the sentence boundary detection package from NLTK. A total
of 40,917 sentences were generated. We chose 70 percent of the annotated data as the
training set, 10 percent as the validation set, and the remaining 20 percent as the test set.
The summary statistics of the training and test set are shown in Table 3. The unit of analysis
of the model is at the sentence level. The input to the model is a sentence while the output is
a sequence of predicted word types in the input sentence. There are three different word
types: B, I, and O, suggesting if a word is the beginning of, inside, or outside of a barrier. The
post clustering analyses aggregate all the identified treatment barriers from all the sentences
in a post.

Baseline Models for Extracting OUD Treatment Barriers

Two classes of baseline methods are selected: conventional machine learning methods and
deep learning methods. The conventional machine learning methods include SVM, Logistic
Regression (LR), Naïve Bayes (NB), and CRF. To use SVM, LR, and NB for extracting OUD
treatment barriers, we first use word2vec to represent each word as a feature vector and then
feed the feature vectors into these classifiers respectively. These methods are implemented
with the commonly adopted machine learning library scikit-learn. CRF is the most common
machine learning sequence labeling model. We use CRFSuite to implement CRF. As
SINDEL modifies standard deep learning models, we also benchmark SINDEL with three

Figure 5. Annotation example.

Table 3. Summary statistics of training and test set.

Statistics Training Set Test Set
Number of posts 2,100 900
Number of sentences 27,128 13,789
483,052 232,797
Number of treatment barrier mentions 1,581 678
180 XIE ET AL.

state-of-the-art deep learning models: RNN, LSTM, and BLSTM. These deep learning
models are implemented using Keras. We also benchmark with attention-based SINDEL,
SINDEL with the left branch only, and SINDEL with the right branch only.

Evaluation Metrics
We adopt the common evaluation metrics in text mining—precision, recall, and F1 score—
to assess the performance. Precision assesses how many OUD treatment barriers that the
model retrieves are correct. Recall measures how many OUD treatment barriers in the
research testbed the model can identify. F1 score is the harmonic average of precision and
recall. In our study, we aim to extract as many OUD treatment barriers as possible and
ensure the extracted barriers are accurate at the same time. Extracting more accurate OUD
treatment barriers could operationalize more effective and complete intervention strategies,
thus preventing unnecessary medical consequences. Therefore, F1 score is the most critical
evaluation metric in our study.

Evaluation of Extracting OUD Treatment Barriers

We evaluate our model on the annotated dataset and repeat the training procedure for each
model 20 times and report the average performance in Table 4. SINDEL outperforms the
conventional machine learning methods (SVM, LR, NB, and CRF) by a very large margin in
F1 score and precision. While CRF achieves the highest precision (78.46 percent) among the
baseline methods, SINDEL still outperforms CRF in precision by 6.85 percent. LR has the
highest F1 score (50.01 percent) among the baseline methods. SINDEL manages to outper­
form LR in F1 score by 53.91 percent. NB achieves the highest recall, because it recognizes
most instances as OUD treatment barriers (very low precision), which is not feasible for
practical use. Since F1 score is the most important in this study, SINDEL is the best model
for OUD treatment barrier mining. We conduct t-tests in R to compare the performance of
SINDEL against the conventional machine learning baseline models. The t-test results
indicate that our proposed SINDEL significantly outperforms all the baseline models (p <
0.001).
We also compare SINDEL with state-of-the-art deep learning methods. The results in
Table 5 show that SINDEL outperforms RNN, LSTM, and BLSTM in all three evaluation
metrics. BLSTM achieves the best results among the deep learning baseline methods due to
the advantages of the bidirectional architecture. SINDEL improves upon BLSTM in preci­
sion by 4.15 percent, recall by 11.96 percent, and F1 score by 8.44 percent. The t-tests for
SINDEL against the deep learning methods indicate the performance improvement of our

Table 4. SImilarity Network-based DEep Learning (SINDEL) vs. conventional machine learning methods.
Method Precision (Percent) Recall (Percent) F1 score (Percent)
SVM 58.10*** 40.09*** 47.40***
LR 45.25*** 61.15*** 50.01***
NB 22.54*** 95.50*** 36.47***
CRF 78.46*** 36.59*** 49.90***
SINDEL (Ours) 85.31 70.14 76.97
Notes: SVM = support vector machine; LR = logistic regression; NB = Naïve Bayes; CRF = conditional random fields. *p < 0.05;
**p < 0.01; ***p < 0.001.
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 181

Table 5. SImilarity Network-based DEep Learning (SINDEL) v.s. deep learning models.
Method Precision (Percent) Recall (Percent) F1 score (Percent)
RNN 75.19*** 48.49*** 58.80***
LSTM 71.90*** 54.48*** 61.77***
SINDEL-Left (BLSTM) 81.91*** 62.65*** 70.98***
SINDEL-Right 80.76*** 69.08 74.43*
SINDEL-ATT 80.02*** 65.75*** 72.18***
SINDEL (Ours) 85.31 70.14 76.97
Notes: RNN = recurrent neural network; LSTM = long short-term memory; BLSTM = bidirectional long short-term memory;
SINDEL-ATT = SINDEL-Attention. *p < 0.05; **p < 0.01; ***p < 0.001.

SINDEL is statistically significant. The superior performance of SINDEL against all the
baseline methods in Tables 4 and Table 5 demonstrates the effectiveness of the proposed
similarity network-based deep learning in extracting OUD treatment barriers.
To show the advantage of the similarity network branch in SINDEL, we conduct ablation
studies. We test the performance of the left branch only (SINDEL-Left) and the right branch
only (SINDEL-Right). The results are shown in Table 5. The right branch (similarity
network branch) significantly outperforms the left branch. Combining the left and right
branches fields the best performance over other baseline models. This result suggests the
similarity network component contributes to the performance gain significantly.
As the attention mechanism shows promising results in deep learning studies, we add
a self-attention mechanism in the right branch of SINDEL to search for the best model
design. The result reported in Table 5 (SINDEL-ATT) indicates that adding the attention
mechanism in SINDEL hampers the performance of SINDEL. Attention mechanism gives
more weights to the important part of a sentence, which is useful in sentence-level predic­
tion such as text classification. However, as we aim to predict a label for each word,
enhancing the important part of a sentence does not benefit the prediction at a smaller
granularity (word level). Attention mechanisms address the long-term dependency issue.
The benefit of attention is not evident in drug forums because irrelevant topics emerge in
long sentences, bringing in more noise than information gain. Therefore, we do not include
the attention mechanism in our model.
In order to check the robustness of SINDEL’s advantages over deep learning methods, we
conduct experiments with varying hyperparameters (the embedding size and the output
dimension of BLSTM) in SINDEL and other baseline models. We choose the commonly
adopted hyperparameter combinations (Online Supplemental Appendix A1) to perform
grid search. Figures 6, Figures 7, and Figures 8 depict the performance of their precision,
recall, and F1 score in this grid search. The results show SINDEL is robust and significantly
outperforms SINDEL-ATT, SINDEL-Right, SINDEL-Left, BLSTM, LSTM, and RNN in all
parameter combinations.
SINDEL utilizes word embedding and the similarity network as the representation. We
utilize word2vec to generate word embedding in the previous analyses. We further test the
robustness of our model by using other commonly adapted word embedding models,
including locally-trained Skip-gram, GloVe, FastText [28], and SeVeN [16]. Table 6
shows the performance of these models.
The word2vec model reaches the best performance among all word embedding models.
The locally-trained Skip-gram is inferior to word2vec because the word embedding training
size is smaller than that of word2vec. GloVe also has lower performance than word2vec
182 XIE ET AL.

Figure 6. Precision of SImilarity Network-based DEep Learning (SINDEL) and deep learning methods with
varying parameters.

Figure 7. Recall of SImilarity Network-based DEep Learning (SINDEL) and deep learning methods with
varying parameters.

because many of the discussions in the entire dataset drift from OUD treatment. The global
representation induces noisy information. Other state-of-the-art embedding models,
including FastText and SeVeN, also result in lower performance than our approach. We
further test the performance of SINDEL with various embedding models in different
hyperparameter settings (the embedding size and the output dimension of BLSTM).
Figures 9, Figures 10, and Figures 11 depict the performance curves of their precision,
recall, and F1 score when the key parameters vary. The hyperparameter description and
statistics are shown in Online Supplemental Appendix A2. The results indicate that the
word2vec model is the best embedding model for SINDEL. The performance is also robust
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 183

Figure 8. F1 Score of SImilarity Network-based DEep Learning (SINDEL) and deep learning methods with
varying parameters.

Table 6. Evaluation of SImilarity Network-based DEep Learning (SINDEL) using different embeddings.
Method Precision (Percent) Recall (Percent) F1 score (Percent)
SINDEL with Skip-gram 80.09*** 68.56** 73.87***
SINDEL with GloVe 83.84* 59.14*** 69.33***
SINDEL with FastText 80.34*** 67.10*** 73.10***
SINDEL with SeVeN 81.42*** 68.71 74.51***
SINDEL (Ours) 85.31 70.14 76.97
Notes: *p < 0.05; **p < 0.01; ***p < 0.001.

in various parameter settings. We, therefore, choose word2vec as the embedding model in
SINDEL. SINDEL incorporates a word similarity network to address the challenge of
various morphs of drugs and treatments. This similarity network representation for each
word consists of a set of semantically similar words. We perform a sensitivity analysis to
select the optimal network size in the similarity network representation. Figure 12 and
Online Supplemental Appendix A3 summarize the sensitivity analysis results. We observe
that the performance of SINDEL gradually improves when the network size grows from 1 to
3 words. Its performance stabilizes from 3 to 8 words with minor fluctuations. The
performance decreases at 9 words due to over-complex model specifications.
Figure 12 shows that the performance of SINDEL is the highest when the network size is
3, 7, or 8 words. To select the best network size in the similarity network representation, we
conduct t-tests to show their performance differences. The t-test results in Table 7 indicate
the performance has no significant difference when the network size is 3, 7, or 8 words.
Therefore, we select 3 similar words in the similarity network representation because the
computational complexity is the lowest and performance is the highest.
The similarity network representation captures the interconnected relationships among
similar morphs, thus identifying more morphs of OUD treatment barriers together with the
original term. In Table 8, we show examples of the additional morphs that the similarity
network could capture. In addition to heroin, the similarity network could capture
184 XIE ET AL.

Figure 9. Precision of SImilarity Network-based DEep Learning (SINDEL) using different word embeddings
with varying parameters.

Figure 10. Recall of SImilarity Network-based DEep Learning (SINDEL) using different word embeddings
with varying parameters.

dopesick, H, and chinawhite. Therefore, OUD treatment barriers mentioned together with
dopesick, H, and chinawhite could be identified as well. Ache is a withdrawal reaction after
treatment, which can be a barrier that prevents patients from receiving treatment. Other
types of barrier symptoms (achey, sore, bites) are also extracted by the similarity network.
Methadone is an OUD treatment medication. Patients may mention it when describing
treatment barriers, such as methadone addiction and withdrawal. Morphs of methadone,
such as meth, M, and MMT, are extracted together to improve the model performance.
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 185

Figure 11. F1 Score of SImilarity Network-based DEep Learning (SINDEL) using different word embed­
dings with varying parameters.

Figure 12. SImilarity Network-based DEep Learning (SINDEL’s) performance with varying number of
similar words.

Table 7. P-Value of T-tests for the Results with Different Network Sizes.
Pairwise T-tests P-value (Precision) P-value (Recall) P-value (F1 score)
3 similar words vs 7 similar words 0.953 0.234 0.188
3 similar words vs 8 similar words 0.669 0.196 0.271
7 similar words vs 8 similar words 0.648 0.888 0.946

Clustering OUD Treatment Barriers

We apply SINDEL to extract OUD treatment barriers from the entire research data. We,
then, leverage K-means to identify the general types of barriers. K-means generated 40
186 XIE ET AL.

Table 8. Examples of morphs captured by the similarity network

representation.
Original Term Similar Morphs Captured by the Similarity Network
Heroin Dopesick, H, chinawhite
Ache Achey, sore, bites
Methadone Meth, M, MMT

clusters of barriers. To choose the optimal number of clusters, a medical expert panel,
including two biomedical researchers, examined these clusters for medical relevance. The
panel merged the clusters that have similar medical relevance. Thirteen clusters are identi­
fied by the expert panel. Table 9 shows the types of OUD treatment barriers. To evaluate the
reliability of the 13 clusters, an independent biomedical researcher read a random sample of
1,000 barriers and labeled the barrier types. Among these 1,000 barriers, 876 (87.6 percent)
share the same barrier type as labeled by the clustering model, indicating good reliability.
The results shed valuable insights to understand patient’s decisions about receiving OUD
treatment. Lack of motivation is the most common barrier to receiving OUD treatments
(24.67 percent). Prior survey shows that self-motivation is a prominent reason for long-
term recovery from opioid addiction [18]. Although these patients are aware of the con­
sequences of addiction, they postpone visits to health providers. To motivate these patients,
community support, family encouragement, and follow-ups from the clinics are essential.
While physically attending support group meetings or counseling may be infeasible for

Table 9. Types of opioid use disorder (OUD) treatment barriers.

Type Description Percentage Examples
Lack of motivation The patient does not have motivation 24.67 Don’t have time to go to a clinic, CANT
to quit opioids stop, don’t want to quit
Lack of medical literacy The patient lacks knowledge of 21.88 Wasting time, dying of boredom, don’t
consequences of addiction have desire to be clean
Concerns about social The patient is concerned about social 12.67 Fails a pre employment drug screen,
stigma and job stigma or afraid of losing jobs new job requires no drug or
opportunities medication use
Afraid of withdraw The patient is afraid of the 12.35 Hate withdrawals, precipitated
reactions withdrawals after quitting withdrawals
Side effects of treatment The patient is concerned about the 9.13 Headaches, migraines, insomnia
side effects of treatment
Reliance because of The patient cannot stop opioids 5.64 I’m sick in pain, chronic pain flares up
chronic pain/fatigue because of chronic pain
Concerns about The patient is concerned about 3.85 Methadone addiction
buprenorphine/ buprenorphine or methadone
methadone addiction addiction
High cost of treatment The patient cannot afford the 2.91 Expensive, unaffordable
treatment or insurance does not
cover
Poor patient-physician The patient does not have good 2.19 Clinic denies me, doc was pissed
relationship relationship with the providers
Enjoy euphoric feeling of The patient enjoys the euphoric 1.70 Like dope, craving
drugs feeling of opioids and does not
want to quit
Depressed mental status The patient is depressed and does not 1.57 Severe depression
want to receive treatment
Lack of accessibility Treatment is not accessible to 0.63 No rehab
patients
Others Others 0.82 Can’t, good
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 187

patients who have chronic conditions and disability, virtual support communities can form
in online forums and provide recovering patients with support beyond that of traditional
counseling.
Lack of health literacy prevails as a common barrier for OUD treatment. Recent studies
in medicine have identified that health literacy has a significant negative association with
opioid misuse and the severity of opioid dependence [49]. Our findings revealed that
patients may not understand the ramifications of opioid addiction due to health literacy
and suffer from prolonged harm caused by OUD. Interventions targeting health literacy
among this group of patients may help to address the opioid public health crisis. Online
forums offer the medium to raise awareness of OUD and quality health resources from
reliable channels, while the data driven approach can help identify the audience in need.
Other types of barriers to OUD treatment include social stigma and fear of job loss,
withdrawal reactions and side effect of treatment, physical reliance, concerns about metha­
done or buprenorphine addiction, cost of treatment, poor patient-physician relationship,
euphoric feeling, mental status, lack of accessibility, and more.
In addition to the barriers that confirm prior literature, we also uncovered new barriers that
have not been noted in prior survey studies, such as side effects of treatment, concerns about
buprenorphine or methadone addiction, poor patient-physician relationships, and depressed
mental status. These barriers have not been identified by survey studies because these barriers
are sensitive and involve personal behavior. The patients are willing to share these undisclosed
opinions in drug forums because of the anonymity. Uncovering these new barriers enables
a deeper understanding of patients’ opioid addiction behavior and facilitates more effective
intervention strategies. For instance, poor patient-physician relationship prevents patients
from seeking addiction treatment. Overlooking this perspective devastates the efforts of other
stakeholders. In order to improve opioid addiction management, health providers need to
foster a healthy relationship with their patients. Depressed patients also have little motivation
to seek treatments. This new finding provides guidance for caregivers to treat the depression
disorder of the patients together with OUD. The new findings further prove the value of social
media platforms in opioid research. Future studies could also harness health social media to
understand many other patient behaviors, such as medication compliance and drug abuse.
As our automated method could continuously collect data and provide up-to-date
insights about emerging treatment barriers, our method could be utilized as a surveillance
tool to unveil more new barriers over time. It also assists proactive decision-making. Key
stakeholders, such as hospitals, health providers, insurance companies, and policymakers,
could ensure the barriers related to their side are properly dealt with for new patients.

Second Case Study in Extracting Barriers to Medication Adherence

In order to test the generalizability of SINDEL, we evaluate SINDEL on a different context
with a different dataset. SINDEL is designed to discover barriers to human decision-making
in textual data. We leverage SINDEL to extract the barriers to medication adherence from
WebMD forum data. We collected 233,325 sentences from 53,180 reviews about 180 drugs
in WebMD. These reviews are patient self-reported experience of taking medications, many
of which contain information of the barriers to medication adherence. In the same manner
of extracting OUD treatment barriers, we extract the phrases about barriers to medication
adherence. We randomly selected 5,400 sentences for annotation. The annotation scheme is
188 XIE ET AL.

Table 10. Evaluation of SImilarity Network-based DEep Learning (SINDEL) in medication

nonadherence case.
Method Precision (Percent) Recall (Percent) F1 score (Percent)
SVM 68.91*** 79.64*** 73.89***
LR 68.73*** 80.63*** 74.21***
NB 41.30*** 71.21*** 52.28***
CRF 94.00*** 46.30*** 62.04***
RNN 83.72*** 81.60*** 82.57***
LSTM 85.43*** 82.24*** 83.76***
BLSTM 88.26 85.94** 87.06***
SINDEL-ATT 87.73 86.56** 87.12***
SINDEL-Right 87.80 87.05*** 87.39***
SINDEL (Ours) 88.13 88.09 88.10
Notes: SVM = support vector machine; LR = logistic regression; NB = Naïve Bayes; CRF = conditional
random fields; RNN = recurrent neural network; LSTM = long short-term memory; BLSTM = bidirectional
long short-term memory; SINDEL-ATT = SINDEL-Attention.

IOB. Five expert annotators independently annotated these sentences in five batches.
A sixth expert annotator reviewed all sentences independently. The Kappa value is 0.98,
indicating excellent reliability. The performance of SINDEL and the benchmarks are
reported in Table 10.
In the medication nonadherence context using WebMD data, SINDEL consistently
outperforms all the baseline models. Using the WebMD data, we further compare
SINDEL with alternative representations, including Skip-gram, GloVe, FastText, and
SeVeN. The results, reported in Online Supplemental Appendices A8 and A9, show that
SINDEL outperforms all the other representations. In order to test the robustness of
SINDEL in different parameter settings in the medication nonadherence context, we
evaluate SINDEL, baseline models, and alternative representations in the 36 hyperpara­
meter settings used in the OUD experiments (Online Supplemental Appendix A7). SINDEL
is able to consistently achieve the best performance, as reported in Online Supplemental
Appendices A4-A6. We also performed robustness checks of different representations
(SINDEL, Skip-gram, GloVe, FastText, and SeVeN) on different hyperparameter settings.
The results, reported in Online Supplemental Appendix A10, suggest SINDEL is the best
model among other representations in all hyperparameter settings.

Discussion
Contributions to IS Literature
This study seeks to address a high-impact health IT problem with significant societal
relevance: opioid use disorder. Following the design science research guidelines, we develop
a novel information system to understand the barriers to OUD treatments and conduct
comprehensive and rigorous evaluations of this information system. This system can be
utilized by multiple stakeholders in IS and healthcare. The design principles in the proposed
method, including model selection, parameter settings, and model architecture, provides
methodological and practical implications for IS research.
This study also fits into the computational data science research [48]. The computational
data science research emphasizes an “interdisciplinary approach in developing novel data
representations, computational algorithms, business intelligence, and analytics methods”
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 189

[48]. Our study develops an interdisciplinary approach that involves a novel computational
algorithm and an analytical solution to a major healthcare problem, thus holding great
potential for generating IS research with significant societal impact. Such impact includes
the capability of advancing the understanding of OUD treatment issues from drug users’
perspective and discovering useful information from multi-scale biomedical data to support
healthcare delivery.
In the same spirit of social media-based IS research [4, 30, 31, 63], we leverage an IT
system from an online health community, opioid forums, to assist the understanding of
opioid addiction. This new IT-enabled application offers unprecedented scale and potential
to deepen the investigation of the opioid crisis.

Methodological Implications for IS

We devise a deep learning framework to extract OUD treatment barriers from social media,
including OUD treatment barrier extraction and clustering. This framework can be adapted to
many other IS research problems aiming to understand barriers to human behaviors. For
instance, this framework could be utilized to extract the barriers to technology adoption, the
barriers to investing in projects, and the barriers to medication adherence, among others. We
develop a fine-grained deep learning model that captures the semantic meaning of words and
the similarity network of text. Our proposed model improves upon standard deep learning
models with a novel similarity network-based representation and a multi-view deep learning
architecture. The similarity network-based component captures the interconnected relation­
ships among morphs in drug forums. As shown in the empirical analyses, the similarity
network helps identify many more OUD treatment barriers that are written as morphs. With
this new method, patient intelligence can be fruitfully exploited from health social media. This
method can be generalized to many other information retrieval tasks that involve morphs,
such as cyber threat identification in hacker forums, aspect mining in product reviews, and
drug surveillance in health social media. Table 11 shows the specific implications.

Contributions to Design Theory

In line with [24], this study proposes a novel analytics method as an IT artifact. This IT
artifact contributes four general design principles to design theory: 1) Incorporating word
networks into language models could disentangle complex word semantics; 2) Designing
multi-dimensional data representations (word embedding and network embedding) could
enrich representation learning; 3) A subsequent clustering module after classification
modules could provide interpretable and actionable insights for prescriptive analytics; 4)
Knowledge discovery on social media could complement survey-based approaches on

Table 11. Methodological implications for other information systems research.

Implications Using SImilarity Network-based
Domain Data Morphs DEep Learning (SINDEL)
Cybersecurity Hacker forums Hacker slang Identify hacking tools
Ecommerce Product reviews Consumer vocabulary Extract product features
Health analytics Health social media Drug street names Extract drug adverse events
190 XIE ET AL.

human behavior understanding. These design principles offer generalized implications for
other design science research, especially in social media analytics. To ensure problem
relevance, we stressed the technical challenges in health social media mining as well as its
significance. Guided by the technical challenges, we summarized the plausible methods and
proposed the best approach through conceptual and empirical model selection. The model
selection process and hyperparameter tuning could be leveraged as a generalized design
guideline for other social media mining studies in IS.

Managerial Implications
This study is among the first attempts to explore drug forums to understand patient decision-
making about OUD treatment. Many new OUD treatment barriers are uncovered from
patient narratives in drug forums. Our findings provide valuable implications for key stake­
holders, including patients, caregivers, health providers, insurers, and policymakers. Poor
medical literacy significantly prohibits patients from receiving OUD treatment. Awareness of
the barriers to OUD treatment can provide patients with social and informational support and
motivate them to participate in online health communities to improve their understanding of
opioid addiction. Caregivers play an important role in improving OUD treatment. Many
patients lack the motivation or information to receive treatment. Caregivers need to actively
seek treatment information and encourage patients to seek medical assistance. Understanding
the treatment barriers also help caregivers seek targeted assistance to connect patients to
practitioners. Poor patient-physician relationship could hinder patients’ attempt to seek OUD
treatment. Health providers could reinforce their relationship with patients and provide
informational support for patients. Our automated model can also actively collect social
media data and identify emerging treatment barriers which assist physicians’ decision-
making. Insurance plan design can encourage or discourage treating OUD. Patients with
high-deductible insurance plans are less likely to receive treatments because of the financial
burden. Insurance plans could encourage patients to receive treatments by reducing the
copayments and deductibles. A comprehensive examination of OUD treatment barriers also
provides a holistic view of the opioid crisis. Policymakers could design intervention policies to
tackle these treatment barriers and ensure stakeholders all work in line with the same
objective. The proposed analytical method also serves as a decision support system for
policymakers to monitor emerging public health problems about opioids. We summarize
the managerial implications for each stakeholder in the Table 12.

Limitations and Future Directions

Our study has several limitations and areas for improvement. First, we incorporated
a similarity network into our model as a multi-view learning framework. Future research
can explore alternative mechanisms to integrate two representations. Second, although our
framework can be generalized to process other text containing morphs, we only tested our
method on opioid forums and WebMD. To extend this research and expand the external
validity, future studies could use our framework to understand the barriers to technology
adoption, barriers to investing in projects, and more. Third, the barriers identified in this
study are the most direct barriers. These barriers could potentially be further caused by
other indirect reasons. As the users only report the most direct barriers, we are not able to
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 191

Table 12. Managerial implications for stakeholders.

Stakeholder IT Artifact Implication
Patients Treatment barriers Raise awareness of barriers and solutions
Caregivers Treatment barriers Raise awareness of barriers and solutions
Physicians Treatment barriers, analytical method Monitor barriers using our method and intervene actively
Insurers Treatment barriers Address barriers resulted from insurance limitation
Policymakers Treatment barriers, analytical method Monitor opioid users’ behavior and public health crisis
using the automated method

observe the hidden barriers. Despite this limitation, the most direct barrier offers the more
direct and effective intervention insights.

Conclusion
OUD is an urgent public health crisis that causes serious medical and financial ramifica­
tions. Understanding OUD treatment barriers forms the premise to alleviate this crisis. Our
research aims to develop a computational approach to understanding OUD treatment
barriers from the patients’ perspective. We designed a novel deep-learning-based approach,
SINDEL, to collect relevant patient discussions from drug forums, extract the OUD treat­
ment barriers, and analyze the types of barriers. This is among the first attempts to analyze
OUD treatment barriers from large-scale health social media data. We identified common
barriers to OUD treatment, such as lack of medical literacy, concerns about job loss,
concerns about withdrawal reactions and side effects, and more. Knowing these barriers,
the key stakeholders, including physicians, patients, policymakers, pharmaceutical compa­
nies, and healthcare systems, could gain rich insights from the patient perspective and take
proactive interventions to avoid harmful outcomes.
In line with the design science research methodology, we rigorously evaluated our model
and compared it with state-of-the-art baseline models in two different contexts and two
different datasets. SINDEL outperforms all the baseline models, attributed to the similarity
network-based component. SINDEL can be generalized in many other information retrieval
tasks involving morphs and applied to extract consumers’ opinion and understand their
decision-making in various business contexts.

Notes
1. We only select drug use-related subforums, such as methadone, opiate, and opioid addiction,
and more. We do not collect the subforums irrelevant to drug use, such as drug laws and forum
announcement.

Acknowledgements
This work was supported in part by the following grants: Grant Nos. 2020AAA0108401, 2017YFC0820105,
2019QY(Y)0101, and 2020AAA0103405 from the Ministry of Science and Technology of China, Grant No.
2017ZX10303401-002 from the Ministry of Health of China, and Grant Nos. 71621002, 72074209,
71974187, and 71472175 from the National Natural Science Foundation of China. Part of the experiments
were run on the El Gato supercomputer that was supported by the National Science Foundation under
Grant No. 1228509.
192 XIE ET AL.

References
1. Abbasi, A.; Albrecht, C.; Vance, A.; and Hansen, J. Metafraud: A meta-learning framework for
detecting financial fraud. MIS Quarterly, 36, 4 (2012), 1293–1327.
2. Abbasi, A.; and Chen, H. CyberGate: a design framework and system for text analysis of
computer-mediated communication. MIS Quarterly, 32, 4 (2008), 811–837.
3. Abbasi, A.; Zhang, Z.; Zimbra, D.; Chen, H.; and Nunamaker, J.F. Detecting fake websites: The
contribution of statistical learning theory. MIS Quarterly, 34, 3 (2010), 435–461.
4. Abbasi, A.; Zhou, Y.; Deng, S.; and Zhang, P. Text analytics to support sense-making in social
media: A language-action perspective. MIS Quarterly, 42, 2 (2018), 427–464.
5. Agarwal, R.; and Dhar, V. Big data, data science, and analytics: The opportunity and challenge
for IS research. Information Systems Research, 25, 3 (2014), 443–448.
6. Andrilla, C.H.A.; Coulthard, C.; and Larson, E.H. Barriers rural physicians face prescribing
buprenorphine for opioid use disorder. Annals of Family Medicine, 15, 4 (July 2017), 359–362.
7. Aral, S.; and Walker, D. Creating social contagion through viral product design: A randomized
trial of peer influence in networks. Management Science, 57, 9 (2011), 1623-1639.
8. Baesens, B.; Bapna, R.; Marsden, J.R.; Vanthienen, J.; and Zhao, J.L. Transformational issues of
big data and analytics in networked business. MIS Quarterly, 40, 4 (2016), 807–818.
9. Baldwin, T.; Kim, Y.-B.; Catherine De Marneffe, M.; Ritter, A.; Han, B.; and Xu, W. Shared
Tasks of the 2015 Workshop on Noisy User-generated Text: Twitter Lexical Normalization and
Named Entity Recognition. Proceedings of the ACL 2015 Workshop on Noisy User-generated
Text, 2015, pp. 126-135.
10. Bojko, M.J.; Mazhnaya, A.; Marcus, R.; Islam, Z.; Filippovych, S.; Dvoriak, S.; Altice, FL. The
future of opioid agonist therapies in Ukraine: A qualitative assessment of multilevel barriers
and ways forward to promote retention in treatment. Journal of Substance Abuse Treatment,
66, (July 2016), 37–47.
11. Bughin, J.; and Chui, M. The rise of the networked enterprise : Web 2.0 finds its payday.
McKinsey Quarterly, 4(2010), 1–6.
12. Capurro, D.; Cole, K.; Echavarría, M.I.; Joe, J.; Neogi, T.; and Turner, A.M. The use of social
networking sites for public health practice and research: A systematic review. Journal of
medical Internet research, 16, 3 (March 2014), e79.
13. Chen, H. Call for papers MISQ special issue on the role of information systems and analytics in
chronic disease prevention and management. MIS Quarterly, 41, 1 (2017), 1–3.
14. Chen, H.; Chiang, R.H.L.; and Storey, V.C. Business intelligence and analytics: From big data to
big impact. MIS Quarterly, 36, 4 (2012), 1165–1188.
15. Council of Economic Advisers. The Underestimated Cost of the Opioid Crisis. Executive Office
of the President of the United States, Council of Economic Advisers; 2017.
16. Espinosa-Anke, L.; and Schockaert, S. SeVeN: Augmenting word embeddings with unsuper­
vised relation vectors. (August 2018), pp.2653–2665.
17. Fang, X.; Hu, P.J.H.; Li, Z.L.; and Tsai, W. Predicting adoption probabilities in social networks.
Information Systems Research, 24, 1 (March 2013), 128–145.
18. Flynn, P.M.; Joe, G.W.; Broome, K.M.; Simpson, D.D.; and Brown, B.S. Recovery from opioid
addiction in DATOS. Journal of Substance Abuse Treatment, 25, 3 (2003), 177–186.
19. Goh, J.M.; Gao, G.; and Agarwal, R. The creation of social value: Can an online health
community reduce rural-urban health disparities? MIS Quarterly, 40, 1 (2016).
20. Gregor, S.; and Hevner, A. Positioning and presenting design science research for maximum
impact. MIS Quarterly, 37, 2 (June 2013), 337–355.
21. Hassamal, S.; Goldenberg, M.; Ishak, W.; Haglund, M.; Miotto, K.; and Danovitch, I.
Overcoming barriers to initiating medication-assisted treatment for heroin use disorder in
a general medical hospital. Journal of Psychiatric Practice, 23, 3 (May 2017), 221–229.
22. He, Y.; Chakrabarti, K.; Cheng, T.; and Tylenda, T. Automatic discovery of attribute synonyms
using query logs and table corpora. In International World Wide Web Conferences Steering
Committee, 2016, pp. 1429–1439.
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 193

23. Hemmatian, F.; and Sohrabi, M.K. A survey on classification techniques for opinion mining
and sentiment analysis. Artificial Intelligence Review, 52, 3 (2019), 1495–1545.
24. Hevner, A.R.; March, S.T.; Park, J.; and Ram, S. Design science in information systems
research. MIS Quarterly, 28, 1 (2004), 75–105.
25. Huang, C.-Y.; Chen, M.-H.; and Ku, L.-W. Towards a better learning of near-synonyms:
automatically suggesting example sentences via fill in the blank. In International World Wide
Web Conferences Steering Committee, 2017, pp. 293–302.
26. Hutchinson, E.; Catlin, M.; Andrilla, C.H.A.; Baldwin, L.-M.; and Rosenblatt, R.A. Barriers to
primary care physicians prescribing buprenorphine. Annals of Family Medicine, 12, 2
(March 2014), 128–133.
27. Jeyapriya, A.; and Selvi, C.S.K. Extracting aspects and mining opinions in product reviews
using supervised learning algorithm. In 2015 2nd International Conference on Electronics and
Communication Systems (ICECS). IEEE, 2015, pp. 548–552.
28. Joulin, A.; Grave, E.; Bojanowski, P.; Douze, M.; Jégou, H.; and Mikolov, T. FastText.zip:
Compressing text classification models. arXiv preprint arXiv:1612.03651. (December 2016).
29. Kallinikos, J.; and Tempini, N. Patient data as medical facts: Social media practices as
a foundation for medical knowledge creation. Information Systems Research, 25, 4 (2014),
817–833.
30. Karahanna, E.; Xu, S.X.; Xu, Y.; and Zhang, N. The needs-affordances-features perspective for
the use of social media. MIS Quarterly, 42, 3 (2018), 737–756.
31. Kitchens, B.; Dobolyi, D.; Li, J.; and Abbasi, A. Advanced customer analytics: Strategic value
through integration of relationship-oriented big data. Journal of Management Information
Systems, 35, 2 (2018), 540–574.
32. Knudsen, H.K.; Abraham, A.J.; and Oser, C.B. Barriers to the implementation of
medication-assisted treatment for substance use disorders: The importance of funding policies
and medical infrastructure. Evaluation and Program Planning, 34, 4 (November 2011),
375–381.
33. Krieger Pharm. D. C. What are opioids and why are they dangerous? Mayo Clinic, (2018).
Retrieved on February 20, 2021: https://newsnetwork.mayoclinic.org/discussion/what-are-
opioids-and-why-are-they-dangerous/
34. Lau, R.; Liao, S.; Wong, K.F.; and Chiu, D. Web 2.0 Environmental scanning and adaptive
decision support for business mergers and acquisitions. MIS Quarterly, 36, 4 (2012), 1239–68.
35. Li, S.; Zhou, L.; and Li, Y. Improving aspect extraction by augmenting a frequency-based
method with web-based similarity measures. Information Processing and Management, 51, 1
(2015), 58–67.
36. Liu, B. Sentiment Analysis and Opinion Mining. Synthesis lectures on human language tech­
nologies, 5, 1 (2012), 1–67.
37. Livingston, J.D.; Adams, E.; Jordan, M.; MacMillan, Z.; and Hering, R. Primary care physicians’
views about prescribing methadone to treat opioid use disorder. Substance Use and Misuse, 53,
2 (January 2018), 344–353.
38. Luo, W.; Zhuang, F.; Cheng, X.; He, Q.; and Shi, Z. Ratable aspects over sentiments: Predicting
ratings for unrated reviews. In Proceedings - IEEE International Conference on Data Mining,
ICDM. 2014, pp. 380–389.
39. Ma, B.; Zhang, N.; Liu, G.; Li, L.; and Yuan, H. Semantic search for public opinions on urban
affairs: A probabilistic topic modeling-based approach. Information Processing and
Management, 52, 3 (2016), 430–445.
40. Mai, F.; Shan, Z.; Bai, Q.; Wang, X. (Shane); and Chiang, R.H.L. How does social media impact
bitcoin value? A test of the silent majority hypothesis. Journal of Management Information
Systems, 35, 1 (2018), 19–52.
41. McKenna, R.M. Treatment use, sources of payment, and financial barriers to treatment among
individuals with opioid use disorder following the national implementation of the ACA. Drug
and Alcohol Dependence, 179, (October 2017), 87–92.
42. Nguyen, K.A.; im Walde, S.S.; and Vu, N.T. Integrating distributional lexical contrast into word
embeddings for antonym-synonym distinction. arXiv preprint arXiv:1605.07766.2016.
194 XIE ET AL.

43. NIDA. Overview | National Institute on Drug Abuse (NIDA). 2018. Retrieved on February 20,
2021: https://www.drugabuse.gov
44. Nogueira, C.; Santos, D.; and Zadrozny, B. Learning Character-level representations for part-of-
speech tagging. In International Conference on Machine Learning, 2014 Jun 18 (pp. 1818-1826).
PMLR.
45. Pew Research. 61% of American adults look online for health information | Pew Research
Center’s Internet & American Life Project. Pew Research Center, 2009. Retrieved on February
20, 2021: https://www.pewresearch.org/internet/2009/06/11/61-of-american-adults-look-
online-for-health-information/
46. Qiu, G.; Liu, B.; Bu, J.; and Chen, C. Opinion word expansion and target extraction through
double propagation. Computational Linguistics, 37, 1 (2011), 9–27.
47. Qu, M.; Ren, X.; and Han, J. Automatic synonym discovery with knowledge bases. In
Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, 2017, pp. 997–1005.
48. Rai, A. Editor’s comments: Diversity of design science research. MIS Quarterly, 41, 1 (2017),
iii–xviii.
49. Rogers, A.H.; Bakhshaie, J.; Orr, M.F.; Ditre, J.W.; and Zvolensky, M.J. Health literacy, opioid
misuse, and pain experience among adults with chronic pain. Pain medicine (Malden, Mass.),
21, 4 (2020), 670–676.
50. Rubenstein, H.; and Goodenough, J.B. Contextual correlates of synonymy. Communications of
the ACM, 8, 10 (October 1965), 627–633.
51. Saboo, A.R. Using big data to model time-varying effects for marketing resource (RE)
allocation. MIS Quarterly, 40, 4 (2016), 911–939.
52. Sha, Y.; Shi, Z.; Li, R.; Liang, Q.; and Wang, B. Resolving entity morphs based on character-
word embedding. Procedia Computer Science, 108, (2017), 48–57.
53. Sharma, A.; Kelly, S.M.; Mitchell, S.G.; Gryczynski, J.; O’Grady, K.E.; and Schwartz, R.P.
Update on barriers to pharmacotherapy for opioid use disorders. Current Psychiatry Reports,
19, 6 (June 2017), 35.
54. Simpson, S.S.; Adams, N.; Brugman, C.M.; and Conners, T.J. Detecting novel and emerging
drug terms using natural language processing: A social media corpus study. JMIR Public Health
and Surveillance, 4, 1 (January 2018), e2.
55. Smith, C.A.; and Wicks, P.J. PatientsLikeMe: Consumer health vocabulary as a folksonomy. In
AMIA Symposium. 2008, p. 682. American Medical Informatics Association.
56. Stieglitz, S.; and Dang-Xuan, L. Emotions and information diffusion in social media -
Sentiment of microblogs and sharing behavior. Journal of Management Information Systems,
29, 4 (2013), 217–248.
57. Stumbo, S.P.; Yarborough, B.J.H.; McCarty, D.; Weisner, C.; and Green, C.A. Patient-reported
pathways to opioid use disorders and pain-related barriers to treatment engagement. Journal of
Substance Abuse Treatment, 73, (February 2017), 47–54.
58. Tang, Q.; Gu, B.; and Whinston, A. Content contribution for revenue sharing and reputation in
social media: A dynamic structural model. Journal of Management Information Systems, 29, 2
(October 2012), 41–76.
59. Trusov, M.; Bodapati, A. V.; and Bucklin, R.E. Determining influential users in internet social
networks. Journal of Marketing Research, 47, 4 (August 2010), 643–658.
60. Samhsa. Results from the 2010 National Survey on Drug Use and Health: Detailed Tables.
(2011). Retrieved on February 20, 2021: https://www.samhsa.gov/data/sites/default/files/
NSDUHNationalFindingsResults2010-web/2k10ResultsRev/NSDUHresultsRev2010.pdf
61. Xie, K.; and Lee, Y.J. Social media and brand purchase: Quantifying the effects of exposures to
earned and owned social media activities in a two-stage decision making model. Journal of
Management Information Systems, 32, 2 (April 2015), 204–238.
62. Yao, Z.; Sun, Y.; Ding, W.; Rao, N.; and Xiong, H. Dynamic word embeddings for evolving
semantic discovery. In Proceedings of the eleventh acm international conference on web search
and data mining, 2018, pp. 673–681.
 JOURNAL OF MANAGEMENT INFORMATION SYSTEMS 195

63. Zhang, D.; Zhou, L.; Kehoe, J.L.; and Kilic, I.Y. What online reviewer behaviors really matter?
Effects of verbal and nonverbal behaviors on detection of fake online reviews. Journal of
Management Information Systems, 33, 2 (April 2016), 456–481.
64. Zhang, L.; and Liu, B. Aspect and entity extraction for opinion mining. In Data mining and
knowledge discovery for big data, Springer Berlin Heidelberg, 2014, pp. 1–40.
65. Zhang, W.; and Ram, S. A comprehensive analysis of triggers and risk factors for asthma based
on machine learning and large heterogeneous data sources. MIS Quarterly
66. Zhang, Y.; Jatowt, A.; Bhowmick, S.S.; and Tanaka, K. The past is not a foreign country:
Detecting semantically similar terms across time. IEEE Transactions on Knowledge and Data
Engineering, 28, 10 (October 2016), 2793–2807.

About the Authors

Jiaheng Xie ([email protected]; corresponding author) is an assistant professor in the Department of
Accounting and MIS at the University of Delaware’s Alfred Lerner College of Business and
Economics. His research interests lie in deep learning, health risk analytics and business analytics.
Dr. Xie’s prior work has been published or is under review at several leading journals, including MIS
Quarterly, Information Systems Research, and Journal of American Medical Informatics Association.
Zhu Zhang ([email protected]) is an assistant professor at the Institute of Automation, Chinese
Academy of Sciences. He received his Ph.D. degree in Computer Science from the University of
Chinese Academy of Sciences. His research interests include data mining, artificial intelligence, health
informatics, and business analytics. He has published a number of papers in journals and conference
proceedings, including ACM Transactions on MIS, Knowledge-Based Systems, and International
Conference on Information Systems.
Xiao Liu ([email protected]) is an assistant professor in the Department of Information Systems,
Arizona State University. Dr. Liu’s research interests are in data science and information system
design, in areas such as health IT, social media analytics and predictive analytics.
Daniel Zeng ([email protected]) is a professor in the Institute of Automation, Chinese Academy of
Sciences, and the University of Chinese Academy of Sciences. He is the head of Shenzhen Artificial
Intelligence and Data Science Research Institute (Longhua). He received his Ph.D. in industrial
administration from Carnegie Mellon University. He serves as the Editor-in-Chief of ACM TMIS
and the associate editor of INFORMS JOC. He is a fellow of IEEE and AAAS.