Understanding Medication Nonadherence from Social Media: A - - PowerPoint PPT Presentation

Understanding Medication Nonadherence from Social Media: A Sentiment-Enriched Deep Learning Approach

Presented by: Xiao Liu Department of Operations and Information Systems University of Utah Joint Work with Jiaheng Xie1, Xiao Fang2, Daniel Zeng1

1 Department of Management Information Systems, University of Arizona 2 Department of Accounting and Management Information Systems, University of Delaware

2

What is Medication Nonadherence

• Definition: patients not take medications as

recommended (Marcum et al. 2013, JAMA)

• Types of nonadherence (Servick 2014, Science)
• Non-fulfillment: not initiate
• Non-conforming: take at the wrong time
• Non-persistence: discontinue after starting

3

Implications of Medication Nonadherence – Patients’ Health

• Detrimental to patients’ health
• Drug withdrawal symptoms (Fox et al. 2014, NEJM)
• Treatment failure, disease worsening (Marcum et al. 2013, JAMA)
• Emergency visit, hospitalization (Traverso & Langer 2015, Nature)

Examples of withdrawal symptoms

4

Implications of Medication Nonadherence – Disease Management

• Critical to disease management
• ≥50% patients not follow therapy (Traverso & Langer 2015, Nature)
• Impact on health: increasing adherence rate ≫ improvements in treatment (Shank 2011, NEJM)
• Costs to healthcare systems
• Account for 19% drug-related visits to emergency rooms (Bailey et al. 2014)
• $290 billion preventable costs annually in US (Traverso & Langer 2015, Nature)

Costs of medication nonadherence

5

Medication Nonadherence Research

Caregivers Patients Pharmaceutical companies

Medical and healthcare system’s perspective

• 1. Harmful impact of

medication nonadherence (Gaebel et al. 2016; Serper et al. 2014)

• 2. Health policy to address

the issue (Viehman et

• al. 2016; Billimek et al.

2015) Perspectives from patients

• Vital for disease

mitigation & management (Adler & Stead 2015, NEJM)

• Why discontinuing

medications? Prior studies

• Focused patient group &

drug class

• Survey: cross-sectional,

expensive

Physicians & healthcare systems

6

Research Motivation

Design a deep learning-based framework for real-time medication nonadherence reason detection in health social media Methodological

• 1. Define reason mining problem
• 2. Sentiment-enriched word representation

– Capture characteristics of patient-generated text and represent noisy patients’ vocabulary – Adjustable based on research context

• 3. Deep learning-based framework for reason extraction

Healthcare IT Investigated medication nonadherence reasons comprehensively by leveraging online health IT platforms

Research objective

7

Literature Review

• Medication nonadherence
• Related work in the medical domain
• Aspect mining
• Approach to understanding factors that lead to certain opinion and

behaviors

8

Medication Nonadherence: Three Research Streams

Limitations:

Category Author Year Method Drug/Drug class Sampl e size Key findings Prevalence of medication nonadherence Viehman et al. 2016 Survey Nafcillin, Oxacillin 224 18% patients discontinued therapy Williams et al. 2014 Survey Diabetes 1,264 16% nonadherence rate Marcum et al. 2013 Survey Cardiovascular 897 40.7% nonadherence rate Du Pan et al. 2012 Cohort Rheumatoid arthritis 1,485 Higher retention rate for patients on non-aTNF-Bio Harmful

• utcomes of

medication nonadherence Gaebel et al. 2016 Survey Schizophrenia 19 Re-exacerbation after drug discontinuation Serper et al. 2014 Interview Liver 105 Adverse clinical outcome if discontinue Miller 2013 Survey Anticholinergic NA Withdrawal symptoms: Irritability, dysphoria, nausea Offord et al. 2013 Cohort Schizophrenia 873 Hospitalization, increased cost if discontinue Reasons for medication nonadherence Mak et al. 2016 Survey Schizophrenia 71 Lack of insight into illness à nonadherence Derya et al. 2015 Survey Bipolar, Schizophrenia 203 Unwilling to use meds, side effect à nonadherence Weid et al. 2015 Survey Diabetes 1,026 Negative expectation, econ burden à nonadherence Mago et al. 2014 Review Bipolar 207 Weight gain, perceived cognitive impairment, tremors, sedation à nonadherence

Table 1. Selected recent studies on medication nonadherence

Category Author Year Method Drug/Drug class Sampl e size Key findings Prevalence of medication nonadherence Viehman et al. 2016 Survey Nafcillin, Oxacillin 224 18% patients discontinued therapy Williams et al. 2014 Survey Diabetes 1,264 16% nonadherence rate Marcum et al. 2013 Survey Cardiovascular 897 40.7% nonadherence rate Du Pan et al. 2012 Cohort Rheumatoid arthritis 1,485 Higher retention rate for patients on non-aTNF-Bio Harmful

• utcomes of

medication nonadherence Gaebel et al. 2016 Survey Schizophrenia 19 Re-exacerbation after drug discontinuation Serper et al. 2014 Interview Liver 105 Adverse clinical outcome if discontinue Miller 2013 Survey Anticholinergic NA Withdrawal symptoms: Irritability, dysphoria, nausea Offord et al. 2013 Cohort Schizophrenia 873 Hospitalization, increased cost if discontinue

Table 1. Selected recent studies on medication nonadherence

9

Medication Nonadherence

Prevalent + serious heath outcomes à need to improve adherence rate

Take preventive actions and improve drug adherence worldwide à need to understand reasons for medication nonadherence comprehensively

10

Medication Nonadherence

Category Author Year Method Drug/Drug class Sample size Key findings Reasons for medication nonadherence Mak et al. 2016 Survey Schizophrenia 71 Lack of insight into illness à nonadherence Derya et al. 2015 Survey Bipolar, Schizophrenia 203 Unwilling to use medication + side effect à nonadherence Weid et al. 2015 Survey Diabetes 1,026 Negative expectation + economic burden à nonadherence Mago et al. 2014 Review Bipolar 207 Weight gain, perceived cognitive impairment, tremors, sedation à nonadherence

Table 1. Selected recent studies on medication nonadherence

Few nonadherence reasons detected Focused drug class & patient group

• Different regimens/effects across drug

classes/patient groups à Varied reasons across drug classes/patient groups Cross-sectional

• Emerging treatment options

à Changing reasons over time

11

Addressing the Limitations

Detect nonadherence reasons comprehensively & continuously Patient and drug-specific nonadherence reasons Need comprehensive & real-time dataset + automated approach

• Learn nonadherence reasons from heterogeneous patient groups

continuously

• Predict customized nonadherence reasons for each group timely
• Discover nonadherence reasons unnoted by previous studies

12

Comprehensive and Real-time Dataset

• Health social media: time-sensitive patients’ feedback about medications
• 61% US adults participating in online health community (Xu et al. 2016)
• Valuable for understanding medication nonadherence
• Social media in healthcare applications
• Peer/social/emotional support (Chiaramello et al. 2016)
• Drug safety surveillance (Xie et al. 2017; Nikfarjam et al. 2015)

à Rarely used by prior medication nonadherence studies

13

Technique to Approach the Problem

• Medication nonadherence reasons (Gellad et al. 2011)
• Attributes of drug: price, effectiveness, etc.
• Attributes of patient: emotion, health literacy
• Attributes of patient-provider relationship
• Other
• Aspect mining
• Extract attributes of objects
• E.g., aspect of an iPhone (price, screen size) à purchase behavior

Aspect mining

14

Aspect Mining

Task Author Year Method Data Object Aspects Performance Aspect extraction Wang et al. 2016 CRF 3,646 reviews Restaurant Food, price, service F-score: 63% Liu et al. 2016 CRF 5.8m reviews Camera Battery life, picture quality F-score: 73% Fang et al. 2015 LDA 5,632 reviews Tourist spot Architecture, park, food Perplexity: 3,375 Jeyapriya & Selvi 2015 Rule-based 100 reviews Camera Pictures, resolution, memory F-score: 80% Li et al. 2015 CRF 6,847 reviews Mouse Durability, compatibility Precision: 65% Brisson & Torrel 2015 CRF 40,160 reviews Smartphone Battery, camera F-score: 68% Marrese et al. 2014 CRF 200 reviews Hotel Price, food, service, pool F-score: 73% Hai et al. 2014 LDA 10,073 reviews Cellphone Screen, battery, fans, money F-score: 56% Patra et al. 2014 CRF 3,045 sentences Restaurant Service, price, food, ambience F-score: 72% Aspect clustering Chen et al. 2014 K-means 36k reviews Camera Battery, service, memory Coherence: -1k Xiong et al. 2016 K-means 1,389 sentences Camera Photo, battery RI: 59% Xiong et al. 2016 K-means 1,389 sentences Camera Photo, battery Entropy: 1.74 Wang et al. 2016 LDA 1,000 reviews Laptop Screen, battery, usefulness Precision: 79%

Table 2. Selected recent aspect mining studies

15

Applications of Aspect Mining

Aspects Food, price, service Battery life, picture quality Architecture, park, food, museum Pictures, resolution, memory Durability, compatibility, sensitivity Battery, camera Price, food, service, pool Screen, battery, fans, money Service, price, food, ambience Battery, service, memory Photo, battery Photo, battery Screen, battery, usefulness

Aspect of interest

Aspect of products (Liu et al. 2016; Li et al. 2015; Hai et al. 2014) à Lead to purchase behavior Aspect of tourist destinations (Fang et al. 2015; Marrese et al. 2014) à Lead to choice of destinations

16

Two tasks in Aspect Mining

Task Method Aspect extraction CRF CRF LDA Rule-based CRF CRF CRF LDA CRF Aspect clustering K-means K-means K-means LDA

Aspect extraction: core task

• Extract aspect expressions in text
• Actual words people used

à Express aspects of a behavior à “Expensive,” “cannot afford”

Aspect clustering

• Identify types of aspects
• Many expressions describing the same aspect:

“expensive,” “cannot afford” à price

• K-means, LDA à standard clustering problem

Tasks

17

Aspect Extraction Approaches

Aspect extraction approaches

Topic model: LDA Noisy topic clusters Cannot find types of nonadherence reasons Arbitrarily provide # of topics and assign labels to topics

Sequence model: CRFs Pinpoint words for each nonadherence reason expression

Not suitable for our study

Fang et al. 2015; Hai et al. 2014; Wang et al. 2016; Liu et al. 2016

18

Limitations of Sequence Models for Aspect Extraction

• Ineffective representation
• Treat words as atomic symbols (Mikolov et al. 2013)
• E.g. “cat” à “id01”, “dog” à “id02”
• Ignore relationships between individual symbols
• “cat”, “dog”: animals
• Need input with correct spelling
• “cat”, “catt”, “kitten”: different representations
• Hand-crafted features (Wang et al. 2016)
• Relatively lower performance with social media text
• Typos, informal words, emoticons

19

Limitations of Sequence Models for Aspect Extraction

• Data Imbalance Issue
• Need balanced data to train parameters
• Ignore minority data class (Cao et al. 2014)
• <10% words about medication nonadherence reason

20

Research Questions

• How to design a deep learning-based approach to
• Detect nonadherence reasons from patients’ perspectives
• Address technical challenges in processing health social media data
• What are the medication nonadherence reasons as reported in

health social media?

21

Design Rationale

• Two tasks to extract aspects of a behavior (Zhang & Liu 2014)
• Aspect extraction + aspect clustering
• Task 1: aspect extraction à expressions regarding nonadherence reasons
• Representation + characteristic of health text
• I. Sentiment-enriched representation
• Capture patients’ emotion + contextual information
• II. Adjustable representation as per characteristics of training text
• Amplify characteristics of health text
• Model structure
• Deep architecture + memory cell + bidirectional structure
• RNN + memory cell + bidirectional à BLSTM

Sentiment- enriched BLSTM (S-BLSTM)

22

Design Rationale

• Task 2: aspect clustering à group similar reasons à types of reasons
• State-of-the-art in literature review: k-means
• Incorporate word embeddings

à Represent semantic meaning of nonadherence reasons

Semantically- enriched k- means (SEKM)

23

Research Design

Figure 1. A deep learning Approach for medication nonadherence detection

24

High-Level Processing Logic

Figure 2. Flow of computational components

• Proposed S-BLSTM model: extract medication nonadherence reason

expressions from text

25

Task 1: Reason Expression Extraction (S-BLSTM)

Input sentence from corpus Integrate sentiment into word representation New representation: sentiment-enriched word representation S-BLSTM layer: two reversed LSTM layers with adjustable representation Softmax classifier to predict nonadherence reasons Figure 3. S-BLSTM architecture for medication nonadherence reason extraction The input expression is nonadherence reason or not

• Patient

ID Drug ID

26

S-BLSTM – Sentiment-Enriched Word Embedding

• Training sequence = [!", !$,…, !%], !&: word '
• Objective: what words likely to co-occur with word of interest
• Skip-gram: given word of interest, maximize probability of generating co-occurring words

Standard representation

• Objective function: L =

" % ∑+," %

∑-./0/.,012 log 6(!+80|!+)

• ;: number of training words
• <: window size of neighboring words

Proposed representation

• Incorporate sentiment information into the

new objective function =>

• => =

" % ∑+," % ? @A@BCDC ∑-./0/.,012 log 6(!+80|!+)

• EFGH'+: sentiment type ∈

{KLE'H'MF, NFOHPQR, NFSQH'MF}

27

S-BLSTM – Sentiment-Enriched Word Embedding

• Output: 300-dimensional vector for each word
• Capture contextual information (∑"#$%$#,%'( log ,(./0%|./)) and sentiment information (

3 4567898)

• Vector: 300 most likely co-occurring words with word of interest
• 300 dim: successfully tested in previous studies (Ma & Hovy 2016; Garten et al. 2015; Dima & Hinrichs

2015)

28

S-BLSTM – Adjustable Representation

• Input: sentiment-enriched embedding vector[!(#), !(&), … , !(()]

Standard approach

• !()) = !#

) , !& ) , … , !+,, )

• à pre-trained embedding

weights à static in the training process

• Input gate: .()) = /.01 2

3!()) + 53ℎ()7#) + 83

• Forget gate: 9()) = /.01 2

:!()) + 5 :ℎ()7#) + 8:

• Output gate: ;()) = /.01 2

<!()) + 5<ℎ()7#) + 8<

• Memory cell: =()) = .())⨀?@AℎB

C 2

D!()) + 5Dℎ()7#) +

8D + 9())⨀=()7#)

• Hidden state: ℎ()) = ;())⨀ tanh(=()))

Proposed approach

• Learning to adjust representation to magnify

characteristics of health text

• !′()) = J⨀! ) = JK!#

) , JL!& ) , … , JMNN!+,, )

• Input gate: .()) = /.01 2

3J⨀!()) + 53ℎ()7#) + 83

• Forget gate: 9()) = /.01 2

:J⨀!()) + 5 :ℎ()7#) + 8:

• Output gate: ;()) = /.01 2

<J⨀!()) + 5<ℎ()7#) + 8<

• Memory cell: =()) = .())⨀?@AℎB

C 2

DJ⨀!()) + 5Dℎ()7#) +

8D + 9())⨀=()7#)

• Hidden state: ℎ()) = ;())⨀ tanh(=()))

29

S-BLSTM – Model Structure

• Forward LSTM: first word à last word
• Hidden state ℎ(#): 128 dim
• Backward LSTM: last word à first word
• Hidden state ℎ′(#): 128 dim
• Output: concat(ℎ(#), ℎ′(#))
• Softmax classifier: whether input expression is nonadherence reason or

not

• Condense useful

information from 300 dim to 128 dim

• Successfully tested

in prior studies (Rao et al. 2015; Chan & Lane 2015)

30

Task 2: Reason Expression Clustering (SEKM)

• S-BLSTM extracts reason expressions (actual words)
• Many expressions describing the same reason

à “expensive”, “cannot afford” à aspect: high price

• Need to cluster reason expressions to identify reason types
• Propose semantically-enriched k-means (SEKM)
• Feature: sentiment-enriched embedding vector !(#)

à Group reasons based on semantic meaning

• Output: medication nonadherence reason clusters
• Why k-means in SEKM
• State-of-the-art in aspect clustering (Saboo et al. 2016; Füller et al. 2014; Hosanagar et al.

2013)

• Easy to implement, high performance (Hadian & Shahrivari 2014)

31

Data Collection

• Test bed: WebMD
• Well-known among patients
• Large-scale and streaming patients’ medication experience
• Used in multiple research areas (Huh et al. 2016; Nowrouzi et al. 2015; Nie et al.

2014)

• Timespan: 2005-2016
• Start of WebMD - collection date
• Data volume: 233,325 sentences à 53,180 patients à 180 drugs

32

Data Pre-Processing & Annotation

• Medication nonadherence indication identification
• Filter threads that indicate medication nonadherence
• Method: BLSTM (state-of-the-art text classification method)
• Used in many studies (Zhou et al. 2016; Liang & Zhang 2016; Fan et al. 2014)
• Accuracy in our dataset: 84.18%
• Annotation
• By 5 master students with information systems background
• Reasons for medication nonadherence: 5,400 sentences
• 11.3% words about medication nonadherence reasons

I have gained 50lbs. I just stopped the abilify and hope to lose the weight gained. Next I hope to stop the Cymbalta.

Reason for medication nonadherence

33

Evaluations

• Baseline models: state-of-the-art in literature review
• Discriminative machine learning model: SVM
• Frequently used text mining technique (Abbasi et al. 2010, 2012; Fang et al. 2013)
• Features: lexicon feature, suffix feature, POS feature, context feature (Björne et al. 2013; Ju et al. 2011;

Asif Ekbal 2010)

• Machine learning sequence model: CRF (CRFSuite)
• State-of-the-art in aspect mining
• Deep learning models: RNN, LSTM, BLSTM
• Evaluate the proposed enhanced deep learning model

• Table 3: average performance of 50 repetitions
• S-BLSTM > baseline models
• Resolve noisy text challenge: retrieved misspelled reasons (“anxiousnous”,“expenssive”)
• Alleviate imbalanced data issue: retrieved rare reasons

34

Model Performance

Table 3. Model performance

Evaluation performed in ten-fold cross validation Table 3. Model Performance Method Precision Recall F1 score SVM 29.30% 53.60% 37.90% CRF 94.00% 46.30% 62.04% RNN 85.70% 84.80% 85.20% LSTM 86.18% 86.84% 86.45% BLSTM 89.24% 87.04% 88.09% Our Approach 87.29% 93.27% 90.18%

35

Model Significance Tests

Table 4. P-value for paired t-tests for S-BLSTM against baseline models

• 50 repetitions for each model à t-tests
• S-BLSTM > baseline models (p<0.05)

Table 4. Pairwise T-tests for S-BLSTM against the Baseline Models Metric S-BLSTM vs SVM S-BLSTM vs CRF S-BLSTM vs RNN S-BLSTM vs LSTM S-BLSTM vs BLSTM Precision < 0.001*** (< 0.001***) 0.001** 0.535 0.006** Recall < 0.001*** < 0.001*** < 0.001*** < 0.001*** < 0.001*** F1 score < 0.001*** < 0.001*** < 0.001*** < 0.001*** < 0.001***

36

Model Performance Explained

• S-BLSTM > SVM/CRF
• Deep architecture retrieved rare reasons, SVM/CRF did not

à E.g., pregnant (0.04%) and cost (0.83%)

• S-BLSTM > RNN
• Memory cell handles long-range sequences
• E.g., in corpus: the shortest sentence = 1 word, the longest sentence =175 words.
• à S-BLSTM handles both short and long-length sequences
• à RNN only handles short sequences

37

Model Performance Explained Cont’d

• S-BLSTM > LSTM
• Bidirectional structure considers word relationships

à Retrieve more reasons when they are related

• E.g., “The drug caused headache, so I cannot sleep.”

à “headache” and “cannot sleep” are related à S-BLSTM retrieved both of them à LSTM just retrieved “headache”

• S-BLSTM > BLSTM
• Sentiment-enriched representation captures semantic + sentiment

à Retrieve more reasons

• E.g., “too expensive” à negative sentiment

à S-BLSTM captures semantic meaning + sentiment à more info à retrieved

38

Types of Medication Nonadherence Reasons

• S-BLSTM à expressions about nonadherence reasons
• SEKM to cluster reasons à identify reason types
• Ten types of medication nonadherence reasons (Table 5)
• Determined by two researchers, reviewed by a MD

Table 5. Types of Medication Nonadherence Reasons Medication Nonadherence Reason Type Description Percentage Examples Adverse event The medication has adverse events or leads to complications. 65.95% It caused a rash on my face. Need to stop. Drug switching The patient switches to another medication by his/herself. 10.40% Stopping Abilify. Switching to Geodon. Low health literacy The patient discontinues the medication because of low health literacy. 5.41% They give me too much of it! Don’t want to risk my life. Social influence The patient stops the medication because his/her peers/caregivers/friends/professionals encourage the patient to stop. 4.90% My mom told me that I should quit this drug. Cost prohibitive for patient The price of the drug is too high, or insurance does not cover. The patient cannot afford. 0.83% I can’t afford this expensive drug anymore. Complex medication plan The medication regimen is complicated. The patient does not like the complicated procedure. 0.48% It’s so annoying to take this drug three time a day. I’m not gonna do it. Medication ineffectiveness The medication is ineffective, so the patient stops it. 0.35% The drug does not help at all! I’m not taking it. Specific population The patient stops/reduces the medication because the patient is pregnant/is a child/has liver disease and more. 0.04% I'm pregnant and had to stop. Others Others 11.64% This is absolutely a nightmare! Stop and get better!

39

Model Performance by Reason Type

• S-BLSTM: accurate and robust for most reason types

à Can detect nonadherence reasons thoroughly in health social media

Table 7. Effectiveness of the Sentiment-Enriched Component by Reason Type Reason Type BLSTM S-BLSTM Reason Percentage Adverse event 92.51% 94.30% 65.95% Drug switching 91.20% 94.07% 10.40% Low health literacy 97.75% 98.41% 5.41% Social influence 91.53% 94.44% 4.90% Cost prohibitive for patient 9.77% 9.95% 0.83% Complex medication plan 99.06% 99.09% 0.48% Medication ineffectiveness 99.73% 98.57% 0.35% Specific population 50.00% 100.00% 0.04% Others 94.73% 95.20% 11.64%

40

Methodological Contributions

• Sentiment-enriched word representation
• Efficiently extract information from patient-generated content
• Adjustable representation based on research context
• Learn characteristics of health text
• Deep learning-based medication nonadherence reason detection

framework

• Precision delivery: patient-specific, real-time detection & intervention
• Generalizable to understand motivation for various behaviors
• Consumer retentions, technology adoption, crowd funding project investment

41

Application Contributions

• Healthcare IT
• 24,832 reasons for medication nonadherence
• Ten types of medication nonadherence reasons
• Unnoted by surveys: overdose risk, negative emotion, switching to other drugs
• Nonadherence reasons for 180 drugs & 60k patients
• Patient and drug-specific nonadherence reasons
• Practical

Nonadherence reasons Preventive measures Financial inability Copay assistance, identification, activation Adverse events Educate ways to mitigate Negative emotion Positive reinforcement & incentive Overdose risk Fixed combinations of medicines Inconvenience Day-of-the-week pill box, single-pill medication

42

Future directions

• Representation
• Incorporating syntactic information (part-of-speech)
• Model structure
• Develop other types of processing unit in RNN (GRU)
• Study context
• Explore individual’s motivations of other behaviors
• Consumer retentions, technology adoption, crowd funding project investment
• Predicting medication nonadherence
• Solutions to address medication nonadherence

Thank You

Xiao Liu Xiao.liu@eccles.Utah.edu Department of Operations and Information Systems David Eccles School of Business

44

Acknowledgment

• This study is supported by
• National Institutes of Health (grant #1R01DA037378-01)
• National Science Foundation (grant #IIS-1553109 and IIS-1552860)

45

References

• Abbasi A, Chen H. CyberGate: a design framework and system for text analysis of computer-mediated communication. Mis Quarterly. 2008 Dec 1:811-37.
• Adler NE, Stead WW. Patients in context—EHR capture of social and behavioral determinants of health. New England Journal of Medicine. 2015 Feb 19;372(8):698-701.
• Ajzen I. The Theory of Planned Behavior Prganizational Behavior and Human Decision Processes.
• Amunategui M, Markwell T, Rozenfeld Y. Prediction Using Note Text: Synthetic Feature Creation with word2vec. arXiv preprint arXiv:1503.05123. 2015 Mar 17.
• Arazy O, Daxenberger J, Lifshitz-Assaf H, Nov O, Gurevych I. Turbulent stability of emergent roles: The dualistic nature of self-organizing knowledge coproduction. Information Systems Research. 2016 Sep 16;27(4):792-812.
• Bailey JE, Hajjar M, Shoib B, Tang J, Ray MM, Wan JY. Risk factors associated with antihypertensive medication nonadherence in a statewide Medicaid population. The American journal of the medical sciences. 2014 Nov 30;348(5):410-5.
• Bapna R, Ramaprasad J, Shmueli G, Umyarov A. One-way mirrors in online dating: A randomized field experiment. Management Science. 2016 Feb 2;62(11):3100-22.
• Barsalou LW. Abstraction as dynamic interpretation in perceptual symbol systems. Building object categories. 2005 May 6;30322:389-431.
• Barsalou LW. Abstraction in perceptual symbol systems. Philosophical Transactions of the Royal Society of London B: Biological Sciences. 2003 Jul 29;358(1435):1177-87.
• Björne J, Kaewphan S, Salakoski T. UTurku: drug named entity recognition and drug-drug interaction extraction using SVM classification and domain knowledge. In Proceedings of the 7th International Workshop on Semantic Evaluation (SemEval 2013) 2013 Jun 14 (pp. 651-659).
• Briesacher BA, Andrade SE, Fouayzi H, Chan KA. Comparison of drug adherence rates among patients with seven different medical conditions. Pharmacotherapy: The Journal of Human Pharmacology and Drug Therapy. 2008 Apr 1;28(4):437-43.
• Brisson L, Torrel JC. Opinion mining on experience feedback: A case study on smartphones reviews. In Research Challenges in Information Science (RCIS), 2015 IEEE 9th International Conference on 2015 May 13 (pp. 187-192). IEEE.
• Bundschus M, Dejori M, Stetter M, Tresp V, Kriegel HP. Extraction of semantic biomedical relations from text using conditional random fields. BMC bioinformatics. 2008 Apr 23;9(1):1.
• Cao P, Zaiane O, Zhao D. A measure optimized cost-sensitive learning framework for imbalanced data classification. InBiologically-Inspired Techniques for Knowledge Discovery and Data Mining 2014 May 31 (pp. 48-75). IGI Global.
• Chalapathy R, Borzeshi EZ, Piccardi M. Bidirectional LSTM-CRF for Clinical Concept Extraction. arXiv preprint arXiv:1611.08373. 2016 Nov 25.
• Chan W, Lane I. Deep Recurrent Neural Networks for Acoustic Modelling. arXiv preprint arXiv:1504.01482. 2015 Apr 7.
• Checchi KD, Huybrechts KF, Avorn J, Kesselheim AS. Electronic medication packaging devices and medication adherence: a systematic review. Jama. 2014 Sep 24;312(12):1237-47.
• Chen H, Chiang RH, Storey VC. Business intelligence and analytics: From big data to big impact. MIS quarterly. 2012 Dec 1;36(4):1165-88.
• Cheng J, Zhang X, Li P, Zhang S, Ding Z, Wang H. Exploring sentiment parsing of microblogging texts for opinion polling on chinese public figures. Applied Intelligence. 2016 Sep 1;45(2):429-42.
• Chiaramello E, Pinciroli F, Bonalumi A, Caroli A, Tognola G. Use of “off-the-shelf” information extraction algorithms in clinical informatics: A feasibility study of MetaMap annotation of Italian medical notes. Journal of Biomedical Informatics. 2016 Oct 31;63:22-32.
• Choi E, Bahadori MT, Sun J. Doctor ai: Predicting clinical events via recurrent neural networks. arXiv preprint arXiv:1511.05942. 2015 Nov 18.
• Chung J, Gulcehre C, Cho K, Bengio Y. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555. 2014 Dec 11.
• Dempe C, Jünger J, Hoppe S, Katzenberger ML, Möltner A, Ladwig KH, Herzog W, Schultz JH. Association of anxious and depressive symptoms with medication nonadherence in patients with stable coronary artery disease. Journal of psychosomatic research. 2013 Feb 28;74(2):122-7.
• Denecke K, Deng Y. Sentiment analysis in medical settings: New opportunities and challenges. Artificial intelligence in medicine. 2015 May 31;64(1):17-27.
• Dima C, Hinrichs E. Automatic Noun Compound Interpretation using Deep Neural Networks and Word Embeddings. IWCS 2015. 2015 Apr 15:173.
• Divita G, Zeng QT, Gundlapalli AV, Duvall S, Nebeker J, Samore MH. Sophia: a expedient UMLS concept extraction annotator. In AMIA Annual Symposium Proceedings 2014 (Vol. 2014, p. 467). American Medical Informatics Association.
• Du Pan, Sophie Martin, et al. "Differential drug retention between anti-TNF agents and alternative biological agents after inadequate response to an anti-TNF agent in rheumatoid arthritis patients." Annals of the rheumatic diseases (2012): annrheumdis-2011.
• eyapriya A, Selvi CK. Extracting aspects and mining opinions in product reviews using supervised learning algorithm. In Electronics and Communication Systems (ICECS), 2015 2nd International Conference on 2015 Feb 26 (pp. 548-552). IEEE.
• Fan Y, Qian Y, Xie FL, Soong FK. TTS synthesis with bidirectional LSTM based recurrent neural networks. In Interspeech 2014 Sep (pp. 1964-1968).
• Fang Q, Xu C, Sang J, Hossain MS, Muhammad G. Word-of-mouth understanding: Entity-centric multimodal aspect-opinion mining in social media. IEEE Transactions on Multimedia. 2015 Dec;17(12):2281-96.
• Fox K, Ford I, Steg PG, Tardif JC, Tendera M, Ferrari R. Ivabradine in stable coronary artery disease without clinical heart failure. New England Journal of Medicine. 2014 Sep 18;371(12):1091-9.
• Frost J, Vermeulen IE, Beekers N. Anonymity versus privacy: selective information sharing in online cancer communities. Journal of medical Internet research. 2014;16(5):e126.

46

References

• Hugtenburg JG, Timmers L, Elders PJ, Vervloet M, van Dijk L. Definitions, variants, and causes of nonadherence with medication: a challenge for tailored interventions. Patient Prefer Adherence. 2013 Jul 1;7:675-82.
• Huh J, Marmor R, Jiang X. Lessons learned for online health community moderator roles: a mixed-methods study of moderators resigning from WebMD communities. Journal of Medical Internet Research. 2016 Sep;18(9).
• Jiang Z, Li L, Huang D, Jin L. Training word embeddings for deep learning in biomedical text mining tasks. InBioinformatics and Biomedicine (BIBM), 2015 IEEE International Conference on 2015 Nov 9 (pp. 625-628). IEEE.
• Keenan TF, Hollinger DY, Bohrer G, Dragoni D, Munger JW, Schmid HP, Richardson AD. Increase in forest water-use efficiency as atmospheric carbon dioxide concentrations rise. Nature. 2013 Jul 18;499(7458):324-7.
• Knight E, Werstine RJ, Rasmussen-Pennington DM, Fitzsimmons D, Petrella RJ. Physical therapy 2.0: leveraging social media to engage patients in rehabilitation and health promotion. Physical therapy. 2015 Mar 1;95(3):389.
• Kuhlmann M. Mildly non-projective dependency grammar. Computational Linguistics. 2013 Jun;39(2):355-87.
• Lakkaraju H, Socher R, Manning C. Aspect specific sentiment analysis using hierarchical deep learning. In NIPS Workshop on Deep Learning and Representation Learning 2014.
• Lee KJ, Hwang YS, Kim S, Rim HC. Biomedical named entity recognition using two-phase model based on SVMs. Journal of Biomedical Informatics. 2004 Dec 31;37(6):436-47.
• Lee YY, Ke H, Huang HH, Chen HH. Combining word embedding and lexical database for semantic relatedness measurement. InProceedings of the 25th International Conference Companion on World Wide Web 2016 Apr 11 (pp. 73-74). International World Wide Web Conferences Steering Committee.
• Levy O, Goldberg Y. Neural word embedding as implicit matrix factorization. In Advances in neural information processing systems 2014 (pp. 2177-2185).
• Li H, Mukherjee A, Si J, Liu B. Extracting Verb Expressions Implying Negative Opinions. In AAAI 2015 Jan 25 (pp. 2411-2417).
• Li Q, Shah S, Fang R, Nourbakhsh A, Liu X. Discovering relevant hashtags for health concepts: A case study of twitter. In Workshops at the Thirtieth AAAI Conference on Artificial Intelligence 2016 Mar 29.
• Liang D, Zhang Y. AC-BLSTM: Asymmetric Convolutional Bidirectional LSTM Networks for Text Classification. arXiv preprint arXiv:1611.01884. 2016 Nov 7.
• Lipton ZC, Kale DC, Elkan C, Wetzell R. Learning to diagnose with LSTM recurrent neural networks. arXiv preprint arXiv:1511.03677. 2015 Nov 11.
• Liu Q, Liu B, Zhang Y, Kim DS, Gao Z. Improving opinion aspect extraction using semantic similarity and aspect associations. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence 2016 Feb 12 (pp. 2986-2992). AAAI Press.
• Liu Y, Ge T, Mathews KS, Ji H, McGuinness DL. Exploiting task-oriented resources to learn word embeddings for clinical abbreviation expansion. In Proceedings of the 2015 Workshop on Biomedical Natural Language Processing 2015 Jul 30 (pp. 92-97).
• Lowry PB, Moody GD, Galletta DF, Vance A. The drivers in the use of online whistle-blowing reporting systems. Journal of Management Information Systems. 2013 Jul 1;30(1):153-90.
• Lowry PB, Zhang J, Wang C, Siponen M. Why do adults engage in cyberbullying on social media? An integration of online disinhibition and deindividuation effects with the social structure and social learning model. Information Systems Research. 2016 Nov 18.
• Ma X, Hovy E. End-to-end Sequence Labeling via Bi-directional LSTM-CNNs-CRF. arXiv preprint arXiv:1603.01354. 2016 Mar 4.
• Magnussen H, Disse B, Rodriguez-Roisin R, Kirsten A, Watz H, Tetzlaff K, Towse L, Finnigan H, Dahl R, Decramer M, Chanez P. Withdrawal of inhaled glucocorticoids and exacerbations of COPD. New England Journal of Medicine. 2014 Oct 2;371(14):1285-94.
• Mago R, Borra D, Mahajan R. Role of adverse effects in medication nonadherence in bipolar disorder. Harvard review of psychiatry. 2014 Nov 1;22(6):363-6.
• Mak K, Lo W, Yeung W, Wong M, Chung D, Chui E, Tam K, Mui J, Chan O, Wong K. Adherence issues to antipsychotic medication in schizophrenia: results from a Hong Kong survey targeted for psychiatrists and case managers. International Journal of Psychosocial Rehabilitation. Vol 20 (1) 88. 2016;104.
• Marcum ZA, Sevick MA, Handler SM. Medication nonadherence: a diagnosable and treatable medical condition. JAMA. 2013 May 22;309(20):2105-6.
• Marcum, Zachary A., et al. "Prevalence and correlates of self-reported medication non-adherence among older adults with coronary heart disease, diabetes mellitus, and/or hypertension." Research in Social and Administrative Pharmacy 9.6 (2013): 817-827.
• Marrese-Taylor E, Velásquez JD, Bravo-Marquez F. A novel deterministic approach for aspect-based opinion mining in tourism products reviews. Expert Systems with Applications. 2014 Dec 1;41(17):7764-75.
• Meducation A. Improving medication adherence in older adults. American Society on Aging and. 2006.
• Mert DG, Turgut NH, Kelleci M, Semiz M. Perspectives on reasons of medication nonadherence in psychiatric patients. Patient preference and adherence. 2015;9:87.
• Mikolov, Tomas, et al. "Distributed representations of words and phrases and their compositionality." Advances in neural information processing systems. 2013.
• Miller. 2013. http://www.psychiatrictimes.com/bipolar-disorder/discontinuing-medications-when-why-and-how. Accessed Nov 28, 2016.
• Miyamoto SW, Henderson S, Young HM, Pande A, Han JJ. Tracking health data is not enough: A qualitative exploration of the role of healthcare partnerships and mhealth technology to promote physical activity and to sustain behavior change. JMIR mHealth and uHealth. 2016 Jan;4(1).
• Nie L, Li T, Akbari M, Shen J, Chua TS. Wenzher: Comprehensive vertical search for healthcare domain. In Proceedings of the 37th international ACM SIGIR conference on Research & development in information retrieval 2014 Jul 3 (pp. 1245-1246). ACM.
• Nikfarjam A, Sarker A, O’Connor K, Ginn R, Gonzalez G. Pharmacovigilance from social media: mining adverse drug reaction mentions using sequence labeling with word embedding cluster features. Journal of the American Medical Informatics Association. 2015 Mar 8:ocu041.

47

References

• Quang D, Xie X. DanQ: a hybrid convolutional and recurrent deep neural network for quantifying the function of DNA sequences. Nucleic acids research. 2016 Jun 20;44(11):e107-.
• Rao K, Peng F, Sak H, Beaufays F. Grapheme-to-phoneme conversion using long short-term memory recurrent neural networks. In2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2015 Apr 19 (pp. 4225-4229). IEEE.
• Rietveld T, van Hout R. The t test and beyond: Recommendations for testing the central tendencies of two independent samples in research on speech, language and hearing pathology. Journal of communication disorders. 2015 Dec 31;58:158-68.
• Rocktäschel T, Weidlich M, Leser U. ChemSpot: a hybrid system for chemical named entity recognition. Bioinformatics. 2012 Jun 15;28(12):1633-40.
• Rubenstein H, Goodenough JB. Contextual correlates of synonymy. Communications of the ACM 1965;8(10):627-33.
• Saboo AR, Kumar V, Park I. Using Big Data to Model Time-Varying Effects for Marketing Resource (Re) Allocation. MIS Quarterly. 2016 Dec 1;40(4):911-39.
• Saha S, Ekbal A. Combining multiple classifiers using vote based classifier ensemble technique for named entity recognition. Data & Knowledge Engineering. 2013 May 31;85:15-39.
• Samha AK, Li Y, Zhang J. Aspect-based opinion mining from product reviews using conditional random fields. In Data Mining and Analytics: Proceedings of the 13th Australasian Data Mining Conference [Conferences in Research and Practice in Information Technology, Volume 168] 2015 Aug 8 (pp. 119-128). Australian

Computer Society.

• Segura-Bedmar I, Suárez-Paniagua V, Martınez P. Exploring word embedding for drug name recognition. In SIXTH INTERNATIONAL WORKSHOP ON HEALTH TEXT MINING AND INFORMATION ANALYSIS (LOUHI) 2015 Sep 17 (p. 64).
• Serre T, Oliva A, Poggio T. A feedforward architecture accounts for rapid categorization. Proceedings of the National Academy of Sciences. 2007 Apr 10;104(15):6424-9.
• Servick K. ‘Nonadherence’: A bitter pill for drug trials. Science. 2014 Oct 17;346(6207):288-9.
• Shrank WH. Our Bulging Medicine Cabinets--The Other Side of Medication Nonadherence. The New England journal of medicine. 2011 Apr 28;364(17):1591.
• Suakkaphong N, Zhang Z, Chen H. Disease named entity recognition using semisupervised learning and conditional random fields. Journal of the American Society for Information Science and Technology. 2011 Apr 1;62(4):727-37.
• Sutskever, Ilya, Oriol Vinyals, and Quoc V. Le. "Sequence to sequence learning with neural networks." Advances in neural information processing systems. 2014.
• Tai KS, Socher R, Manning CD. Improved semantic representations from tree-structured long short-term memory networks. arXiv preprint arXiv:1503.00075 2015.
• Tang D, Wei F, Yang N, Zhou M, Liu T, Qin B. Learning Sentiment-Specific Word Embedding for Twitter Sentiment Classification. InACL (1) 2014 Jun (pp. 1555-1565).
• Traverso G, Langer R. Perspective: Special delivery for the gut. Nature. 2015 Mar 26;519(7544):S19-.
• Van Vleck T, Wilcox AB, Stetson PD, Johnson SB, Elhadad N. Content and structure of clinical problem lists: a corpus analysis. In AMIA 2008 Nov 6.
• Viehman, J. Alexander, et al. "Adverse Events Lead to Drug Discontinuation More Commonly among Patients Who Receive Nafcillin than among Those Who Receive Oxacillin." Antimicrobial agents and chemotherapy 60.5 (2016): 3090-3095.
• Wang P, Xu B, Xu J, Tian G, Liu CL, Hao H. Semantic expansion using word embedding clustering and convolutional neural network for improving short text classification. Neurocomputing. 2016 Jan 22;174:806-14.
• Wang R, Huang W, Chen W, Wang T, Lei K. ASEM: mining aspects and sentiment of events from microblog. In Proceedings of the 24th ACM International on Conference on Information and Knowledge Management 2015 Oct 17 (pp. 1923-1926). ACM.
• Wang S, Chen Z, Liu B. Mining aspect-specific opinion using a holistic lifelong topic model. In Proceedings of the 25th International Conference on World Wide Web 2016 Apr 11 (pp. 167-176). International World Wide Web Conferences Steering Committee.
• Wang W, Tan G, Wang H. Cross-domain comparison of algorithm performance in extracting aspect-based opinions from Chinese online reviews. International Journal of Machine Learning and Cybernetics. 2016:1-8.
• Wang W, Pan SJ, Dahlmeier D, Xiao X. Recursive Neural Conditional Random Fields for Aspect-based Sentiment Analysis. arXiv preprint arXiv:1603.06679. 2016 Mar 22.
• Wang, Peilu, et al. "Part-of-Speech Tagging with Bidirectional Long Short-Term Memory Recurrent Neural Network." arXiv preprint arXiv:1510.06168 (2015).
• Weidenbacher HJ, Beadles CA, Maciejewski ML, Reeve BB, Voils CI. Extent and reasons for nonadherence to antihypertensive, cholesterol, and diabetes medications: the association with depressive symptom burden in a sample of American veterans. Patient preference and adherence. 2015;9:327.
• Wöllmer M, Metallinou A, Katsamanis N, Schuller B, Narayanan S. Analyzing the memory of BLSTM neural networks for enhanced emotion classification in dyadic spoken interactions. In2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2012 Mar 25 (pp. 4157-4160). IEEE.
• Woo CW, Koban L, Kross E, Lindquist MA, Banich MT, Ruzic L, Andrews-Hanna JR, Wager TD. Separate neural representations for physical pain and social rejection. Nature communications. 2014 Nov 17;5.
• Wu Y, Denny JC, Rosenbloom S, Miller RA, Giuse DA, Xu H. A comparative study of current Clinical Natural Language Processing systems on handling abbreviations in discharge summaries. In AMIA 2012 Nov.
• Wu Y, Xu J, Zhang Y, Xu H. Clinical abbreviation disambiguation using neural word embeddings. ACL-IJCNLP 2015. 2015 Jul 30:171.
• Xiong S, Ji D. Exploiting flexible-constrained K-means clustering with word embedding for aspect-phrase grouping. Information Sciences. 2016 Nov 1;367:689-99.
• Xu Y, Armony M, Ghose A. The effect of online reviews on physician demand: A structural model of patient choice.

48

Appendix – Medication Nonadherence Indication Identification

Precision Recall F1-score F2-score Posts indicating nonadherence 84.16% 81.58% 82.85% 82.08% Posts with no nonadherence indication 82.13% 84.65% 83.37% 84.13% Average 83.15% 83.11% 83.13% 82.12%

20,977 threads (128,003 sentences) indicating medication nonadherence

8,000 posts to train BLSTM 1,000 posts to test BLSTM Table A1. Performance of BLSTM text classification model

49

Appendix – Supplementary Analyses

• Identified 24,832 medication nonadherence reason expressions (10 types)
• Whether nonadherence reasons are complexed by different classes of drugs?
• What nonadherence reasons are the most commonly reported?

50

Appendix – Supplementary Analyses

Whether nonadherence reasons are complexed by different classes of drugs? (Figure A1- A3 in appendix)

• More diabetes patients (12.66%) quit because of overdose risk
• Need to take constantly à hypoglycemia
• More mental disease patients (16.10%) quit because of negative emotions
• Unstable mental and emotional status
• More diabetes patients (10.59%) switch to other drugs
• Many alternative diabetes medicines à find the most suitable medicine

51

Appendix – Supplementary Analyses

• Many patients with

diabetes are concerned with

• verdose risks
• Need to take

constantly à hypoglycemia

Figure A1. Medication nonadherence reasons by drug class

Drug class

52

Appendix – Supplementary Analyses

Figure A2. Medication nonadherence reasons by drug class

Drug class

• Many patients with

mental diseases have negative emotions, so they discontinued medication

• Unstable mental

status

53

Appendix – Supplementary Analyses

Figure A3. Medication nonadherence reasons by drug class

Drug class

• Many patients with

diabetes switch to

• ther medications
• Many

alternative diabetes medicines à find the most suitable medicine

54

Appendix – Supplementary Analyses

What nonadherence reasons are the most commonly reported?

• Logistic regression: ! "#$%&'($ℎ#*#'+# ,#(-&'- =

/ /012 ∑456

67 87984:;<=>?4

• Randomly sampled 400 posts from corpus (' ≥ (

BC D )F à enough statistical power)

• 121 indicating nonadherence, 279 do not indicate nonadherence
• Independent variables: the count for each type of reasons in a post

Estimate

• Std. Error

z value Pr(>|z|) (Intercept)

• 2.91571

0.31632

• 9.218

< 2e-16 *** Adverse event 0.25013 0.05794 4.317 1.58e-05 *** Caregivers’ suggestion

• 0.88093

0.30628

• 2.876

0.00402** Negative emotion 0.56459 0.18337 3.079 0.00208** Ineffective 19.045 2160.67511 0.009 0.99297 Insurance does not cover

• 0.28797

1.20389

• 0.239

0.81095 Overdose risk 0.03004 0.15625 0.192 0.84752 Patient's choice to quit 19.44369 1525.99804 0.013 0.98983 Price is too high

• 0.42732

0.83321

• 0.513

0.60805 Switch to other drugs 1.52494 0.32967 4.626 3.73e-06 ***

Table A2. Logistic regression results

Self-perception is strong: adverse event, emotion

55

Appendix – Differed Reasons by Drug Class

• 180 unique drugs à nonadherence reasons differ

Drug class Overdose risk High price Advers e event Negative emotion Caregivers’ influence Ineffective Inconvenience No insurance Drug switch Pregnant Allergic 8.65% 1.05% 58.09% 14.02% 1.65% 2.23% 3.08% 0.98% 10.15% 0.09% Arthritis 9.54% 0.52% 59.75% 14.46% 2.36% 1.63% 3.35% 0.17% 8.21% 0.02% Cancer 10.43% 0.39% 62.16% 13.69% 2.40% 1.04% 3.01% 0.14% 6.68% 0.06% Diabetes 12.66% 0.41% 59.25% 11.78% 2.22% 0.81% 2.18% 0.08% 10.59% 0.02% Heart 10.37% 0.84% 62.58% 13.17% 2.21% 1.05% 3.05% 0.15% 6.55% 0.04% High blood pressure 9.45% 1.06% 64.24% 11.13% 1.87% 0.99% 2.33% 0.50% 8.41% 0.01% Infectious 8.48% 0.39% 64.76% 13.29% 2.26% 1.11% 2.45% 0.05% 7.17% 0.04% Kidney 9.63% 0.30% 63.76% 12.70% 2.26% 1.28% 2.92% 0.10% 7.00% 0.05% Mental health 10.79% 0.52% 57.83% 16.10% 2.42% 1.62% 3.25% 0.16% 7.26% 0.05% Respiratory 10.77% 0.71% 60.65% 13.99% 2.52% 1.07% 2.88% 0.19% 7.18% 0.04% Seizure 10.04% 0.71% 60.32% 14.63% 2.50% 1.33% 2.98% 0.20% 7.25% 0.05%

Table 6. Medication nonadherence reasons by drug class

56

Appendix – Error Analysis

• Sources of errors in the test set

(13,765 words in test set)

• False positive (0.81%)
• False negative (1.58%)

Figure 9. Sources of false positive and false negative