Scholars’ Mine
Scholars’ Mine
Masters Theses
Student Theses and Dissertations
Spring 2015
Online diagnosis of diabetes with Twitter data
Online diagnosis of diabetes with Twitter data
Farheen Ali
Follow this and additional works at: https://scholarsmine.mst.edu/masters_theses
Part of the Computer Sciences Commons
Department:
Department:
Recommended Citation
Recommended Citation
Ali, Farheen, “Online diagnosis of diabetes with Twitter data” (2015). Masters Theses. 7383.
https://scholarsmine.mst.edu/masters_theses/7383
This thesis is brought to you by Scholars’ Mine, a service of the Missouri S&T Library and Learning Resources. This
work is protected by U. S. Copyright Law. Unauthorized use including reproduction for redistribution requires the
permission of the copyright holder. For more information, please contact scholarsmine@mst.edu.
ONLINE DIAGNOSIS OF DIABETES WITH TWITTER DATA
by
FARHEEN ALI
A THESIS
Presented to the Faculty of the Graduate School of the
MISSOURI UNIVERSITY OF SCIENCE AND TECHNOLOGY
In Partial Fulfillment of the Requirements for the Degree
MASTER OF SCIENCE IN INFORMATION SCIENCE
AND TECHNOLOGY
2015
Approved by
Dr. Fiona Fui-Hoon Nah, Advisor
Dr. Sriram Chellappan, Co-Advisor
Dr. Keng L. Siau
Dr. Michael Gene Hilgers
Copyright 2015
Farheen Ali
All Rights Reserved
iii
ABSTRACT
Innovation in technology enables people to communicate, share information and
look for their needs by just sitting in rooms and going through some clicks. While social
media has played a very important role in connecting people worldwide, its potential has
stretched beyond the innovative idea of connecting people through their social networks.
While many thought there was no meeting point for the healthcare sector and social
media, it was a surprise when research and innovations have shown that social media
could lay a very significant role in the health care sector.
Research has been done in developing models that could use social media as the
data source for tracking diseases. Most of these analyses are based on models that
prioritize strong correlations with seasonal and pandemic kinds of diseases over the
health conditions of a specific individual user.
The aim of this research is to develop a diabetes detecting tool at the individual
level using a sample of Twitter IDs that have been collected from the Twitter search
using the query –‘recently diagnosed’ and ‘diabetes’. Based on text analysis of social
media posts using Fisher’s exact test, without any medical settings, this thesis
investigates the feasibility of diagnosing and classifying diabetes via machine learning
techniques, Naive Bayes and Random Forest classifiers. It was found that more than half
(20/30 ≈ 67%) of the users in the sample mentioned being tested positive for diabetes,
about 27% (8/30) of the users mentioned the symptoms and got involved in diabetes
related discussions, but did not mention about being tested positive and rest 4% had no
mention of symptoms or diabetes.
iv
ACKNOWLEDGMENT
My first thanks and heartfelt gratitude goes to Dr. Fiona Fui-Hoon Nah, my
advisor, for giving me the freedom to pursue my own interests and for trusting me on the
same. I could not have completed this thesis without her valuable suggestions and those
brainstorming meetings, where she taught me how to assess a problem and find a best
possible solution to it. I would like to thank her for being so patient with me, and helping
and guiding me to improve this thesis and in bringing it to this shape.
I would also like to thank my co-advisor, Dr. Sriram Chellappan, for introducing
me to the concept of Health Diagnosis via Social Network, offering me his invaluable
assistance despite his busy schedule, and for discussing with me his innovative ideas.
Without his motivation and support, I wouldn’t have been able to learn about this topic
and get a deeper understanding. I would also like to thank Dr. Keng L. Siau and Dr.
Michael Gene Hilgers for being part of my thesis committee and taking time to review
this work. This thesis would not be possible without the generous help of Raja Ashok
Bolla, who helped me by providing the tweets from the filtered Twitter IDs.
I saved the last for people closest to my heart – my family. I’m very thankful to
my parents, Dr. Mir Firman Ali and Shahnavaj Begum, and my siblings, Dr. Syed Irfan
Ali and Dr. Nasreen Ali, for helping me understand the medical terms and concepts
related to diabetes and for being patient with me while I dragged I.T. to medical science
and questioned a few traditional concepts. I specially wish to acknowledge Dr. Sekh
Ansar Alli, my brother-in-law, for encouraging me to pursue a master’s degree. If it
weren’t for him, I would have missed out on this amazing experience.
v
TABLE OF CONTENTS
Page
ABSTRACT
……………………………………………………………………………………………………….. iii
ACKNOWLEDGMENT………………………………………………………………………………………. iv
LIST OF ILLUSTRATIONS
……………………………………………………………………………….. vii
LIST OF TABLES
…………………………………………………………………………………………….. viii
SECTIONS
1. INTRODUCTION
……………………………………………………………………………… 1
1.1. PROBLEM DESCRIPTION ……………………………………………………….. 1
1.2. SOCIAL MEDIA AND HEALTHCARE: AN OVERVIEW ………….. 3
1.3. RESEARCH QUESTION AND MAJOR CONTRIBUTIONS
………… 5
1.4. THESIS ORGANIZATION…………………………………………………………. 7
2. RESEARCH METHODOLOGY …………………………………………………………. 9
2.1. FISHER’S EXACT TEST
…………………………………………………………… 9
2.2. NAIVE BAYES CLASSIFIER ………………………………………………….. 10
2.3. RANDOM FOREST ………………………………………………………………… 11
2.4. RESEARCH APPROACH
………………………………………………………… 13
3. TWITTER DATA PROCESSING
……………………………………………………… 14
3.1. COLLECTION OF TWEETS ……………………………………………………. 14
3.2. CLEANING AND PARSING DATA
…………………………………………. 17
3.3. CONDUCTING STATISTICAL ANALYSIS …………………………….. 17
vi
4. MACHINE LEARNING TECHNIQUE AND RESULTS …………………….. 21
4.1. NAIVE BAYES CLASSIFIER ………………………………………………….. 21
4.2. RANDOM FOREST METHOD OF CLASSIFICATION
……………… 25
5. CONCLUSION ……………………………………………………………………………….. 27
6. FUTURE WORK …………………………………………………………………………….. 28
APPENDICES
A. JAVA CODE TO COUNT THE WORDS
………………………………………….. 29
B. RAW DATA USED FOR THE FISHER’S EXACT TEST
…………………… 34
C. MATLAB CODE USED FOR RANDOM FOREST CLASSIFICATION 37
D. JAVA CODE TO GET USER STATUS…………………………………………….. 41
BIBLIOGRAPHY
………………………………………………………………………………………………. 49
VITA ……………………………………………………………………………………………………………….. 54
vii
LIST OF ILLUSTRATIONS
Figure
Page
3.1. Fisher’s Exact Test On Diabetes & Sleep
………………………………………………………… 18
3.2. Fisher’s Exact Test On Diabetes & Water
……………………………………………………….. 19
3.3. Fisher’s Exact Test On Diabetes & Rash
…………………………………………………………. 19
3.4. Fisher’s Exact Test On Diabetes & Tired ………………………………………………………… 20
4.1. Out-of-bag v/s Number Of Trees Grown Plot ………………………………………………….. 26
viii
LIST OF TABLES
Table
Page
3.1. Sample Tweets Collected
………………………………………………………………………………. 15
4.1. Training Data For Naive Bayes Classifier ……………………………………………………….. 21
4.2. Probability Table From Training Data
…………………………………………………………….. 23
1
1. INTRODUCTION
This section begins by stating the problem description and motivation for
conducting this research. This is followed by the main research question and a very brief
outline of the proposed research approach. The section closes with an outline of this
thesis along with the major research contributions.
1.1 PROBLEM DESCRIPTION
A human body consumes energy to perform different daily tasks. The source of
this energy is the food that is consumed. An organ called the pancreas, in a human body,
lying near the stomach, produces a hormone called insulin, which helps glucose to reach
all the cells of a human body. Diabetes is a metabolic disease, in which either the body
fails to make sufficient insulin or cannot utilize the insulin the way it should, which in
return causes sugar to build up in the body. Diabetes, if not controlled, causes
complications and effects heart, nerves, eyes, feet and kidneys [1].
The early common symptoms of diabetes include [2]:
Frequent urination
Feeling very thirsty
Frequently feeling hungry
Extreme fatigue
Blurry vision
Cuts/bruises that are slow to heal
Weight loss – even though a person eats more (type 1)
Tingling, pain, or numbness in the hands/feet (type 2)
2
According to statistics, approximately 366 million people suffer from diabetes
and it is estimated that by the end of 2030, the count will rise to 552 million. It is a
known fact that an early diagnosis of diabetes can help prevent progression to later
complications. Research states that about 183 million people presently have diabetes and
are unaware of it. One of the types of diabetes, type 2, can be evident in people for about
9 – 12 years without their knowledge and can cause complications during treatment.
Early detection of diabetes is crucial for active management for people who have
been newly diagnosed and have not developed complications yet [3]. It is unlikely to
expect everybody to be aware of the early symptoms of diabetes and visit a doctor.
However, in today’s world, according to “Worldwide Social Network Users: 2013
Forecast and Comparative Estimates”, approximately one in four people across the globe
use social networks, and this number is believed to have risen from 1.47 billion in 2012
to 1.73 billion in 2014, with an estimated 18% increase [4].
Twitter is one of the most famous online social networking services, with an
estimated 310,000,000 monthly visitors and 500 million users worldwide [5]. Within a
character limit of 140, it allows its users to post their thoughts and opinions, and gives its
registered users the privilege to read and comment [6]. Twitter provides its users a
platform to converse on almost every topic known to man, and thus, people started
discussing their health intentionally or unintentionally as well. This has intrigued many
researchers to look into the most common diseases people discuss and the scope of the
possible diagnosis virtually.
3
This study, hence, focuses on a potential system which can help a healthcare
professional to track his/her patients’ Twitter posts and diagnose diabetes in accordance
with the symptoms they post, with the help of social network analysis and text analysis.
1.2 SOCIAL MEDIA AND HEALTHCARE: AN OVERVIEW
Social media is a group of internet-based applications developed using Web 2.0
technology that offer opportunities for users to generate, share, receive, and comment on
social content among multiuser through multisensory communication [7, 8, 9, 10, 11].
Research has shown that there is a relationship between personality traits and engagement
with social media [12].
Social media brings a novel dimension to health care, as it proposes a medium to
be employed by the public, patients, and health professionals to communicate about
health issues with the possibility of potentially improving health outcomes. Social media
is changing the nature and speed of health care interaction between individuals and health
organizations. The general public, patients, and health professionals are using social
media to communicate about health issues [13] including health promotion and health
education [14, 15, 16, 17]. Social media has widened access and increased awareness
among those who may not easily access health information via traditional methods, such
as younger people, ethnic minorities, and lower socioeconomic groups [18, 19, 20].
Colineau and Paris [21] from their research have reported that people prefer using health-
related social networking sites to discuss sensitive issues and complex information with
health professionals. One of the advantages of social media is that it lets the health
4
information reach its audience via various other modes than just text; for example, videos
can be used to replace text and can be useful when literacy is low [22].The very famous
video sharing website YouTube allows users to share, upload and view videos online for
free, has been used by the general public to share and learn about medications,
symptoms, and diagnoses [21] and by patients to share personal cancer stories [23].
Social media adoption rates have shown variations in accordance to the
geographic locations; for example, in Europe the percentage of German hospitals using
social networks are in “single figures”, whereas approximately 45% of Norwegian and
Swedish hospitals are using LinkedIn, and 22% of Norwegian hospitals use Facebook for
health communication [24]. In the United States, on the other hand, 61% of adult search
online and 39% use social media such as Facebook for health information [25]. The
growth in popularity of social media among the general public has caused the research
and evolution of many applications within health contexts, ranging from the World
Health Organization using Twitter during the influenza A (H1N1) pandemic, with more
than 11,700 followers [26], to medical practices [27], and health professionals obtaining
information to inform their clinical practice [28, 29].
There is a range of social media platforms available currently that can facilitate a
dialogue between patients and health professionals [21, 30]. For example, sites such as
PatientsLikeMe enable patients to easily converse with others and share health
information and advice including information on treatment and medication [31, 32].
Famous social networking sites such as Twitter and Facebook are being used by the
general public, patients, and health professionals to share their experience of disease
management, exploration, and diagnosis [33]. Blog sites create a space where people can
5
access tailored resources [34] and provide health professionals with an opportunity to
share information with patients and members of the public [35, 21]. Asthma groups are
using MySpace to share health information, in particular personal stories and experiences
[36, 37].
Nowadays, social media is been used by many researchers to collect data on
patient experiences and opinions such as symptoms, physician’s performance etc. [34,
38]. With the help of these new modes of interactions, social media can monitor public
response to health issues [20], track and monitor disease outbreaks [39], identify target
areas for intervention efforts [40], and disseminate pertinent health information to
targeted communities [41]. Health professionals can aggregate data about patient
experiences from blogs and monitor the public reaction to health issues.
1.3 RESEARCH QUESTION AND MAJOR CONTRIBUTIONS
In the light of the reasons described in section 1, the main research question
addressed by this thesis is as follows:
Is it possible to observe diabetes based on text analysis of social media even if
the individual does not intentionally discuss his/her health?
While there is a lot of existing work on prediction of seasonal and pandemic
diseases, to the best of the author’s knowledge, an attempt to diagnose a non-seasonal as
well as a non-pandemic disease, like diabetes, based on an individual’s post on Twitter
has never been done before. Considering the fact that non-pandemic diseases are
6
extremely complex due to their similarity of early symptoms with many other diseases, it
becomes very difficult to possibly predict one without having actually met the patient in
person [42]. This thesis, therefore, presents an original work with the following as its
major contributions:
Approach to deal with the problem statement: One of the previous works done
on healthcare and social media includes the prediction of Influenza and Influenza-
like (ILI) activity in the USA, prior to the generation of Control and Prevention
(CDC) report and the source of the data has been Wikipedia usage and individual
based tweets.
Both of the above approaches demonstrate the diagnosis of seasonal and
pandemic diseases where the trends and estimation of the time period plays an
important role in the determination [43]. This research, on the other hand, focuses
on diseases which are not seasonal, yet the individuals who have been diagnosed
with these diseases have shown similar trends of mentioning the symptoms in their
tweets.
Approach to solve the problem statement: Twitter provides its API (Application
Programming Interface) to researchers and other web developers, and hence allows
a web platform to access and share information from one another.
My approach towards solving this problem involves, with the help of these
Twitter APIs, collecting all the individual posts, filtered from the IDs based on the
Twitter search query which involves keywords ‘recently diagnosed’ and ‘diabetes’
and then sorting them as cases and then comparing them to a sample of randomly
selected subset of people without the attribute (the controls) [44]. The entire
7
Twitter history of these individuals was extracted, and a model was developed to
count the number of times they posted the symptoms (such as sleep, water, eye,
rash, tired, etc.) related to diabetes [45], which being the early symptoms of
diabetes, the trends for these symptoms match the curve obtained from searching
the keyword ‘diabetes’ in Google trends. The more the number of keywords
mentioned in the posts, over a period of time, the greater the probability of a
person being possibly diabetic.
Diabetes, if known in the early stage, can help to take prior precautions and
keep the blood sugar level in control: A solution to this problem statement is
important since it would provide a patient more time, if detected earlier, to control
diabetes and prevent it from getting worse.
1.4 THESIS ORGANIZATION
Excluding this section, the remainder of the thesis is organized in the form of four
sections.
Section 2 describes the research methodology along with the necessary processes
and infrastructure used to obtain the outcome from the hypothesis. This section also
describes the research approach used, by dividing the entire thesis into three phases and
summarizing each of them.
Section 3 describes the process used to collect, clean and parse the relevant data.
This section closes by presenting a detailed account of various statistical insights
obtained by performing statistical analysis on the extracted data.
8
Section 4 is solely devoted to the main research task of building classification
classifiers. This section begins with sorting the training data for Naive Bayes classifier
and then creating the probability table. The later portions of this section describes how to
build a Random Forest method of classification. The section concludes by presenting
performance metrics such as the accuracy and precision for the proposed classification
models.
Finally, Section 5 concludes the thesis by summarizing all the four sections and
the results obtained. In addition, several potential approaches are described for improving
the classification accuracy, intended as a guide to future researchers who wish to extend
and build on this thesis.
9
2. RESEARCH METHODOLOGY
This section begins with the description of the types of analysis performed in this
study along with the detailed description of the necessary research infrastructure. This is
followed by a brief description of the proposed research methodology.
2.1 FISHER’S EXACT TEST
Named after its inventor, Sir R. A. Fisher, Fisher’s exact test is a statistical
significance test used for a 2 X 2 contingency table in order to compare the binomial
probabilities and to test for independence of 2 classifications. It is believed that Fisher’s
exact test helps to exactly calculate the deviations from a null hypothesis, independent of
the sample size or the sample characteristics, hence it falls under the class of exact test.
Although it is valid for all sample sizes, because of the above grounds, this test is
preferred when sample sizes are small [46]. The purpose of using the Fisher’s exact test
is to classify the categorical data in order to determine the significance of contingency
between them. The null hypothesis for Fisher’s exact test states that assuming each
observation is classified into exactly one cell and the rows and columns are fixed, the
comparative proportions of two variables are independent of each other. In simpler terms,
there is no affiliation between the rows and columns of a 2 X 2 contingency table, such
that the probability of a subject being in a particular row is not determined by being in a
particular column [47]. Provided the margins are fixed, the Fisher’s exact test when
applied to a table with cells a, b, c & d and the marginal totals (a + b), (c + d), (a + c) and
(b + d):
10
Where
is the binomial coefficient.
2.2 NAIVE BAYES CLASSIFIER
Descending from the family of simple probabilistic classifiers, Naive Bayes is a
popular method for text classification i.e. it judges the belonging of documents in their
respective categories (such as sports or politics, healthy or sick etc.) on the basis of word
frequencies as the features [48]. Based on the Bayesian theorem, this classifier assumes
the presence (or absence) of a particular feature of a class is unrelated to the presence (or
absence) of any other feature. For example, an orange is a fruit with distinctive features
of orange in color, round and about 4’ in diameter. Now irrespective of other features
present or the fact that these features may be dependent on each other, a Naive Bayes
classifier would consider all of these properties to independently contribute to the
probability that the given fruit is an orange. This type of classifier is henceforth useful in
medical diagnosis, since it would work very well with diseases showing similar
symptoms. Further, it is also capable of working well with a small amount of training
data to estimate the parameters (means and variances of the variables) necessary for
classification. Other advantages of using Naive Bayes classifier include its non-
sensitivity to irrelevant features, its capability to handle real, discrete and streaming data
11
and most importantly it is fast to train and classify. Naive Bayes classifier is particularly
suited when the dimensionality of the input is higher. Parameter estimation of this model
uses the method of maximum likelihood [49].
P(c|x) is the posterior probability of class (target) given predictor (attribute).
P(c) is the prior probability of class.
P(x|c) is the likelihood which is the probability of predictor given class.
P(x) is the prior probability of predictor. [14]
2.3 RANDOM FOREST
Random forest is an ensemble learning method developed by Leo Breiman and
Adele Cutler [50], which during its training period constructs a magnitude of decision
trees and outputs the resultant mode of the classes (classification) or mean prediction
(regression) of the individual trees. Dietterich first came up with the idea of randomized
node optimization, where instead of deterministic optimization, a randomized procedure
12
was used to select decisions from each node. Usually used for classification and
regression, this method is better than the decision trees since they do not over fit their
training data by providing too many parameters relative to the number of observations
[50]. Unlike the standard tree methods of classification, in the case of random forest
method, the best among a subset of predictions are chosen randomly to split each node
[51]. This method is considered to be more user-friendly since it has only two parameters
(the number of variables in the random subset at each node and the number of trees in the
forest), and is usually not very sensitive to their values. The concept behind this method
is growing a forest of trees and inducing the variation among the trees by projecting the
training data into a randomly chosen subspace before fitting with each other. It averages
multiple deep decision trees during this process with an intention to reduce the variance
and during this it also boosts the performance of the final model [52]. An estimate of the
error rate can be obtained based on the training data by calculating the out-of-bag error.
Out-of-bag error
While in the process of retaining the training set by sampling with replacement
for the current tree, about one-third of the cases are left out of the sample and not used in
the construction of the kth tree. This left out data is termed as out-of-bag data and is
utilized to get an unbiased estimation of the classification error as the trees are added to
the forest. Once the tree is built, the entire set of data is made to go through the tree and
in the meanwhile the proximities for each pair cases are computed. The proximity
increases by one if any of the two cases occupy the same terminal node. The proximities
are then normalized in the end by dividing with the number of trees. Now, let us consider
j is a class that scored more points every time case n was out of the bag. Thus, the
13
proportion of the time that j doesn’t match with the class of n averaged over all cases is
the out-of-bag error estimate [51, 52].
2.4 RESEARCH APPROACH
This thesis is accomplished by a three phase procedure.
Phase I – Data collection from social network analysis: A Twitter search query,
including keywords, “diagnosed, diabetes” is used to search for all the Twitter IDs
of people who posted about being diagnosed with diabetes. These IDs were then
filtered to find out only those prospects who posted about being diagnosed
recently. The resultant Twitter IDs were made to pass through a Java program,
which would process posts of each ID, one at a time and count the number of
times the symptoms of diabetes have been mentioned in the past.
Phase II – Data preprocessing and statistical analysis: Fisher’s exact test was
used to find the similarity between the symptoms, used as the keywords for
selecting the prospective Twitter profiles, and diabetes. The Twitter profiles were
narrowed down on the basis of these keywords used in the posts along with a
mention of being diagnosed with diabetes.
Phase III – Diabetes classification using Machine Learning Techniques:
Leveraging on the statistical insights from phase II, two machine learning
techniques (Random Forest classification method & Naive Bayes’ classifier) will
be employed to perform diabetes classification.
14
3. TWITTER DATA PROCESSING
This section provides a detailed explanation of the data collection from Twitter,
data processing and statistical analysis over the collected data. The analysis of the posts
on Twitter used to determine the symptoms of diabetes consists of the following steps:
1. Collection of tweets.
2. Cleaning and parsing of data.
3. Conducting statistical analysis of the extracted data.
3.1 COLLECTION OF TWEETS
Twitter welcomes developers to explore its platform for research purposes. An
unofficial Java library for the Twitter Application Program Interface (Twitter API),
provides application automation so that it can be integrated with Twitter. A few
healthcare professionals were contacted and consulted in order to have a better
understanding about diabetes and its early symptoms. Tweets were collected on the basis
of the symptoms suggested by the physicians. A keyword strategy was adopted for the
collection purposes. The following symptoms were derived from the discussions:
sleeping disorder, obesity, water loss in the body, susceptibility to heat, and the
redundant need of eating food. From these early symptoms, the following keywords
were then derived: sleep, weight, water, heat, hungry; and these keywords were looked
up in the Twitter search to find the IDs of people who have posted the same keywords in
the past along with the phrases ‘recently diagnosed’ and ‘diabetes’. Only those prospects
who have mentioned at least 2 or more symptoms in their Twitter posts (for accuracy
15
purpose, mentions of minimum two symptoms were looked up) along with the mention
of being recently diagnosed with diabetes, were taken into account and then passed
though the Java code to count the number of times these symptoms were posted and the
dates of their postings. A sample of tweets collected, where users have unintentionally
tweeted their symptoms and with time, eventually mentioned about diagnosed diabetic,
are shown in Table 3.1.
Table 3.1. Sample Tweets Collected
S.
No.
Date
Tweet
1.
Mon Apr 14 12:12:57 CDT 2014
Hurting, need more sleep
Mon Jul 21 01:58:04 CDT 2014
Finally ate something today, but I did drink
a lot of water, as always
Wed Jul 30 21:55:29 CDT 2014
I wilt in this heat
Sun Jul 20 17:55:37 CDT 2014
I am so hungry, having a 🙁 kinda day.
Fri Sep 12 00:57:28 CDT 2014
Kinda like my name:) This diabetes sight
has some great quotes. Recently found it,
and was recently diagnosed.
16
2.
Sun Mar 03 17:45:55 CST 2013
Sleep, who needs it sef. Back in the groove
Wed Jun 05 23:15:29 CDT2013
My weight loss journey started on Monday
Fri Aug 23 09:52:32 CDT 2013
Why y run? Couldn’t stand the heat
Wed Aug 13 00:06:35 CDT2014
Recently diagnosed with diabetes, this
story scares me
3.
Sat Aug 02 02:24:30 CDT 2014
Okay for real, someone please
pray/will/voodoo me to sleep.
Wed Sep 24 23:09:42 CDT 2014
So either I freeze to death tonight or I die of
heat exhaustion. Either way, I will not live
to see 8 am.
Mon Sep 15 11:18:20 CDT 2014
Love feeling so hungry and nauseous at the
same time.
Fri Oct 17 19:18:36 CDT 2014
I was recently diagnosed with diabetes so I
am trying to be good.
The results from the execution of the search query included both relevant and
irrelevant accounts (such as the accounts by diabetes community, diabetes association
and some by the nurses or diabetes physicians). The accounts which were irrelevant were
Table 3.1. Sample Tweets Collected (cont.)
