Analyzing Public Reactions during the MPox Outbreak: Findings from Topic Modeling of Tweets

1. Introduction

Monkeypox (MPox), caused by the monkeypox virus, which belongs to the Poxviridae family, Chordopoxvirinae subfamily, and Orthopoxvirus genus [1], is a re-emerging zoonotic disease. In the shape of a brick-like virion ranging from 200 nm to 250 nm, the MPox virus has a large genome of about 200 kilobase pairs encoding approximately 190 proteins [2]. Two clades of MPox, clade 1 and clade 2, show a 0.5% genomic difference, with clade 1 having a 1-12% case-fatality rate and clade 2 having a 0.1% case-fatality rate [3,4]. The first case of human MPox was recorded in a 9-month-old boy in the Democratic Republic of the Congo (DRC) in 1970 [1]. After the first case in 1970, 59 cases were reported in West and Central Africa in the next decade, with a 17% mortality rate in children under 10 [5,6,7,8,9]. The World Health Organization (WHO) monitored MPox cases post-1980. Between 1981 and 2017, there were multiple outbreaks of MPox in DRC due to clade 1, with the fatality rate being between 1-12%, primarily due to inadequate health systems [10,11,12,13,14]. From 2003-2022, few travel-related cases were reported outside endemic countries, but the number of cases was not very high [15,16,17,18,19,20]. However, a global outbreak of the MPox virus started on May 7, 2022 [21], and on July 23, 2022, the WHO declared MPox a Global Public Health Emergency (GPHE) [22]. This outbreak is linked to a new lineage, B.1 (clade 2b), with a higher mutation rate. This outbreak has resulted in 110 countries reporting about 87,000 cases and 112 deaths so far [23,24,25,26].

The MPox virus can enter hosts via respiratory or dermal routes. As a result, infection may occur in airway epithelial cells, keratinocytes, fibroblasts, and endothelial cells [27,28,29]. The incubation period of MPox is 5–21 days, and common symptoms include fever (between 38.5°C and 40.5°C), headache, and myalgia. A distinguishing feature of the MPox infection is the presence of swelling at the maxillary, cervical, or inguinal lymph nodes [30]. Currently, no FDA-approved treatments for MPox exist. At present, in the United States, there are three vaccines - JYNNEOS, ACAM2000®, and APSV [31], that are available. Out of these three vaccines, JYNNEOS has been approved by the FDA for smallpox and monkeypox in adults at high risk [32]. Tecovirimat is effective against Orthopoxvirus in animals but untested in human Mpox [33]. Human trials indicate minor side effects and ongoing randomized controlled trials are assessing its safety and efficacy for MPox [34]. Other potential treatments for MPox include VIGIV, cidofovir, and brincidofovir, with proven in vitro and animal efficacy but limited availability [35].

Since the first case of this outbreak, various policy-making bodies of the world have taken measures to contain the spread of the MPox virus. For instance, the New York City Health + Hospitals (NYC H+H), an integrated healthcare system, has been pivotal in the fight against emerging pathogens in the New York City region of the United States. As a result of the work done by NYC H+H, a total of 99,079 New Yorkers have received at least one shot of the MPox vaccine [36]. However, the uncertainty and challenges surrounding asymptomatic transmission of viruses such as Mpox [37,38] and the absence or inadequacy of appropriate and effective transmission-based personal protective equipment (PPE), such as N95 masks, face shields, gowns, and extended cuff examination gloves, may result is a possible resurgence of MPox [39,40]. The fifth meeting of the WHO’s International Health Regulations (IHR) Emergency Committee on the Multi-Country Outbreak of Mpox took place on May 10, 2023. At this meeting, the committee acknowledged remaining uncertainties about the disease, regarding modes of transmission in some countries, poor quality of some reported data, and continued lack of effective countermeasures in the African countries, where mpox occurs regularly [41]. On May 15, 2023, the US Centers for Disease Control and Prevention (CDC) warned about the potential resurgence of MPox cases in the US [42,43].

In today’s Internet of Everything era [44], social media platforms provide a seamless and virtual means for users to connect, communicate, and collaborate with each other. The popularity of social media platforms has been growing exponentially in the last few years. At present, 4.9 billion people use social media, and this number is expected to rise to 5.85 billion by 2027 [45]. Among the numerous social media platforms that have been used by the public in the last decade and a half, Twitter has been highly popular amongst all age groups. Twitter has over 368 million active monthly users [46]. Twitter stands out as the social media site that journalists choose to use [47] and is among the sites with the highest global adoption rates [48]. Twitter has been highly popular amongst researchers for studying, analyzing, modeling, and interpreting social media communications related to a wide range of research problems from different domains. During the virus outbreaks that took place in the last few years, such as COVID-19, H1N1, flu, Ebola, Zika Virus, Middle East Respiratory Syndrome (MERS), Measles, and West Nile Virus, just to name a few [49], topic modeling of Tweets helped to understand the perception, preparedness, response, views, and opinions of the general public during these virus outbreaks. Prior works in this field have primarily focused on the sentiment analysis and content analysis of Tweets about this MPox outbreak. Since the first case of this outbreak, only a couple of studies have been published that have focused on topic modeling of the Tweets. However, these studies have multiple limitations centered around (1) the limited time range of the Tweets that were analyzed and (2) the limited number of Tweets that were analyzed. Addressing this research gap serves as the main motivation for this work. The rest of the paper is organized as follows. A comprehensive review of recent works in this field is presented in Section 2. Section 3 describes the methodology that was followed for this work. It is followed by Section 4, which presents the results and highlights the novel findings of this work. Section 4 is followed by the conclusion section that summarizes the contributions of this paper and outlines the scope for future work in this area.

2. Literature Review

Mining and analysis of Tweets for the investigation and exploration of different research questions has been of significant interest to researchers from a wide range of disciplines in the last few years. While misinformation presents some challenges, it is still crucial to understand web behavior on Twitter and its implications for real-world decision-making. Therefore, this section is section is divided into three parts. Section 2.1 presents a brief review of mining and analysis of Tweets for interdisciplinary research. Section 2.2 discusses the recent advances related to the study and analysis of Tweets in the field of Healthcare. Section 3.3 outlines the latest works that focused on the mining and analysis of Tweets about the MPox outbreak.

2.1. A Brief Review of Recent Works Related to the Mining and Analysis of Tweets for Interdisciplinary Research

Shaheer et al. [50] analyzed tweets calling for tourism boycotts in China, Kerala (India), Spain, and South Africa due to concerns about animal abuse and identified three strategies used to mobilize support for the boycotts. Abu Samah et al. [51] proposed a web-based dashboard to visualize customer sentiment towards Malaysian airline companies, as expressed on Twitter. Similar to the tourism industry, the entertainment industry and politics have also benefitted from the mining and analysis of Tweets. Bodaghi et al. [52] explored the differences between the web behaviors of actors who spread fake news and those who spread the truth on Twitter. This study showed that while fake news has much better modularity and intra-to inter-links ratio, truth tweeters generally have higher page rank centrality. Ante’s study [53] focused on the analysis of Elon Musk’s Twitter activity on cryptocurrency to highlight the significance of the so-called “Musk Effect”, suggesting the seriousness of issues like market manipulation by influential individuals. Collins et al. [54] conducted a comparative analysis of over 2,000 tweets from the first two U.S. presidents of the “Twitter era”, Barack Obama and Donald Trump, to assess the impact of their online correspondence on America’s image abroad and its soft power. An interesting finding of this study was that the tone of presidential tweets can have significant and divergent effects on the perception of the U.S., even internationally. Beyond the border of the United States, Berrocal-Gonzalo et al. [55] found that politainment, the phenomenon of trivializing political information for entertainment purposes, manifested on Twitter during the Spanish general elections in April 2019, which prevented the creation of meaningful debates or interactions surrounding the elections on the platform.

The Big Data of conversations on Twitter is also considerably informative regarding the unremitting controversies underlying human society and human rights. Following the United States Supreme Court decision in Dobbs vs. Jackson Women’s Health Organization that overturned abortion rights, Chang et al. [56] presented a large-scale Twitter dataset collected on the abortion rights debate in the U.S. Similarly, Peña-Fernández et al. [57] analyzed the polarization produced in social media debates regarding the rights of feminism and transgender, specifically the use of the term “TERF” (trans-exclusionary radical feminist) on Twitter. The findings of this work show that online debates are poorly inclusive, suggesting the prevalence of community isolation. Goetz et al. [58] analyzed sentiments in food security-related Tweets in the U.S. during the early stages of the COVID-19 pandemic, from which they found that keywords of negative emotions were statistically correlated with contemporaneous food insufficiency rates reported in the Household Pulse Survey. Tao et al. [59] conducted a comparative study of posts on Twitter and Weibo regarding the Russian-Ukrainian War to reveal the differences in the topics of posts between the two platforms and to call for humanitarianism and peace. As can be seen from this brief review, mining, and analysis of Tweets holds the potential for the investigation of research questions across different disciplines. In the context of different virus outbreaks that the world has witnessed in the last decade and a half, healthcare-based research using Tweets has emerged as a crucial utilization of this vast potential of mining and analyzing Tweets. Some recent works are briefly reviewed in Section 2.2, which is followed by a dedicated review of recent works related to the analysis of Tweets about the MPox outbreak.

2.2. A Brief Review of Recent Works Related to the Mining and Analysis of Tweets for Healthcare Research

The Big Data of conversations and information exchange from social media platforms, specifically Twitter, has the potential to improve the efficiency, accuracy, and coverage of the healthcare systems in different geographic regions. Skovgaard et al. [60] analyzed Twitter discussions on personalized medicine. Their study revealed an intriguing distinction in attitudes between the professionals and the general private users, with the former considering personalized medicine as a promising future and the latter being concerned about new infrastructures and their implementation. Thakur et al. [61,62] developed a framework to address loneliness and social isolation in the elderly using Twitter data. Cevik et al. [63] performed a comprehensive study to analyze the sentiments of Tweets about Parkinson’s disease. Kesler et al. [64] applied topic modeling and qualitative content analysis to comments related to cancer-related cognitive impairment (CRCI), revealing the importance of coping mechanisms. Klein et al. [65,66] demonstrated the potential of using Twitter data to identify the start and end of the 40-week prenatal period, making it a valuable resource for observational studies on potential risk factors in pregnancy. Thackeray et al. [67] analyzed how Twitter is used during Breast Cancer Awareness Month (BCAM). The findings showed that while organizations and celebrities emphasized fundraisers, early detection, and diagnoses, the general public constituted the majority of the tweets that did not promote any specific preventive behavior.

Tweets provide a timely and authentic record of public perception and understanding of human health crises. The comprehensive study by Russell et al. [68] highlighted the need for regulatory changes to restrict online advertising volume about gambling activities during the COVID-19 lockdown. To investigate how the Dengue epidemic was reflected on Twitter, Gomide et al. [69] proposed an active surveillance methodology based on volume, location, time, and public perception. The analysis found a high correlation between the number of cases reported by official statistics and the number of tweets posted during the same period. The work done by Radzikowski et al. [70] reported that news organizations had a higher impact than health organizations in communicating health-related information. By examining the use of Twitter data to track public sentiment and disease activity related to H1N1 or swine flu, Signorini et al. [71] found that estimates of influenza-like illness derived from Twitter chatter accurately tracked reported disease levels from governmental organizations. Similar conclusions have been drawn from studies that focused on the recent outbreaks of listeriosis and cholera [72,73]. With the realization of Twitter’s ability to provide real-time information, Voss et al. [74] collected and analyzed tweets tagged with #WestNileVirus and #WNV and concluded that unusually higher temperatures and mosquito activities led to an increase in tweet numbers about WNV. In response to the 2015 U.S. Salmonella outbreak in cucumbers imported from Mexico, Bolotova et al. [75] were able to draw some insightful conclusions by analyzing Twitter communications at the time of CDC’s official announcements and the official release of the first recall of cucumbers. To inform health promotion efforts, Porat et al. [76] analyzed the content and source of popular tweets related to a diphtheria case in Spain. The most notable conclusion from their study is their suggestion for healthcare organizations to collaborate with popular journalists, news outlets, and science authors to address public concerns and misinformation through the outlet of social media like Twitter. As can be seen from this brief review, the study, analysis, and interpretation of multimodal characteristics of Tweets about different virus outbreaks has helped in the timely advancement of research in the field of Healthcare. Section 2.3 specifically highlights the recent works in this field that focused on the analysis of Tweets about MPox.

2.3. Review of Recent Works related to the Mining and Analysis of Tweets about MPox

Knudsen et al. [77] studied 262 Tweets to describe MPox risks to students. The results showed that credentialed Twitter users were 4.6 times more likely to tweet inaccurate information about MPox. Zuhanda et al. [78] studied 5,000 Tweets about MPox posted on August 5, 2022, to perform sentiment analysis. The results showed that 51.92% of the Tweets had a negative sentiment, and 48.08% of the Tweets had a positive sentiment. Ortiz-Martínez et al. [79] performed a study of 100 top Tweets about MPox posted on May 24, 2022. The findings showed that most of the Tweets were posted by informal individuals or groups (60%), followed by healthcare or public health (32%), and news outlets or journalists (8%). The work by Rahmanian et al. [80] involved studying 384,560 Tweets posted between May 16, 2022, and May 22, 2022. The findings showed that the majority of these Tweets were posted by individuals from the United States and Canada. Cooper et al. [81] studied Tweets containing the word “monkeypox” posted between May 1, 2022, and July 23, 2022. The results showed that a total of 48330 Tweets were posted by LGBTQ+ self-identified advocates or allies. The work of Ng et al. [82] focused on studying 352,182 Tweets about MPox posted between May 6, 2022, and July 23, 2022. The authors performed topic modeling of these Tweets and derived three themes - concerns of safety, stigmatization of minority communities, and a general lack of faith in public institutions. Bengesi et al. [83] mined over 500,000 multilingual tweets related to MPox and performed sentiment analysis. Olusegun et al. [84] studied 800,000 Tweets about MPox and used NRCLexicon to predict and measure the emotional significance of each Tweet. Farahat et al. [85] performed sentiment analysis on a total of 8532 Tweets about MPox posted between May 22, 2022, and August 5, 2022. The findings of sentiment analysis showed that 48% of the Tweets were neutral, 37% of the Tweets were positive, and 15% of the Tweets were negative. Sv et al. [86] studied a total of 556,402 Tweets about MPox posted between June 1, 2022, and June 25, 2022. The results of topic modeling showed that among the Tweets about MPox that had negative sentiments, there was a range of topics that were represented, such as deaths caused by the MPox virus, the severity of the virus, lesions caused by the virus, whether the virus is airborne, vaccines for the virus, and whether the virus will lead to the next pandemic. Mohbey et al. [87] developed a CNN and LSTM-based model to perform sentiment analysis of Tweets about MPox, and the accuracy of the model was found to be 94%. A dataset of Tweets about the MPox outbreak was developed by Nia et al. [88]. The work of Iparraguirre-Villanueva [89] focused on the detection of polarity in conversations on Twitter about MPox. The results showed that 45.42% of people expressed neither positive nor negative opinions, while 19.45% expressed negative and fearful feelings about MPox. AL-Ahdal [90] studied 15,936 Tweets about MPox posted by individuals from Germany. The results showed that the public displayed an impersonal feeling toward MPox.

As can be seen from this review and the research questions that were investigated in these papers, the major focus areas of these works have been sentiment analysis and content analysis. Only a couple of works have focused on the topic modeling of Tweets about MPox. However, those works have multiple limitations centered around the limited time range of the Tweets that were analyzed and the limited number of Tweets that were analyzed (discussed in detail in Section 4). This study aims to address this research gap. Therefore, topic modeling of 601,432 Tweets about the 2022 Mpox outbreak, which was posted on Twitter between May 7, 2022, and March 3, 2023, was performed in this study. Section 3 outlines the step-by-step methodology that was followed for the system design and implementation. The results and novel contributions of this work are presented and discussed in Section 4.

3. Methodology

This section presents the methodology that was followed for the system design and implementation. This section is divided into three parts. In Section 3.1, a technical overview of RapidMiner [91] is presented as RapidMiner was used for this work. Section 3.2 presents the description of the topic modeling architecture that was used in this work. Section 3.3 outlines the steps that were followed for the implementation of this topic modeling architecture, along with the specifics of the system design.

3.1. Technical Overview of RapidMiner

RapidMiner is a Data Science software platform that allows the development and implementation of different algorithms. It enables its users to visually design data workflows and build predictive models using a graphical user interface (GUI). RapidMiner allows the development of applications, workflows, and algorithms, which are known as “processes” that consist of multiple “operators”, which may be built-in and/or user-defined. Every “process” and each of its “operators” are associated with a specific functionality to make the entire “process” work. RapidMiner provides a wide range of built-in “operators” that may be directly used for the implementation of various tasks. There also exist certain “operators” that may be used to modify the functionality or features of other “operators”. The platform also allows developers to create their own “operators” which can be shared via the RapidMiner Marketplace. RapidMiner is developed on a client-server model with public and private cloud infrastructures. There is a free version and an enterprise version. The free version has a data processing limit of 10,000 rows for any “process”, and the enterprise version does not have any limits on the number of rows in a data file that it can process [91,92]. For this research work, the educational license of RapidMiner (available to researchers in academia upon request) was used in RapidMiner Studio 10.1.001. With the educational license, RapidMiner Studio can be used to process any number of rows in any dataset. The following represent a few notable characteristics of RapidMiner Studio [93,94]:

• It supplies pre-built “operators” encompassing distinct functions that can be directly employed or customized for the creation and execution of algorithms and applications related to Machine Learning, Data Science, Artificial Intelligence, and Big Data.
• RapidMiner is developed using Java, which ensures that RapidMiner “workflows” retain the Write Once Run Anywhere (WORA) attribute of Java.
• The platform permits the installation of various extensions to facilitate seamless connectivity and integration of RapidMiner “workflows” with other software and hardware environments.
• Scripts developed in programming languages, like Python and R, can also be imported into a RapidMiner “workflow” to supplement its functionalities.
• The software enables the creation of new “operators” and effortless dissemination of the same within the RapidMiner community.
• RapidMiner consists of “operators” that enable it to establish connections with social media platforms, such as Twitter and Facebook. Such connections facilitate the extraction of tweets, comments, posts, reactions, and other relevant social media interactions.

3.2. Description of the Topic Modeling Architecture for System Design

Latent Dirichlet Allocation (LDA) [95] is a generative probabilistic model used in the field of Natural Language Processing and Machine Learning. It is commonly used for topic modeling, which is the task of identifying topics within a collection of documents. In LDA, the mixture of topics is derived from a consistent Dirichlet prior, which is the same across all documents. The procedure [96] for creating a corpus is outlined as follows (in this context, the focus is on smoothed LDA). Thereafter, the likelihood of generating a corpus can be represented as shown in Equation (1).

• Select a multinomial distribution ${\varnothing }_z$ for each topic z from a Dirichlet distribution with parameter $\beta$
• For each document d, pick a multinomial distribution ${\theta }_d$ also from a Dirichlet distribution with parameter $\alpha$
• In document d, for each token word w, pick a topic z $\in$ {1….K} from the multinomial distribution ${\theta }_d$
• Pick a word w from the multinomial distribution ${\theta }_z$

  $$P\left(Do{c}_1,\dots .,Do{c}_N|\alpha ,\beta \right)=\int \int {\displaystyle \prod }_{z=1}^{K}P\left({\varphi }_z|\mathsf{\beta }\right){\displaystyle \prod }_{d=1}^{N}P\left({\varphi }_d|\mathsf{\alpha }\right)\left({\displaystyle \prod }_{i=1}^{N_d}{\displaystyle \sum }_{z_i=1}^{K}P\left(z_i|\mathsf{\theta }\right)P\left(w_i|z,\mathsf{\theta }\right)\right)d\mathsf{\theta }d\varphi$$

  (1)

To use language models for information retrieval in an LDA, an approach using the query likelihood model is used, where each document is scored by the likelihood of its model generating a query Q. This is shown in Equations (2) and (3). In Equation (2), D represents a model for documents, Q stands for the query, and q denotes an individual term within the query Q. P(Q|D) signifies the probability of the document model generating the query terms, following the assumption of ‘bag-of-words’. This assumption considers that terms are independent when given the documents. $P\left(q_i|D\right)$ is specified by the document model with Dirichlet smoothing. In Equation (3), P(w|D) is the maximum likelihood estimate of word w in document D, and P’(w|coll) is the maximum likelihood estimate of the same word w in the entire collection. However, directly implementing the LDA model hurts retrieval performance. So, in a prior work in this field [96], the original document model (Equation 3) was combined with the LDA model to construct a new LDA-based document model, as shown in Equation (4). The LDA model introduces a novel document representation centered around topics. After obtaining the posterior estimates for θ and φ, the word probability within a document can be computed using Equation (5), where $\widehat{\mathsf{\theta }}$ and $\widehat{\mathsf{\varphi }}$ are the posterior estimates of $\theta$ and $\varnothing$ [96].

$$P\left(Q|D\right)={\displaystyle \prod }_{q\in Q}^{}P\left(q|D\right)$$

(2)

$$P\left(w|D\right)=\frac{N_d}{N_d+\mu }{P}_{ML}\left(w|D\right)+\left(1-\frac{N_d}{N_d+\mu }\right)P_{ML}\left(w|coll\right)$$

(3)

$$P\left(w|D\right)=\lambda \left(\frac{N_d}{N_d+\mu }{P}_{ML}\left(w|D\right)+\left(1-\frac{N_d}{N_d+\mu }\right)P_{ML}\left(w|coll\right)\right)+\left(1-\lambda \right)P_{lda}\left(w|D\right)$$

(4)

$$P\left(w|d,\widehat{\mathsf{\theta }},\widehat{\varphi }\right)={\displaystyle \sum }_{\mathrm{z}=1}^{\mathrm{K}}P\left(w|z,\widehat{\varphi }\right)P\left(z|\widehat{\mathsf{\theta }},d\right)$$

(5)

In this research work, SparseLDA was implemented, as prior work has shown that it is 20 times faster than the traditional LDA [97]. In the SparseLDA framework [97], given an observed word type w, the probability of topic z in document d can be computed using Equation (6).

$$P\left(z=t|w\right)\alpha \left({\alpha }_t+n_{\left(t|d\right)}\right)\frac{\beta +n_{w|t}}{\beta V+n_{.|t}}$$

(6)

However, Equation (6) requires the calculation of the unnormalized weight q(z) for all topics to determine the normalizing constant for the distribution ${\displaystyle \sum }_{\mathrm{z}}q\left(z\right).$ A simpler approach involves storing a significant portion of the computation needed to calculate the normalization constant. Through the reorganization of terms within the numerator, Equation (5) can be portioned into three distinct sections, as shown in Equations (7) to (10). Here, the first term is constant for all documents, and the second term is independent of the current word type. Moreover, the ${\displaystyle \sum }_{z}q\left(z\right)$ corresponds to the sum across topics for each of the three components in Equation (7) [97].

$$P\left(z=t|w\right)\alpha \frac{{\alpha }_t\beta }{\beta V+n_{.|t}}+\frac{n_{t|d}\beta }{\beta V+n_{.|t}}+\frac{\left({\alpha }_t+n_{t|d}\right)n_{w|t}}{\beta V+n_{.|t}}$$

(7)

$$s={\displaystyle \sum }_t\left(\frac{{\alpha }_t\beta }{\beta V+n_{.|t}}\right)$$

(8)

$$r={\displaystyle \sum }_t\left(\frac{n_{t|d}\beta }{\beta V+n_{.|t}}\right)$$

(9)

$$q={\displaystyle \sum }_t\left(\frac{\left({\alpha }_t+n_{t|d}\right)n_{w|t}}{\beta V+n_{.|t}}\right)$$

(10)

This divides the full sampling mass into three buckets. Now, $U~\mathcal{U}\left(0,s+r+q\right)$ can be sampled. If U < s, it would imply hitting the smoothing-only bucket. Thereafter, the process involves stepping through each topic and calculating and adding $\frac{\beta {\alpha }_t}{\beta V+n_{.|t}}$ for that topic until it is greater than ‘x’. For the document bucket, s < r < (s+r), the process involves iterating through the set of topics that satisfies $n_{t|d\ne 0}$. The constant ‘s’ only changes when the hyperparameters are updated. Conversely, the constant ‘r’ is solely influenced by document-topic counts. This permits the computation of ‘r’ once at the start of each document and subsequently modifies it by subtracting and adding values related to the prior and current topics in each Gibbs update. This process takes constant time, independent of the number of topics. The topic word constant, ‘q’, changes with the value of w, so old computations cannot be easily recycled. However, the performance can be significantly improved by splitting q into two components, as shown in Equation (11). With this Equation, the coefficient $\frac{\left({\alpha }_t+n_{t|d}\right)}{\beta V+n_{.|t}}$ can be cached for every topic. Calculating ‘q’ for a specific ‘w’ involves performing a single multiplication operation for each topic, where $n_{w|t}$ is non-zero. Given that $n_{t|d}$ = 0 for all topics within any document, the coefficients vector will predominantly comprise only $\frac{\left({\alpha }_t\right)}{\beta V+n_{.|t}}$.Therefore, this allows the optimization of the LDA model by storing these coefficients across documents, refreshing values only for topics with non-zero counts in the current document, and reverting these values to α-only values upon finishing the sampling process for that document [97].

$$q={\displaystyle \sum }_t\left[\frac{\left({\alpha }_t+n_{t|d}\right)}{\beta V+n_{.|t}}\times n_{w|t}\right]$$

(11)

3.3. Description of the System Design and Implementation

This section describes the system design and its implementation, as well as the dataset that was used for performing this research work. The dataset that was used comprises 601,431 Tweet IDs of Tweets about MPox posted between May 7, 2022, and March 3, 2023 [98]. This dataset complies with the FAIR principles of scientific data management [99]. This dataset contains only Tweet IDs, and the standard procedure for working with such datasets is that the dataset is hydrated to obtain the text of the Tweets and related information. However, this dataset was developed by the first author of this paper, so all the Tweets were already available, and hydration of the Tweet IDs was not necessary. Figure 1 shows the system design in RapidMiner. This is a “process” that was developed in RapidMiner Studio 10.1.001 (with the Educational License) to set up this system, and this “process” comprises different “operators” with different functionalities.

In this Figure, the “MPox_Tweets-Data” “operator” represents the Tweets from the dataset described earlier in this section. These Tweets were imported into RapidMiner Studio to develop this “process”. All 601,431 Tweets about MPox posted between May 7, 2022, and March 3, 2023, were used for the development of this LDA model. Thereafter, separate “operators” were developed to perform the different steps of the data processing. The data processing comprised the following steps, and for each of these steps, a separate “operator” was developed in this RapidMiner “process”. For steps (a), (b), (c), and (d) of the data preprocessing, different regular expressions were developed and applied to define the functionalities of these “operators”.

a)

Removal of characters that are not alphabets.

b)

Removal of URLs.

c)

Removal of hashtags.

d)

Removal of user mentions.

e)

Detection of English words using tokenization.

f)

Stemming and Lemmatization.

g)

Removal of stop words

h)

Removal of numbers

After completion of the data pre-processing, an LDA model was developed and implemented in RapidMiner as per the architecture of the parallel topic model and SparseLDA (described in Section 3.2) by customizing and utilizing the “Extract Topics from Data (LDA)” operator in RapidMiner Studio 10.1.001. The number of iterations for optimization was set to 1000, and the frequency of hyperparameter optimization was set to 10. As discussed in prior works in this field [100,101,102,103], the average coherence value of an LDA model serves as a key indicator for the determination of the optimal number of topics. So, this “process” (shown in Figure 1) was repeatedly run by varying the number of topics from 2 to 50, and the average coherence value of the model was computed and recorded for each of these runs. The results of the same to deduce the optimal number of topics, as well as the specific topics that were identified in the Tweets, are presented in Section 4.

4. Results and Discussions

This section presents the results of this work. As stated in Section 3.3, the LDA model (shown in Figure 1) was run by varying the number of topics from 2 to 50 to determine the optimal number of topics based on the analysis of the average coherence value for each run. Table 1 represents the average coherence value of this LDA model from each run. This table was compiled by varying the number of topics from 2 to 50. An analysis of the same is presented in Figure 2.

From Table 1 and Figure 2, the optimal number of topics was determined to be 4, as the LDA model produced the highest coherence score for the same. It is worth mentioning here that negative values of coherence scores are not unusual for an LDA model that follows a system architecture as described in this paper, and a recent work [104] followed a similar system architecture for the LDA model that was developed and the authors of that work obtained negative values for the coherence scores for all the number of topics, and the lowest value of the coherence score reported was as low as about -11.5.

After determining that the optimal number of topics for this LDA model is 4, this model was run by setting the number of topics as 4, and the characteristics from the output were observed and analyzed. For each Tweet, the RapidMiner “process” computed a confidence value for each topic and then predicted a topic for that Tweet based on the highest confidence value. This is shown in Figure 3. To avoid an output table comprising 601,431 rows, this Figure shows a random selection of 17 rows from this output table. In Figure 3, the attributes – confidence(Topic_0), confidence(Topic_1), confidence(Topic_2), and confidence(Topic_3), represent the confidence values of each Tweet belonging to Topics 0, 1, 2, and 3, as computed by the LDA model. The last attribute – Tweet_Text in this Figure, shows the Tweet (after data preprocessing) that was analyzed. For each Tweet, the highest of these confidence values was used to predict the topic for the same. For instance, in the first row in Figure 3, it can be seen that confidence(Topic_0) has the highest value, so the predicted topic for this Tweet was Topic 0. In a similar manner, the LDA model predicted the topics for all the Tweets of this dataset. The Tweets belonging to each of these topics – Topic 0, Topic 1, Topic 2, and Topic 3 were studied to understand the underlying themes of conversation that represented each of these topics.

Based on this study, the broad themes which represented these topics were observed to be “Views and Perspectives about MPox”, “Updates on Cases and Investigations about Mpox”, “MPox and the LGBTQIA+ Community”, and “MPox and COVID-19”. Table 2 represents a random selection of five Tweets for each of these Topics. In Table 2, these Tweets are presented in “as is” form, i.e., in the manner in which they were originally posted on Twitter to provide better context as compared to the preprocessed version of these Tweets.

Figure 4 shows an analysis of the number of Tweets per topic. As can be seen from Figure 4, Topic 0, or the theme of Views and Perspectives about Mpox, was most popular on Twitter (in terms of the number of Tweets posted) during this time range of May 7, 2022, to March 3, 2023. It was followed by Topic 2, or the theme of MPox and the LGBTQIA+ Community. This was followed by Topic 3 (or the theme of MPox and COVID-19) and Topic 1 (or the theme of Updates on Cases and Investigations about Mpox), respectively.

Next, a comparative study with prior works in this field is presented. Table 3 presents a summary of recent works in this field that focused on the mining and analysis of Tweets about MPox. As can be seen from Table 3, most of the works in this field have focused on Sentiment Analysis and Content Analysis of Tweets. At the same time, a couple of works [82,86] also exist in this field where topic modeling of Tweets about Mpox was performed. As can be seen from Table 3, topic modeling of Tweets about MPox is a research area that remains less explored and less investigated. A comparison of the different characteristics of this work with these two works [82,86] is presented in Table 4.

Table 4 outlines the limitations in a couple of similar works [82,86] in this field and highlights how the work of this paper addresses the same. Specifically, the limitations in these two works can be listed as follows:

• Limited time range of the analyzed Tweets: The time range of the Tweets that were analyzed in these works represents Tweets that were posted only during certain months of the 2022 MPox outbreak. One of the works [82] included Tweets that were posted on the day the first case of the 2022 MPox outbreak was recorded (May 7, 2022), but the other work [86] didn’t. Furthermore, none of these works have analyzed Tweets posted after July 23, 2022.
• Limited number of Tweets: The number of Tweets that were investigated in these works are 352,182 and 556,402, respectively. These numbers represent a fraction of the total number of Tweets that have been posted since the first recorded case of the 2022 Mpox outbreak on May 7, 2022.

These limitations that exist in similar works in this field are addressed in this paper. First, the dataset that was used for developing the LDA model comprises Tweets about MPox, which were posted on Twitter between May 7, 2022, and March 3, 2023 – a time range that is greater than the time ranges of the similar works shown in Table 4. Second, a total of 601,432 Tweets were analyzed in this study. This is much higher than the number of Tweets which were analyzed in similar works in this field. Thus, to summarize, the time range of the Tweets and the number of Tweets that were analyzed in this study further support the scientific contributions of this work.

5. Conclusions

In the last decade and a half, the world has experienced the outbreak of a range of viruses such as COVID-19, H1N1, flu, Ebola, Zika Virus, Middle East Respiratory Syndrome (MERS), Measles, and West Nile Virus, just to name a few. In today’s Internet of Everything era, the popularity of social media platforms has been growing exponentially. Social media platforms have served as virtual communities during the outbreak of such viruses in the past, allowing people from different parts of the world to share and exchange information, news, perspectives, opinions, ideas, and comments related to the outbreaks. Researchers from different disciplines have analyzed this Big Data of conversations related to virus outbreaks on social media platforms such as Twitter using concepts such as Topic Modeling to understand the underlying themes of conversations and information exchange of the general public. The recent outbreak of the MPox virus has resulted in a tremendous increase in the utilization of social media platforms such as Twitter. Prior works in this field have primarily focused on sentiment analysis and content analysis of Tweets about MPox, and a couple of works in this field that have focused on topic modeling have multiple limitations. This paper aims to address this research gap and makes two scientific contributions to this field. First, it presents the results of performing topic modeling of 601,432 Tweets about the 2022 MPox outbreak, which was posted on Twitter between May 7, 2022, and March 3, 2023. These results indicate that the conversations related to MPox during this time range may be broadly categorized into four distinct themes - Views and Perspectives about MPox, Updates on Cases and Investigations about MPox, MPox and the LGBTQIA+ Community, and MPox and COVID-19. Second, the paper presents the findings from the analysis of the Tweets that focused on these topics. The results show that the theme that was most popular on Twitter (in terms of the number of Tweets posted) during this time range was Views and Perspectives about MPox. It is followed by the theme of MPox and the LGBTQIA+ Community. This theme is followed by the themes of MPox and COVID-19 and Updates on Cases and Investigations about Mpox, respectively. As per the best knowledge of the authors, no similar work has been done in this field thus far. Future work in this area would involve collecting more Tweets and repeating this study in a few months to evaluate and interpret any variations in terms of the themes of conversations on Twitter related to MPox.