This study falls into three main stages. These include (1) the first stage involves acquiring data on Twitter; (2) the second stage concentrates on the initial operation of preprocessing, which was conducted for cleaning and filtering out irrelevant information (i.e., punctuation, stop words, and retweet symbols) from the tweets; (3) the third stage involves utilizing NLTK’s VADER classifier and identifying the most frequent terms used in public conversations on thalassemia screening. The application of the scoring method was carried out to the results of VADER to assess the method’s capability of classifying the collected tweets to a three-point measurement (either positive or negative, and neutral). The architecture of the introduced approach is shown in Fig. 1.

The data in the context of this study comprise a total of 3,376 English tweets, which were gathered between February 2009 and September 2020, applying the TWINT application. The collected tweets were crawled utilizing several search keywords, e.g., “thalassemia screening.” Fig. 2 shows the number of tweets posted each year, starting from 2009 until 2020. It is observed from the Figure that the number of tweets published related to thalassemia increases over time. This confirms how people are aware of thalassemia screening and its impact on society.

A tweet can be defined as a specific microblog message, which is posted on the Twitter platform and is confined to a total of 280 characters. Often, users do not apply the proper language structure when they post opinions or sentiments about a specific topic. Instead, slang, misspelling, different emoticons, and abbreviations, sometimes puns, complicates the analysis of these structures. Complex textual data significantly influence the performance of analyzing sentiments because the quality of the output depends on the input [29]. Therefore, a preprocessing steps series have been performed for removing irrelevant data from the collected tweets because the cleaner the information is, the more appropriate this information will be for mining, as well as feature extraction, thereby improving the findings’ accuracy and precision [30,31]. For the preprocessing, we have utilized Python’s natural language processing (NLP) toolkit (NLTK). Initially, a given regular expression, i.e., (Regex) in Python is applied for detecting and eliminating uniform resource locator (URL) links, retweet symbols (RT), user mentions (@), and decoding hypertext markup language (HTML) Entities with equivalent characters. Hashtags (#) describe the subject of the tweet. They carry useful information about the topic of the tweet, included as a particular part of a given tweet, except for the symbol “#”, which is deleted [31]. Once the preprocessing steps are complete, a dataset is now ready for the classification of sentiments by using the algorithm of VADER. Various NLTK functions are also used as additional preprocessing steps and are required to calculate terms’ frequency and thus to visualize in Word Cloud. These steps include:

Convert the tweets to lowercase to reduce redundancy.

The analysis of sentiment can be generally used for examining the context polarity, whether a positive or negative opinion and a neutral sentiment. During this phase, a largely approved human-validated sentiment analysis instrument was used, i.e., Valence Aware Dictionary for Sentiment Reasoning (VADER). This analysis tool was used because it represents a lexicon-based sentiment analysis engine, which is specially tailored for expressions on social media sentiments [22]. It is a wholly open-sourced lexicon under the MIT License. We use the VADER analyzer to examine both the sentiment’s polarity and intensity in each tweet. This operation results in four sentiment scores, including positive, negative, neutral, and compound. The compound score is a unidimensional metric with a specific value between −1, referring to (extremely negative), in addition to 1, referring to (extremely positive), which is a very helpful metric to measure the overall sentiments in each tweet. Table 2 demonstrates the compound threshold value utilized for classifying tweets as positive or negative tweets and neutral tweets.

Based on Table 2, 0.05 and −0.05 were used as threshold values to explain compound values, i.e., positive values, negative values, as well as neutral values. If the compound value is above or equal to 0.05, it signifies a positive value. If it is less than or equivalent to −0.05, it signifies a negative value. Otherwise, it signifies a neutral value.

Although VADER has been substantiated by researchers on typical tweets according to Hutto et al. [22], VADER’s performance in categorizing tweets linked to public health issues, especially thalassemia screening, requires additional validation. Accordingly, validation has been performed by selecting a randomly selected subset of a total of 320 tweets obtained from an original set of thalassemia-screening tweets. The sentiment polarity of each of the 320 tweets is interpreted by domain experts as actual results. However, the poor performance of the F1 score (<0.7) has been observed based on the classification of VADER, whereby the primary reason is detected.

In this original lexical dictionary of VADER, some words appeared in tweets that can reverse the polarity score from negative to positive and vice versa, relying on their score values in VADER. For example, In the tweet “I did thalassemia test; unfortunately, it was positive,” this tweet shows a positive compound score of 0.296. In fact, it is originally a negative sentiment. Its polarity score has been changed to positive due to the word ‘positive’ appearing in the tweet with a high sentiment value (i.e., 2.6) which has affected the final compound score to be positive at the end. Similarly, the word ‘negative’ refers to thalassemia screening results containing a high negative sentiment value (i.e., −2.7).

This can affect the sentiment polarity towards the negative side once it is a part of thalassemia screening conversations, affecting sentiment classification accuracy. In addition, we have also removed the word ‘cancer’ from the original lexical dictionary since these words do not contain any positive or negative connotations as they are a part of the health domain. Consequently, that has resulted in a more favorable F1 score of 0.818. Thus, other classification variations between using VADER and utilizing a human rater refer to challenging obstacles faced in classifying sentiments, including ambiguity, sarcasm, as well as mixed sentiments, which seem quite difficult for the human rater to identify. Eventually, this study presented the VADER analyzer’s adjusted version to measure thalassemia screening tweets’ sentiment scores.

TWINT, NLTK (Natural Language Toolkit), and VADER constitute the three main tools, which were applied in this study. TWINT has been applied for collecting the raw data from Twitter. NLTK, a text analysis Python library, has been utilized for data pre-processing activities to prepare it for sentiment classification. VADER, a sentiment lexicon, has been used to categorize the relevant tweets’ sentiments into ternary classes.

TWINT represents a sophisticated Twitter scraping instrument, which is written in Python and aims at enabling scholars to scrape stored tweets collected from profiles on Twitter without using API authentication. TWINT utilizes the search operators of Twitter for scraping the tweets. Additionally, it aims to scrape tweets, which are linked to hashtags, topics, and trends. It also aims to sort out sensitive data from the Tweets’ emails and users’ phone numbers in a helpful way. Moreover, TWINT provides queries to Twitter, and enables scraping the users’ followers on Twitter, special tweets, which are loved by users, and their likes and follows without using an API authentication or even browser emulation [32]. TWINT is a standard Python library used in various text analysis research as a primary tool for data acquisition [26,33].

Natural Language Toolkit, i.e., NLTK represents a Python library. It provides a specific base for building Python programs, as well as data classification. The toolbox also plays a key role in converting textual data to a certain format, where sentiments are extracted. NLTK mainly aims to perform natural language processing via performing analysis of human language data. Also, NLTK provides different functions for pre-processing data to perform all NLP methodologies, such as part-of-speech tagging, tokenizing, lemmatizing, stemming, parsing performing sentiment analysis for specified datasets. By doing so, the available data can be fit for feature extraction and mining [34].

Vader represents a gold-standard sentiment lexicon; it is mainly attuned to contexts on social media platforms. It aims to combine several lexical features considering 5 generalizable rules: (1) Punctuation, (2) Capitalization, (3) Degree modifiers, (4) Polarity shift due to Conjunctions, and (5) Catching Polarity Negation. The rules are syntactic and grammatical conventions used by people when they highlight or express the intensity of sentiment. Hutto et al. [22] compared VADER efficiency to eleven benchmark models, involving linguistic inquiry and word count (LIWC), affective norms for english words (ANEW), the General Inquirer, SentiWordNet, and other techniques of machine learning, including Naive Bayes, Maximum Entropy, as well as SVM algorithms. They concluded that the VADER sentiment lexicon generalized more favorably over contexts than other models. Botchway et al. [26] and Kumaresh et al. [25] examined various lexicons in text classification and concluded that VADER had beaten other lexicons in terms of accuracy. Indeed, the VADER sentiment lexicon discriminates itself from other models as being more sensitive to some sentiment phrases in certain social media texts. VADER enjoys the capability of generalizing more efficiently in other domains. Besides common dictionary words, it also gives sentiment scores on emoticons, slang (nah, meh, giggly, etc.), and acronyms (LOL, OMG, etc.) [22].

Primarily, there were 3,375 tweets were found during the data collection period associated with the thalassemia screening topic. The baseline monthly thalassemia screening tweets volume has fluctuated between (20) and (50) tweets, whereby an explosive number of tweets were observed in May due to world thalassemia day, celebrated every year on May 8th. The tweets memorialized thalassemia victims and encouraged those who are struggling to live with the disease, raising awareness among vulnerable people against thalassemia risks and encouraging them to do screening early. The volume of tweets gradually dropped down and returned to the baseline, as illustrated in Fig. 3.

Fig. 4 displays the most 25 unigram words that are frequently used while people were having conversations about thalassemia screening. As observed in Fig. 4 above, many words are relevant to blood, screen, get, marriage, and patient. Similarly, Fig. 5 displays the most 25 frequent bigrams and common words, such as (make, mandatory), (blood, marriage), (pregnant, women), (get, marry), and (blood, transfusion). For example, the terms “Getting Married”, “Pregnant Woman”, “Genetic”, “Carrier” and “Prevent” were observed in both bigram and unigram graphs, emphasizing the dire need for thalassemia screening for people who will get married and for pregnant women, which allowed them to access their status of thalassemia at an early stage. Therefore, they have adequate time to identify the risks and prevent the disease from being passed from parents to children through genes. Moreover, the terms, “Blood Transfusion”, “Blood Donation”, “Blood Group” and “Iron Deficiency” emphasized an urgent need for blood donation because of the lack of iron and the periodic need for thalassemia patients for blood transfusions.

Fig. 6 displays the top co-occurring bigrams visualization in the dataset as a network diagram with the help of a Python package named NetworkX. Based on the graph in Fig. 6, it is easier to understand the relationship between the words (nodes), which frequently appeared together by drawing the edge that determines the words’ connectedness to each other. For instance, combinations of connected terms in sickle-cell-anemia disease were observed, indicating that most patients of thalassemia have severe forms of the disease and suffer from chronic anemia (sickle cell disease). Additionally, a combination of “hiv-hepatitis” terms was observed with the co-occurrence of words for three main reasons:

Some tweets complained that most people do not pay attention to thalassemia screening as much as they are concerned about HIV and hepatitis tests.

Fig. 7 maps the countries mentioned in public conversations regarding thalassemia screening. This map helped determine the correlation between countries with a high rate of thalassemia prevalence and active conversations over thalassemia screening on Twitter. Table 3 demonstrates the 10 most countries mentioned during people’s conversations about thalassemia screening.

As remarked in [35], more than 70,000 babies are born with thalassemia worldwide each year. This defect is observed more often in the Indian subcontinent, the Mediterranean, Southeast Asia, and West Africa. Also, most children with thalassemia are born to women in countries, where people receive a low income. Thalassemia was first confined to the tropics and subtropics. Currently, it is commonly found worldwide, and that is because of the substantial population migrations. Moreover, Beta-thalassemia is most prevalent in Mediterranean, African, and Southeast Asian descent, whereas Alpha thalassemia is much more widespread among African and Southeast Asian descent.

From the figure above, we can observe the highest countries mentioned in tweets, including India and Pakistan. They were expressed in the extracted data with 181 and 164 tweets, respectively. According to the study conducted by [36], most of the available info about thalassemia in the South Asian region originated from studies in India. Because of severe heterogeneity, the variable frequency of the beta-thalassemia heterozygote or carrier ranging between 1%–10% was registered in different areas of India. Nevertheless, the general prevalence of beta-thalassemia carriers was in India between 2.78% and 4%. This represents nearly 30–48 million of India’s population, in other words, approximately 5–12 million carriers of thalassemia with a rate, ranging between 5% and 7% in Pakistan [37–39].

The algorithm of VADER lexicon-based was utilized for extracting features via importing SentimentIntensityAnalyzer from vaderSentiment in the NLTK library. The utilized method “polarity_scores ()” from SentimentIntensityAnalyzer () module has provided four sentiment scores as the output scores, comprising positive, negative, as well as neutral scores, with the compound scores. Regarding the compound scores, every sentiment orientation of the tweet has been identified in accordance with its values. These scores are in a dictionary in Python, whereas the compound scores are extracted for measuring whether each tweet is a positive tweet, a negative tweet, or a neutral tweet with a value ranging from [−1 to +1]. Table 4 shows the tweets’ sentiments with corresponding values after using the threshold (Table x) to classify tweets into three predefined categories. The values of [1, 0, −1 ] were set to refer to positive public sentiments, neutral, as well as negative public sentiments [40].

Fig. 8 displays the sentiment results by the number of tweets of each class in a bar chart using the matplotlib.pyplot Python library.

Based on the number of tweets by sentiment in Fig. 8 and Table 5, there are about 3376 tweets, incorporating all sentiment categories. Most tweets referred to positive and neutral sentiments regarding thalassemia screening. Remarkably, (1569) 46.5% of the obtained tweets indicated positive sentiments, while (1302) 38.6% of these tweets indicated neutral opinions, and (503) 14.9% referred to negative sentiments. The positive tweets were the highest in number compared with other classifications as the value of the given compound threshold expressed positive sentiments regarding thalassemia screening.

Fig. 8 and Table 5 illustrates the word cloud of the negative and positive tweets labeled through VADER to manifest the dominant words for each class. The search keywords (i.e., thalassemia and screening), and stop words (i.e., they, I, do, them) were removed because they appeared in almost all tweets. This helped Wordcloud focus on terms only and be more productive in gaining insights precisely. It was observed that the most frequent words appearing in both Figures are domain-related, such as blood, disease, disorder, alpha, beta, genetic, etc. as we expected. By looking at the Wordcloud of positive tweets, some words were observed like prevent, awareness, free, support, mandatory, marriage, camp, and pregnancy. The results showed considerable solidarity and support among the thalassemia community on Twitter translated by organizing free test camps, spreading awareness among the public, and making mandatory tests before marriage and during pregnancy. The Wordcloud of negative tweets showed the following words: child, suffer, HIV, and negligence, which highlighted the suffering of children from thalassemia during blood transfusion who were infected with HIV due to negligence in blood screening.

A Word Cloud also called a tag cloud, or a text cloud is a tool for visually summarizing a vast amount of text [25]. In this study, Fig. 9 showed the word cloud to visualize the tweets for every sentiment classification based on the VADER scores after removing the meaningless information.

Fig. 10 demonstrates the distribution of the negative and positive compound scores as a histogram plot after neglecting the neutral ones.

Based on Fig. 11, the negative opinions do not correspond to the positive and negative sentiments. The positive tweets have a significant numerical advantage over the negative ones in terms of the compound score. The mean of the positive tweets is 0.4758, as opposed to negative tweets, which have an average of about −0.4234.

As shown in Fig. 11, a box plot gives graphical information concerning the positive and negative tweets such as batch location, dispersion, and the data set’s skewness. Also, a boxplot drew attention to specific possible outliers. Therefore, it is easier to compare and reflect on these 4 features of our data sets.

Assessment of batch location: The figure above indicates that the median compound score of the obtained positive tweets was larger compared with the negative tweets.

Regarding the model evaluation of the proposed approach, we used four performance measures, such as True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). These four performance measures represent the confusion matrix and are shown in Fig. 12. The confusion matrix represents a mixture of predicted and actual observations to measure the current algorithm's effectiveness by determining precision and recall, in addition to the F1 Score.

■ Precision

Precision represents the observations’ rate, which correctly forecasted positive to total forecasted positive ones.

(1)Precision=TP/(TP+FP)

■ Recall

Recall, which is referred to as sensitivity, signifies the observations’ rate, which is correctly forecasted positive to the entire real class observations.

(2)Recall=TP/(TP+FN)

■ F-measure

F-measure or F-score refers to a performance measure, which combines precision with recall by determining the weighted harmonic mean covering accuracy flaws with the skewed data.

(3)F−measure=2∗(Recall∗Precision)/(Recall+Precision)

Fig. 12 shows the confusion matrix of 320 randomly selected sentiments. They were manually labeled by three domain expert annotators at the tweet level for identifying ternary classes (actual labels) to be compared to the results of the VADER classifier (predicted labels). According to the obtained results, the total number of the actual positive sentiments, the neutral sentiments, as well as the negative sentiments is 137, 130, 53, correspondingly. When the VADER lexicon was used, the model anticipated that the positive attitudes number will be 148, 109 for neutral sentiments, and 63 for negative opinions.

Table 6 compares several related studies about public health regarding precision (P), recall (), and F-score (F1). It is, therefore, concluded that the proposed approach achieved promising results, which outperformed many proposals with 0.829 precision, 0.816 recall, and 0.818 F1 score. However, holding a comparison between approaches is a difficult task for a couple of reasons. First, these approaches are oriented toward various sentiment analysis stages [41]. Second, the proposed methods are based on various approaches (i.e., machine learning or lexicon-based), topics, as well as domains, whereby the linguistic resources utilized vary in size and context, making the comparison between the proposals so tricky. Therefore, to perform an appropriate comparison between the proposals, the same corpus in all evaluations is necessitated.

Vader is a gold-standard sentiment lexicon that is particularly developed for microblogs such as Twitter. It combines the lexical features with consideration for five generalizable rules (punctuation, capitalization, degree modifiers, constructive conjunction, and Tri-gram examination to identify negation), which are grammatical and syntactic conventions that humans use to express and emphasize sentiment intensity. Indeed, VADER has outperformed individual human raters at correctly classifying the sentiment of tweets into positive, neutral, or negative classes by 0.96 to 0.84 of the F1 measure (Hutto & Gilbert, 2014). The VADER sentiment lexicon discriminates itself from others in that it is more sensitive to sentiment expressions in social media contexts and generalizes better to other domains. Besides common dictionary words, it also gives information on emoticons, slang (nah, meh, giggly, etc.), and acronyms (LOL, OMG, etc.) (Hutto & Gilbert, 2014). I have chosen VADER in the present project because it is becoming broadly adopted as it was even implemented as a component of the NLTK python library. However, several limitations can be outlined in this study. Firstly, the introduced approach handles English tweets only although a wealth of information about thalassemia is adequately available in different languages. It is, therefore, highly recommended that this method is applied to other languages, e.g., Arabic. Secondly, the general opinion lexicon is inadequate to capture the meanings of health-related texts.