Mapping the “X” Debate: Water Fluoridation Sentiment Analysis With Advanced Machine Learning.

3 Results

3.1 Engagement Metrics Analyses

A total of 78,914 English-language tweets related to CWF were extracted from January 2014 to December 2023, with 72,309 tweets finally analyzed after data processing. The analysis revealed a gradual decline in the proportion of tweets related to CWF, with minor fluctuations reaching a peak in 2016–2017 and a moderate and substantial decline following 2018. Despite the decrease in tweet generation, engagement activities, such as retweets and favorites, exhibited significant variability (Figure 2). This implies that, despite the topic’s ongoing interest, engagement levels are primarily driven by specific tweets or events. In addition, most tweets had a length of 10 to 30 characters, with 15 characters being the most common tweet length. The frequency of tweets progressively decreases as the length of a sentence increases, with only a small number of tweets exceeding 80 characters (Figure 3).

3.2 Textual Sentiment Analyses

3.2.1 Sentiment Trends and Pattern

Of the tweets regarding CWF, 37.4% conveyed a negative emotion, 34.4% were positive, and 28.2% were neutral, indicating a relatively equal distribution of sentiments (Figure 4). Furthermore, reactions to water fluoridation-related events are reflected in the proportional variation of positive and negative sentiments. In 2016, 2018, and 2020, substantial changes in public sentiments were observed, suggesting that these events in the years significantly impacted public opinion. These annotations indicated a strong correlation between sentiment trends and historical events, emphasizing the impact of societal events on public emotions (Figure 5).

3.2.2 Word Sentiment Analysis

The presentation of the most favorable (positive) terms the users have used to describe CWF indicates its social acceptance. Words like “safe,” “prevent,” “healthy,” and “beneficial” were employed most frequently to underscore the health advantages of water fluoridation. Words like “essential” and “effective” emphasize its importance and efficiency. The users’ high regard and trust for the scientific foundation of fluoridation are indicated by the frequency with which the terms “approve,” “trust,” and “recommend” were used (Figure 6).

In contrast, examining unfavorable (negative) terms illustrates divergent viewpoints. Concerns regarding the safety of water fluoridation are raised by the phrases “toxic,” “harmful,” and “dangerous.” The terms “risk,” “side effects,” and “controversial” all serve to emphasize the ongoing discourse concerning potential adverse effects. The utilization of the terms “oppose,” “fight,” and “ban” implies that there is a concerted effort to oppose fluoridation policies. The use of terms such as “dispute,” “criticize,” and “skeptic” suggests that the narratives that advocate for fluoridation have been approached with caution by some users (Figure 7).

3.2.3 Word Co-Occurrence Network Analysis

Word co-occurrence network analysis investigated the relationships between keywords associated with water fluoridation in tweet discussions. The primary focus of the debate was on central hubs, including “fluoride,” “water,” and “health,” which have high node degrees, indicating they were highly connected to other keywords. “Safety issues” (including ‘risk,’ “safety,” and “toxicity”) and “dental health” (with phrases such as “teeth,” “dental,” and “toothpaste”) were notable clusters that indicated specific debate topics. “Infection” “filters,” and “filtration” indicated concerns regarding implementing safety measures and water purity. The correlations between “benefit,” “protection,” and “treatment” suggested the necessity of educational and lobbying initiatives (Figure 8).

3.2.4 Topic Modeling and Issue Tweet Analysis

The recurring topics in “X” conversations on CWF are illustrated in the tabulated list of prominent terms from 2014 to 2023 (Figure 9). The persistent concerns regarding the health effects and community acceptance are indicated by the consistent periodic recurrence of critical issues, including “health,” “community,” “dental,” and “decay.” The term “children” has frequently been used for years, suggesting concerns regarding children’s dental health. In 2023, the terms “toxic” and “safety” were used to underscore periods of elevated apprehension regarding the potential risks of fluoridation. In addition, specific years exhibit a concentration on subjects such as “benefits,” “research,” and “cavities,” suggesting a trend toward evaluating the scientific credibility and efficacy of fluoridation methods. The variations in public discourse regarding CWF over the years are underscored by these patterns.

3.2.5 Model Performance Evaluation

Table 1 summarizes the performance evaluation of various sentiment analysis models. The highest accuracy (88.8%) and AUC value (0.95) were exhibited by the Logistic Regression model, indicating its exceptional performance in accurately classifying sentiments. Naive Bayes exhibited promising results, obtaining an AUC value of 0.85 and an accuracy rate of 78.6%. The Random Forest algorithm exhibited a high AUC value of 0.90 and a similar level of accuracy at 77.3%, indicating its high and reliable performance. The Decision Tree model showed the lowest performance, with an AUC value of 0.75 and an accuracy of 77%. The results of this study suggest that the model is less reliable than the others (Table 1).

TABLE 1. Model performance and evaluation details.

Model	Accuracy (%)	Precision (%)	Recall (%)	F1 score (%)	AUC value
Logistic regression	88.83	88.87	88.83	88.82	0.95
Decision tree	77.09	77.14	77.09	77.07	0.75
Naive bayes	78.63	78.79	78.63	78.59	0.85
Random forest	77.31	78.80	77.31	77.12	0.90

• Note: This table summarizes the performance metrics of various sentiment analysis models used in the study. Metrics such as accuracy, precision, recall, F1 score, and area under the curve (AUC) are provided to evaluate each model’s effectiveness in classifying sentiments related to CWF discussions on “X.”

4 Discussion

Our sentiment analysis of community water fluoridation (CWF) on “X” over the past decade revealed a detailed understanding of public opinion. The study showed that 37.4% of tweets expressed negative sentiments toward CWF, compared to 34.4% positive and 28.2% neutral sentiments. The slight disparity between positive and negative attitudes suggests the subject is polarizing, with public opinion almost equally divided (Data S2).

We found significant variations in public involvement in CWF discussions, with surges in 2015–2016 and a more marked decrease after 2018. The data indicates that the public’s interest in CWF has slowly diminished, possibly overshadowed by other health-related debates. An investigation on media coverage of health matters revealed that health subjects frequently vie for attention, with more dramatic or urgent issues typically eclipsing others [22]. This phenomenon is particularly evident in talks around digital health, where prominent subjects such as COVID-19 and mental health have overwhelmingly dominated public conversations on social media platforms, reducing focus on topics like CWF [23]. Nevertheless, specific tweets or occurrences continue to generate substantial interaction, highlighting the significance of context in debates on social media.

We found that the higher spikes in negative sentiment often corresponded with significant CWF-related events. For example, in 2015, the U.S. Department of Health and Human Services updated its guidelines for optimal fluoride levels in drinking water, leading to a surge in tweets discussing the changes and their implications [24]. Similarly, in May 2013 and May 2016, the rejection of a proposal to add fluoride to the drinking water in Portland, Oregon, and Albuquerque, New Mexico, sparked a significant increase in Twitter activity [25].

Our findings align with previous research by Oh et al. [11], which reported a significant proportion of negative sentiments towards water fluoridation on Twitter. Furthermore, Mertz and Allukian [5, 6] highlighted the role of mis(dis)information in amplifying negative sentiments online. Rodrigues et al. [26] also found that these negative sentiments towards health interventions are more likely to become polarized, shared, and retweeted, contributing to mis(dis)information.

Our research emphasizes the importance of analyzing neutral sentiments, which can swing toward either positive or negative based on new information. Neutral perceptions, comprising 28.2% of the sentiments, represent a significant portion of public opinion. While we did not analyze this subset to determine whether it reflects ambivalence, lack of knowledge, or another sentiment type, this represents an area for future investigation. Understanding the underlying drivers of neutrality could provide deeper insights into public perceptions of CWF. For instance, neutral sentiments might indicate a lack of awareness or engagement with the topic, reflecting opportunities for targeted educational campaigns. Johnson et al. [25] highlighted that neutral opinions could shift dramatically with new information, potentially altering the balance of online views and influencing public discourse. Monitoring these neutral sentiments is essential for anticipating and managing changes in public opinion on CWF.

Our word sentiment analysis identified key positive terms like “safe,” “prevent,” and “healthy,” indicating support for the health benefits of CWF. Conversely, negative terms such as “toxic,” “harmful,” and “dangerous” reflect ongoing safety concerns. Similar to our results, Oh et al. observed that negative terms like “poison” and “waste” were frequently used, reflecting significant public concern over the safety and potential risks of fluoridation [11]. This consistency suggests that public skepticism toward CWF remains persistent despite varying methodologies and timeframes.

The co-occurrence network analysis highlighted central themes, particularly clusters around “safety issues” and “dental health.” This is consistent with findings by Mackert et al. [12], who observed that concerns about safety and potential side effects were prevalent in social media discussions on dental health interventions. Additionally, topic modeling revealed ongoing concerns about health effects and community acceptance of CWF, with significant attention to children’s dental health. We found that in recent years, there has been sustained apprehension about potential risks associated with CWF, which aligns with findings by Oh et al. [11], noting that safety and health impacts were dominant themes in Twitter discussions. Furthermore, Kanchan and Gaidhane [23] found that the perceived credibility of sources and the emotional tone of messages significantly impacted public engagement and sentiment.

Performance evaluation is a critical parameter for analyzing the accuracy of various sentiment analysis models. In our study, we selected Logistic Regression, Naive Bayes, Decision Tree, and Random Forest due to their proven effectiveness in text classification. Logistic Regression was chosen for its simplicity, interpretability, and ability to capture linear relationships in high-dimensional data [27, 28]. Naive Bayes is efficient for handling sparse data in natural language processing, despite its simplifying assumption of feature independence [29]. Random Forest’s ensemble approach reduces overfitting and effectively captures non-linear relationships [30]. In contrast, Decision Trees, though interpretable, are more prone to overfitting in high-dimensional contexts [31]. Overall, our experimental results showed that Logistic Regression outperformed the other models, aligning with previous research. This selection reinforces the relevance of these models in addressing sentiment analysis challenges.

Additionally, the recent Cochrane systematic review on water fluoridation (2024) highlights that while CWF remains effective in reducing dental caries, the magnitude of its effect may be lower than earlier estimates, in part due to additional fluoride sources now widely available [28]. Importantly, this review underscores the equity debate, noting that communities with less access to alternative fluoride interventions may still benefit substantially from CWF. This resonates with our findings, where neutral sentiments—representing a potential “swing” group—could be influenced by targeted messaging about both the protective and equitable aspects of fluoridation policies. Policymakers should thus consider not only the efficacy but also the fairness and reach of CWF initiatives when designing public health strategies.

Given our key finding that public sentiment towards CWF shows a nearly equal distribution of negative, positive, and neutral sentiments, we recommend developing comprehensive monitoring systems and strategies to manage public sentiment about CWF effectively. Engaging with the public, communicating transparently with health authorities, and using credible influencers can help balance the discussion. Real-time sentiment analysis and tracking key accounts can quickly spot and address new trends, reducing mis(dis)information and limiting negative sentiments from spreading [23, 26]. Additionally, developing reliable fact-checking systems specifically for CWF concerns is crucial. These systems should educate the public, address their concerns, and use automated tools to flag mis(dis)information. Educational campaigns that explain the benefits and safety of CWF, backed by scientific evidence, can help clear up misconceptions and build trust [11, 26].

This study has several limitations that warrant consideration. First, temporal and demographic biases may have influenced the dataset. Changes in Twitter’s user demographics over the years, as well as region-specific events, could have impacted the observed sentiment trends. While these biases were not explicitly controlled for, we addressed them partially by contextualizing significant sentiment spikes and linking them to major events, such as policy debates and media coverage. Future studies should integrate demographic and geographic metadata to mitigate these biases and enhance the precision of analytical insights. Second, the analysis was limited to English-language tweets, potentially excluding important discussions in other languages. This limitation may restrict the generalisability of the findings to non-English-speaking populations. Additionally, while automated sentiment analysis tools are efficient and effective for processing large datasets, they may not fully capture the nuance of human emotions and context. Incorporating manual validation alongside automated tools in future studies could improve the accuracy and reliability of sentiment classification. Third, the study’s timeframe is a potential limitation. Social media trends evolve rapidly, and analyzing a static time period may not fully capture the dynamic nature of public discourse. Expanding the study period or conducting longitudinal studies would provide deeper insights into how sentiments change over time and help identify emerging trends. Finally, this study focused solely on a single platform, “X” (formerly Twitter). While “X” provides valuable insights, expanding the analysis to include other social media platforms, such as Facebook and Instagram, would offer a more comprehensive view of public sentiment. Different platforms attract diverse user demographics and may host varied discussions on CWF. We recommend that future research explore multiple platforms to capture a broader spectrum of public opinion and enhance the generalizability of findings.

Moreover, while we examined tweet content and engagement, we did not investigate the profiles or credibility of the users posting these tweets—whether they were general users, key opinion leaders, bots, or institutional accounts. Distinguishing these user types and assessing their credibility could yield valuable scientometric insights, particularly regarding how polarizing narratives originate or are amplified. As our primary objective was to explore overall sentiment and topical trends surrounding CWF, future studies should integrate user-level analyses to better understand and address mis(dis)information.

Overall, the study highlights the evolving nature of public sentiment about CWF on “X” and emphasizes the importance of real-time monitoring to guide public health strategies. By using machine learning to analyze social media data, this research provides actionable insights for improving health communication and policymaking in today’s digital age. Practically, the findings suggest that public health policymakers should develop targeted interventions to address misinformation and engage neutral audiences. Strategies could include creating localized, culturally relevant educational campaigns and leveraging trusted community influencers to amplify accurate information. Establishing partnerships with social media platforms to monitor and flag misleading content is another actionable step. These efforts can help build trust in CWF and ensure informed decision-making within communities. While this study offers valuable information, more research on CWF-related sentiments across various social media platforms is needed to ensure a holistic understanding and more effective public health strategies.