This study did not require institutional review board approval from the Council of Ethics in Human Research of the University of Waterloo because federal regulations do not apply to research using publicly available data that do not involve human subjects.

On June 5, 2022, the Twitter application programming interface [29] was used to extract 23,436 tweets published in English between May 2016 and May 2022 with the keyword “fluoride-free.” It followed the parameters of a prior study about fluoride-related misinformation on Instagram [16].

The CSV file data set was uploaded and processed to enhance the quality of the tweets for subsequent analysis [30]. Initially, duplicate tweets were identified and eliminated, resulting in a data set containing 21,169 unique tweets. Next, tokenization was performed to break down the text into smaller units, such as individual words. Then, all tokens were converted to lowercase to ensure uniform treatment, irrespective of capitalization, and to prevent potential discrepancies in the analysis. Stop words were removed to eliminate common words that contribute little meaning to the analysis. This included articles (eg, “a,” “an,” and “the”), prepositions (eg, “on,” “in,” and “at”), and conjunctions (eg, “and,” “but,” and “or”). Additionally, unspecific or irrelevant words were removed, such as non-English characters, keywords, verbs, and digits. Lastly, punctuation marks, symbols, and special characters were eliminated to refine the data set further and ensure consistency in subsequent analysis.

Furthermore, exploratory data preprocessing was performed to assess the impact of both the most frequent and rare words in the data set. Notably, term frequency filtering did not exhibit substantial improvements on latent Dirichlet allocation (LDA) topic modeling. Therefore, the decision was made to retain the most frequent and rare words to enhance the contextual relevance of the resulting topics (Multimedia Appendix 1).

LDA topic modeling is an unsupervised machine learning probabilistic and word count–based algorithm used to effectively determine salient terms and extract emerging groups (topics) of a particular corpus (data set), including social media posts concerning health issues [31,32]. Even though this analysis does not provide a complete meaning of social media posts, it contributes to a good overview of their main issues [33]. Notably, a detailed description of the LDA model is provided elsewhere [34].

The preprocessed 21,169 tweets were subjected to LDA topic modeling techniques using the Python-based Gensim and spaCy libraries in a Google Colaboratory interface to identify the salient terms and topics [31,32]. The number of ideal topics was defined according to the coherence score proposed by Nikolenko and colleagues [35].

In this context, the coherence values were computed for K topics, where K ranged from 2 to 50, before eventually narrowing the consideration range down to 3 to 15 cases. Remarkably, higher coherence scores probabilistically represent better-quality topic modeling, facilitating the interpretation of outputs. An investigator (ML) validated the topic model with the highest coherence score by rating the model on the meaningfulness of its topics, how well publications within a single topic are related, and how publications of different topics are distinct.

Facing the definition of the number of topics, the similarity between them was calculated through an intertopic distance map. It is a visualization of the issues in a bidimensional space, where the bubbles represent topics that emerged from the topic modeling analysis and are dimensionally proportional to the number of contained words. In this graph representation, the axes are based on 2 principal components to facilitate the interpretation of the distance between topics. Synthetically, the LDA algorithm uses a principal component analysis to reduce N-dimensional vectors to bidimensional vectors and achieve (x,y) coordinates.

Moreover, the same investigator (ML) manually assessed a sample of tweets depicting each of the most representative word groups that determined specific issues. Lastly, additional data visualization was performed regarding the total count of each topic of fluoride-free record and its relevance over time, using Elastic Stack software through the Elasticsearch and Kibana features [36-38].

The purpose of this study was to analyze the fluoride-free content on Twitter. Although studies have investigated fluoride-related content on Facebook, Instagram, and Twitter [16,28,39-42], to the best of our knowledge, this is the first study that focused on fluoride-free content on Twitter using topic modeling methods. The findings demonstrated that fluoride-free content found on Twitter mainly concerned a perceived healthy lifestyle, followed by the consumption of natural and organic oral care products and recommendations of fluoride-free products and measures. The outcomes are in accordance with previous studies [16,28,39-42]. We found that the interest in falsehoods gradually decreased between 2016 and 2019 and subsequently rose again from 2020 onward. The use of topic modeling analysis enabled the examination of a large Twitter data set, which would be difficult to achieve with traditional human-based analysis. In this sense, developing artificial intelligence–grounded systems capable of detecting and analyzing a massive volume of fluoride-related misinformation is essential.

Despite a decline in “fluoride-free” tweets between 2016 and 2019, the antifluoridation movement has had a negative impact on dental public policies in English-speaking countries. This has led to some regions discontinuing community water fluoridation to meet specific demands of users, despite community-based intervention being a key factor in promoting oral health equity [43]. For example, in Canada, only approximately 39% of the population has access to fluoridated tap water, with uneven distribution across the country (71% living in Ontario vs 1.7% living in British Columbia) [44]. Research has shown that dental caries prevalence is significantly higher in areas where fluoridation has been discontinued. An observational study conducted on schoolchildren in Alberta found that dental caries prevalence in primary dentition (ie, deciduous teeth and baby teeth) was significantly higher in Calgary (64.8%; fluoridation cessation in 2011) compared to Edmonton (55.1%; still fluoridated) [45]. In contrast, oral care companies have developed fluoride-free products to meet consumers’ preferences [16,46]. Interestingly, the increase in “fluoride-free” tweets in 2020 coincided with the onset of the COVID-19 pandemic. The infodemic scenario during the pandemic has led to increased health information-seeking behavior among users, including dental-related issues [47,48].

False health beliefs usually originate from the same mechanisms of accurate health beliefs, influenced by diverse cognitive, social, and affective drivers such as intuitive thinking (the absence of analytical thinking or deliberation), cognitive failures (neglecting or forgetting sources and evidence), illusory truth (familiarity, fluency, and cohesion), reliance on source cues (elites, in-group, and attractive), emotion (emotive content and emotional state), and worldview (personal views and partisanship) [49,50]. Notably, users tend to uncritically interact with posts associated with their health beliefs, reinforcing the genesis of echo chambers on social media [51,52]. This scenario propitiates the emergence of denialist groups, characterized by their high level of engagement with health misinformation publications. In this context, anti–community water fluoridation groups’ digital activity exceeds those of their counterparts by 60 times [40]. In addition, developing health beliefs have a significant association with lifestyle and impacts on its changes [53], which could explain the high demand for this topic in our study.

The findings may contribute to elaborating educational public health actions regarding eHealth and mobile health solutions, aiding communities in avoiding antifluoridation content on social media. Likewise, these results can orient fact-checking agencies to promptly clarify deceitful web-based publications about fluoride to their audience, addressing the concerns of users to support fluoride-containing measures. For example, issues concerning fluoride overexposure must be clarified, reinforcing safeguard actions for storing and using fluoridated products in early childhood [19,25]. Certainly, the adoption of reliable web-based health information by users can enhance individual and shared decision-making, thus contributing to better health care outcomes [54,55]. In this context, government agencies and organizations should be aware of the importance of investments in measures designed to improve the acquisition of information by the population [56,57]. Moreover, artificial intelligence–based systems may support public health actions to mitigate the propagation of false fluoride content on the web. Specifically, these solutions can automatically label publications and down-promote them until an in-depth investigation by social media companies. With insight into the underlying social media issues surrounding antifluoridation content, health professionals can ensure adequate communication with their patients based on person-centered care principles [58]. Equally important, legislators must discuss resolutions to control the spread of misinformation, specifically the role and duties of users and social media managers. Since the antifluoridation movement is substantially present in contemporary society, fluoride refusal will frequently be reported in dental offices [28].

This study has some limitations. First, we used only data from Twitter for the proposed analyses. As a result, the findings are influenced by platform characteristics (eg, limitation in the word count of posts), differentiating them from other social media. Despite that limitation, previous studies also focused on Twitter data to investigate health topics on social media because of the advantages relating to data collection and the number of active users [13,59]. Second, we collected the Twitter data set from a specific keyword, which may have limited the extrapolation of outcomes to all antifluoridation content. Nevertheless, we defined the search strategy and collected the data based on a previous fluoride misinformation study [16]. Third, although the results indicate the need to control information disorder on Twitter more effectively, it is improbable that the company will modify its policies in this area due to recent changes in ownership. Fourth, trend analysis did not consider the influence of possible platform-related confounders, such as the increase or decrease in the overall number of tweets on Twitter or changes in the content moderation policy over time. However, trend patterns are likely more related to users’ interest in fluoride-free content than Twitter’s characteristics. Lastly, these findings were based only on English-language tweets, influencing the screening of fluoride-free content by cultural and regional aspects. In contrast, falsehoods tend to be found in English since it is the most spoken language in the world.

The public concerns toward a healthy lifestyle, including the adoption of natural and organic oral care products, seem to be the main motivation of the recent increase of “fluoride-free” tweets, which can be boosted by the propagation of fluoride falsehoods noted on the web. Therefore, public health authorities, health professionals, and legislators should be aware of the spread of fluoride-free content on social media to create and implement strategies against their potential health damage for the population.