We conducted a retrospective observational infodemiology study using annual TCIM-related mention counts from public content posted on X [18]. Word- and phrase-level time series from X have been used in previous health research and infodemiology studies [19,20]. We analyzed aggregated annual mention counts for public X posts containing at least one term from the study lexicon; no participants were recruited. The observation period ran from January 2015 to December 2024 and was used to document decade-long raw mention patterns, including a global pandemic episode (COVID-19 pandemic), during which TCIM-related discussions and misinformation on social media have been reported [2].

The lexicon aimed to cover TCIM-related mentions at 3 semantic levels: approaches, disciplines, and techniques, with remedy-related labels treated as part of the technique level when relevant. This strategy was intended to reduce omissions caused by differences in terminology across communities and discourse contexts [21]. The lexicon was compiled from three sources: (1) institutional references, including World Health Organization (WHO) reports on TCIM [22] and information sheets from the French Ministry of Health [23], to frame the scope of TCIM; (2) encyclopedic references from Wikipedia, which is often used in health-related text mining as a source of vocabulary and lexical variants [24]; and (3) triangulation with 3 generative AI tools, DeepSeek, ChatGPT (OpenAI), and Microsoft Copilot, used as controlled brainstorming tools to propose candidate terms. The use of generative models to assist in refining search strategies and search terms has been discussed in recent literature [25,26]. Triangulation was used to improve lexical coverage and reduce dependence on a single source [27].

The exact prompt used for this step is provided in Textbox S1 in Multimedia Appendix 1. It was executed independently for each AI tool on April 21, 2025. The prompt focused on terms likely to be written, searched for, used as hashtags, or discussed on X (formerly Twitter). It prioritized terms that were widely recognized in public discourse, likely to generate substantial mention volumes, and relevant to TCIM. It excluded names of people, brands, clinics, commercial products, and biomedical treatments not generally addressed within the TCIM field. The prompt also specified that AI-generated outputs should not be interpreted as evidence of clinical effectiveness, public health relevance, frequency of use, or real-world use. The AI-assisted step identified 104 candidate TCIM-related terms in at least one AI tool, including 39 terms identified by all 3 AI tools (Table S1 in Multimedia Appendix 1). AI outputs were used only to generate candidate terms; final inclusion in the main lexicon required review by the authors and consistency with the study scope.

In total, the study retained 39 English seed terms: 10 (25.6%) for approaches, 16 (41%) for disciplines, and 13 (33.3%) for techniques. The retained terms are “yoga,” “meditation,” “mindfulness,” “ayurveda,” “essential oils,” “homeopathy,” “reiki,” “chiropractic,” “tai chi,” “acupuncture,” “intermittent fasting,” “traditional medicine,” “hypnotherapy,” “music therapy,” “ketogenic diet,” “kinesiology,” “reflexology,” “art therapy,” “acupressure,” “phytotherapy,” “qi gong,” “shiatsu,” “traditional Chinese medicine,” “osteopathy,” “naturopathy,” “shamanism,” “energy healing,” “EMDR (eye movement desensitization and reprocessing),” “sophrology,” “integrative medicine,” “forest bathing,” “neurofeedback,” “energy medicine,” “complementary medicine,” “St John’s wort,” “moxibustion,” “thermal therapy,” “Buteyko,” “anthroposophic medicine.”

These English seed terms were translated into Spanish and French to reduce Anglocentric bias and improve coverage across 3 widely represented languages on X [17], without assuming perfect semantic equivalence across languages. Relevant lexical variants were included when applicable. The 39 English terms, their Spanish and French translations, and the lexical variants used in the searches are provided in Table S2 in Multimedia Appendix 1.

Data were obtained from Brandwatch Consumer Research (Cision Ltd), a social listening platform that supports historical queries and aggregated exports. Brandwatch was selected because it provides large-scale historical coverage, standardized queries, and has been used in previous academic work on social data mining in health [28,29]. Compared with partial API feeds, provider-based access can reduce some coverage constraints associated with sampled data streams [30]. In Brandwatch terminology, a “mention” corresponds to a content item returned by a query. According to Brandwatch documentation, Brandwatch is an official X partner and indexes public posts as mentions, including reposts, retweets, and quote posts [31]. The study team received an Excel spreadsheet containing aggregated annual mention counts. The primary analytic variable was the raw, nonnormalized annual mention count for each term, stratified by language: English, Spanish, and French. Language classification was provided in the Brandwatch export and was not manually assigned post hoc by the study team. Because only aggregate exports were available, postlevel validation of language classification, including code-switching errors, was not possible. No annual denominator, such as total X activity, language-specific posting volume, or active users by language, was available.

This study analyzed provider-generated aggregate counts derived from publicly available X content. The study team did not access individual-level posts, usernames, profile information, or other directly identifiable personal data. No users were contacted, recruited, or interacted with. Because the analyses were restricted to nonidentifiable aggregate data by year, term, and language, informed consent was not required. Data handling and reporting followed applicable ethical and privacy principles for internet research [32].

All analyses were based on raw, nonnormalized annual mention counts. Descriptive analyses were conducted in 2 complementary ways. First, for each language, mention counts were aggregated across the 2015 to 2024 period to summarize the overall distribution of TCIM-related terms within each language-specific corpus. Second, annual mention counts were aggregated across English, Spanish, and French to examine temporal patterns in the predefined TCIM-related corpus. These descriptive results were reported using tables and figures, in accordance with infodemiology approaches to time-series reporting [18].

Exploratory trend analyses were restricted to the 2015 to 2019 prepandemic window to avoid fitting a single linear trend across the full 2015 to 2024 period, which included the COVID-19 pandemic period and may have involved a structural break [33]. The complete 2015 to 2024 series was retained for descriptive reporting. For each of the 39 terms within each language, an ordinary least squares linear regression was fitted with annual mention count as the dependent variable and calendar year as the independent variable. Because only 5 annual observations were available and multiple term-language comparisons were performed, these analyses were treated as exploratory and hypothesis-generating rather than confirmatory.

Two prespecified filters were applied to identify exploratory linear trend signals. First, goodness of linear fit was assessed using the coefficient of determination (R²) and only terms with R²≥0.70 were retained. Second, the slope was evaluated using an exact permutation test, enumerating all 5!=120 permutations of the 5 annual values across years. The 2-sided P value was calculated as the proportion of permuted slopes at least as extreme as the observed slope. Terms were retained as showing an exploratory linear trend signal when both criteria were met: R²≥0.70 and P≤.05. No multiplicity adjustment was applied because these analyses were designed to identify exploratory signals rather than provide confirmatory evidence.

For retained terms, changes from 2015 to 2019 were summarized using the estimated slope and relative change. Uncertainty around 2015 to 2019 relative changes was summarized using 95% bootstrap CIs obtained by nonparametric resampling with replacement and the percentile method, using the 2.5th and 97.5th percentiles [34]. Bootstrap CIs are reported in Figures S1-S3 in Multimedia Appendix 3.

Across the 2015 to 2024 period, the raw, nonnormalized mention counts did not show a sustained aggregate increase in the predefined TCIM-related corpus on X. No sustained aggregate surge was evident in the raw descriptive series during the COVID-19 pandemic period, although term-level fluctuations were observed. This may partly reflect informational saturation during the pandemic [35]. Prominent topics such as vaccines, lockdowns, and pandemic-related policies dominated online health discussions on X [36,37]. However, the analysis could not assess the effect of platform-level changes on indexed mention volumes, including the 2023 rebranding and changes in governance or visibility algorithms. Year-on-year variations were mixed. Between 2019 and 2020, meditation showed the largest absolute increase in raw mention counts, rising from 3,627,557 to 5,563,506 mentions (1,935,949/3,627,557, 53.4%). Ayurveda showed the largest relative increase, rising from 214,901 to 615,525 mentions (400,624/214,901, 186.4%; Table S3 in Multimedia Appendix 2). These changes should be interpreted as isolated term-level fluctuations rather than evidence of a broader TCIM surge. They may be compatible with broader wellness and stress management vocabularies circulating online. However, the data do not allow inference about motivations, use, therapeutic intent, or endorsement. Similarly, the decrease in yoga mentions between 2015 and 2019 may reflect several nonmutually exclusive mechanisms, including lexical diversification, changes in platform activity, or shifts toward other platforms [38,39]. These mechanisms could not be tested in this study.

Over the decade, disciplines and techniques were more frequently mentioned than broad approaches within the predefined lexicon. This finding should be interpreted descriptively because the semantic groupings were heterogeneous and contained different modalities of practice: 10 approaches, 16 disciplines, and 13 techniques. High-volume labels such as yoga, meditation, and mindfulness are also culturally broad wellness terms that may dominate aggregate counts regardless of category logic. Therefore, this result should not be interpreted as evidence that disciplines or techniques are generally more visible than broad approaches across TCIM discourse overall. Rather, it indicates that, within this specific lexicon and Brandwatch-indexed X corpus, practice and wellness labels were more frequently captured than broad umbrella labels. From a monitoring perspective, this pattern suggests that a harmonized classification of evidence-based TCIM programs could improve their recognition by digital tools and their traceability within health and prevention organizations [40,41]. This aligns with a global health policy context aimed at strengthening evidence, quality, safety, and access to TCIM [42].

The predominance of English in TCIM-related mentions (170,100,867/191,994,848, 88.6%), compared with Spanish (17,171,541/191,994,848, 8.9%) and French (4,722,440/191,994,848, 2.5%), is consistent with previous literature [14]. These cross-language differences may reflect variations in platform activity, public policies, and the influence of country-specific traditional medicines [21], but these mechanisms were not directly tested. Consequently, mention frequency in each language should not be interpreted as evidence of the views, attitudes, or practices of the corresponding population. Language should be considered a proxy for linguistic space, not a direct proxy for culture, nationality, or comparable discourse communities. Translated terms may also differ in meaning and use across languages.

A central issue for interpretation is the heterogeneity of the TCIM umbrella concept. The predefined lexicon included broad systems of medicine, lifestyle and wellness labels, body-mind practices, diet-related terms, clinically framed interventions, and contested modalities. Therefore, the aggregate corpus should not be interpreted as a homogeneous measure of TCIM discourse. The concentration of 80.3% (154,093,932/191,994,848) of cumulative mentions in yoga, meditation, and mindfulness indicates the dominance of wellness and body-mind vocabularies within the selected lexicon. It does not imply that all TCIM domains were equally visible or that these high-volume terms were always used in explicitly medical, therapeutic, or TCIM-related contexts. In a geopolitical period marked by war, infectious disease, misinformation, climate change, occupational pressure, and resource constraints, this predominance may align with broader online wellness and self-care vocabularies, particularly those related to stress management and mindfulness [7]. More broadly, these patterns may be discussed in relation to online interest in nonpharmacologic or complementary options and concerns about trust in pharmaceutical manufacturers, although existing evidence remains population- and context-specific, including cardiovascular-risk populations [43]. However, the present data cannot show that users turned toward approaches perceived as more “natural,” or whether posts reflected mistrust, criticism, or patient empowerment. These findings suggest that social media constitute an important space for tracking the circulation of TCIM-related health vocabularies [20]. Further qualitative, sentiment, and user-level analyses would be needed to examine therapeutic preferences, stress management narratives, and broader relationships to conventional medicine.

The study measured the frequency of predefined TCIM-related mentions on X, not actual practice use, sentiment, or user intent. Mentions may reflect positive, neutral, or negative opinions, as well as promotion, criticism, controversy, humor, media coverage, advertising, or reposting. Without sentiment, contextual, or user-level analyses, changes in volume cannot be interpreted as evidence of endorsement, rejection, or real-world trends in TCIM practices. This limitation is widely recognized in Twitter-based public health research, where representativeness biases, platform dynamics, and context collapse can affect conclusions [19].

In addition, the counts were raw and nonnormalized. We did not have access to annual denominators, such as total X activity, language-specific posting volume, or active users by language. Therefore, the observed trajectories cannot be interpreted as normalized changes in public interest. Apparent stability in raw mention volumes may also be shaped by algorithmic visibility dynamics, platform rebranding, moderation changes, or shifts in user activity rather than by actual levels of public interest. Relevant discussions may also have migrated to other platforms, such as TikTok (ByteDance), Instagram (Meta Platforms), Reddit, or health-specific online communities. The growing role of “influencers” in the dissemination of TCIM content should also be considered [44].

Because Brandwatch indexes reposts, retweets, and quote posts as mentions, high-volume terms may partly reflect recirculation or amplification of a limited number of posts rather than independent contributions from many users. Moreover, we could not determine whether the users behind the mentions were consumers, patients, caregivers, health professionals, advocates, opponents, organizations, automated accounts, or bots.

The heterogeneity of the TCIM umbrella construct is also a limitation of the aggregate analysis. Because the 39 retained terms refer to different types of practices and vocabularies, they should not be interpreted as equivalent units of TCIM discourse. The lexicon was intentionally restricted to 39 commonly recognized terms. This approach improves feasibility and transparency but does not capture the full anthropological, cultural, doctrinal, and lexical diversity of TCIM-related discourse.

Future studies should examine predefined subcategories and broader lexicons to better distinguish wellness, clinical, traditional, and contested TCIM-related vocabularies.

The multilingual comparison also has limitations. Language was used as a proxy for linguistic space, but it cannot be equated with culture, nationality, or comparable discourse communities. Although lexical variants were considered for selected terms when relevant, translated terms may still not map perfectly onto equivalent semantic uses across English, Spanish, and French.

The exploratory trend analyses were also limited by the small number of annual observations in the 2015 to 2019 window. Even with the exact permutation test, a single year could influence the slope, R², and whether a term-language combination met the predefined exploratory criteria. In addition, no multiplicity adjustment was applied across term-language comparisons. These results should therefore be interpreted as hypothesis-generating exploratory trend signals rather than confirmatory evidence of linear change.

Finally, the study reflects only one social media platform. The observed mention patterns may therefore be specific to X and should not be generalized to other digital spaces where TCIM-related discussions may follow different visibility, audience, and dissemination dynamics [45].

Over the past decade, social media have become an important space for health-related information and discussion. Between 2015 and 2024, raw TCIM-related mention counts on X were highly concentrated in a small set of wellness and body-mind practice labels and were predominantly indexed in English compared with Spanish and French. No sustained aggregate surge was evident in the raw descriptive data during the COVID-19 pandemic period, although specific terms showed short-term fluctuations. These findings should therefore be interpreted as platform-specific raw mention patterns, not as evidence of normalized public interest, sentiment, attitudes, professional practice, or actual TCIM use. Monitoring health-related vocabularies on social media may help inform strategies to improve the dissemination of accurate health information, mitigate misinformation, and strengthen health communication among public and scientific communities [46].