Health information dynamics and audience responses to avian influenza on YouTube: Implications for digital public health communication

Abstract

Effective management of infectious disease relies not only on biomedical interventions but also on the circulation and interpretation of health-related information. Narratives, whether accurate or misleading, influence public perceptions, adherence to protective measures, and policy debates, as illustrated during COVID-19, measles resurgences, and vaccine controversies. YouTube, as a global platform combining an information repository and interactive discussion space, exerts persuasive influence through its audiovisual formats. Avian influenza, a recurrent zoonotic threat with episodic media attention, provides a valuable case for analyzing evolving health narratives. We analyzed 11,465 YouTube videos and associated comments related to avian influenza published between 2020 and 2025. Thematic structures were identified using an MPNet-based topic modeling approach, and sentiment analysis was conducted with XLM-RoBERTa to assess emotional tone across content and audience interactions. Temporal patterns of narrative clusters were also examined. Video content predominantly conveyed factual health information aligned with epidemiological updates and preventive measures. In contrast, user comments displayed substantial interpretive activity, including contestation, speculation, and the adoption of controversial frames concerning virus origins, institutional trust, public health interventions, and alternative treatments. These discussions were characterized by predominantly negative sentiment, indicating that audiences frequently recontextualize online information, potentially fostering ambiguity and misinformation. YouTube serves as both a disseminator of factual health information and a forum for audience reinterpretation, highlighting the platform's dual role in digital health communication. Understanding these dynamics is crucial for public health authorities seeking to design interventions that mitigate misinformation and promote effective protective behaviors during emerging infectious disease threats.

Significance statement.

Online platforms shape how people learn about new health risks. In this study, we use avian influenza as an example to show how YouTube discussions go beyond science and animal health. They also include politics, food prices, media stories, personal beliefs, and distrust of institutions. We found that videos often share general health information, while comments often add doubt, fear, blame, or misleading claims. This matters because these online discussions can affect how people understand disease risk and whether they trust public health advice. By studying these discussions at a large scale, our work shows why public health groups need to follow online health conversations and respond with clear, timely, and trustworthy communication.

Introduction

Emerging infectious diseases continue to represent a major global public health concern, requiring public intervention to mitigate them. The H1N1 influenza pandemic, Ebola virus disease, and COVID-19 have demonstrated the vulnerability of health systems to novel and reemerging pathogens (1), particularly zoonotic agents capable of crossing species barriers (2). Alongside these biological threats, the rapid expansion of social media platforms has introduced new challenges for infectious disease prevention and control. Digital platforms have profoundly transformed how health information is produced, accessed, and interpreted (3). A growing proportion of the population now relies on online sources to obtain health-related information. Platforms such as Facebook, Instagram, X (formerly Twitter), TikTok, and YouTube host extensive health content produced by both expert and nonexpert users (4). While these platforms offer opportunities for health promotion and the dissemination of public health recommendations, they also raise essential questions about the quality and influence of health-related information on health-related behaviors, particularly during major health crises.

Indeed, during health emergencies, social medias create an overabundance of information, more commonly referred to as an “infodemic” (5). The World Health Organization defines this phenomenon as an excess of information, including false or misleading content, circulating in physical and digital environments during a health emergency (6). False information without intent to deceive is called misinformation, whereas deliberate misleading content is disinformation (6). Evidence accumulated during the COVID-19 pandemic demonstrated how the participatory and decentralized nature of social media enables the rapid spread of rumors and conspiracy narratives, contributing to reduced adherence to public health measures and increased social polarization (7). During this crisis, while authorities promoted mask wearing, vaccination, and physical distancing, online misinformation and disinformation generated doubt regarding vaccine safety and the existence of the virus itself (7). Repeated exposure to health misinformation has been associated with impaired health literacy, reinforcement of false beliefs, and decreased trust in public institutions, ultimately compromising the effectiveness of public health interventions (4). These dynamics challenge traditional models of health communication and highlight the need for systematic approaches to analyze and understand online health-related discourse at scale.

Among social media platforms, YouTube plays a particularly influential role in shaping public perceptions of health risks (8). Its audiovisual format and algorithmic recommendation systems promote the repeated exposure of users to specific narratives, amplifying both authoritative information and anecdotal or erroneous claims (9). Previous research has documented the widespread circulation of health-related misinformation on YouTube (10), positioning the platform as a critical space for studying the construction and dissemination of health narratives. Despite growing interest in online misinformation surrounding human infectious diseases such as COVID-19, Ebola, and Zika (11), considerably less attention has been paid to zoonotic animal diseases with pandemic potential. Avian influenza represents a particularly salient case, given its recent geographic expansion, the increasing detection of infections in mammals since 2021, and persistent uncertainties surrounding animal-to-human transmission (12). Control of avian influenza relies on complex surveillance and mitigation strategies at the animal–human interface (12), where misinformation may undermine risk perception, biosecurity measures, and public health preparedness. Beyond its direct health implications, avian influenza has also generated substantial economic and societal impacts (12), further intensifying public attention and online discourse.

In this study, we examine how avian influenza is framed and discussed on YouTube using a natural language processing-based analytical framework. By applying topic modeling and sentiment analysis to large-scale video content and comments, we aim to identify dominant narratives, detect potentially misleading information, and explore how online discourse may shape public risk perceptions. This work contributes to a better understanding of how digital platforms influence the social construction of emerging infectious disease threats and provides evidence to inform more effective public health communication strategies in digital environments.

Results

Data overview

After six iterative refinement steps designed to exclude off-topic texts using our MPNet-based topic modeling procedure, as described in the Materials and methods section, analyses were conducted on a final corpus of 188,874 texts, comprising 10,855 videos and 143,385 comments. Video transcripts originated from 6,495 unique channels, and comments were posted by 114,617 distinct users. A complementary analysis of user contribution patterns indicated that comment participation was highly dispersed. Among the comments retained for analysis, n = 76,087, 65,412 unique users were identified, with a median of one comment per user. Most users, 89.2%, posted only a single comment, while the top 1% of users accounted for 4.8% of comments and the top 10 users for <0.5%. Recurrent participation across videos was limited, with 92.6% of users commenting on only one video. We analyze the temporal distribution of videos and comments (Fig. 1). Temporal variations in online activity broadly mirrored major phases of H5N1 epidemiological dynamics. From 2020 onwards, a sustained increase in both video production and user engagement was observed, with pronounced peaks in 2024–2025, coinciding with intensified epizootic activity and expanded cross-species transmission (13, 14). Detailed correspondence with epidemiological events is provided in Table S1.

Figure 1.

Distribution of the number of videos and comments per month for the period 2020–2025 and association with H5N1 epidemiological events.

Topic analysis

Topic overview

Topic modeling identified 36 topics across the corpus, of which one cluster (topic −1) captured off-topic or semantically diffuse content (15). Topic −1 accounted for 52% of the corpus. Its lexical profile substantially overlapped with that of topic 0, including general avian influenza–related terms as well as conversational markers (e.g. “uh,” “um,” “know,” and “like”). However, these segments did not exhibit sufficient internal coherence to support a stable thematic grouping and were therefore not retained for subsequent analyses (see Materials and methods). The remaining 35 topics were grouped into 12 thematic categories using hierarchical clustering (Fig. 2). The complete list of 35 extracted topics, including their representative keywords, relative sizes, and representative texts, is provided in Table S2. Among the 12 macro-categories defined manually (Table 1), “Avian influenza epidemiology” dominated the corpus, representing 66% of retained texts (n = 60,238). The predominance of the “Avian influenza epidemiology” category was largely driven by topic 0, the largest retained cluster. While sharing a similar general lexical profile with topic −1, topic 0 displayed sufficient internal consistency to be retained as a cluster, while remaining broad and low specificity. Its lexical profile reflects a general layer of epidemiological discourse, encompassing overlapping dimensions such as outbreak reporting, human and zoonotic risk, agricultural contexts, and animal–host dynamics, rather than a clearly delineated thematic category. The lexical diagnostic is provided in Tables S3 and S4. Other prominent categories included “Political and media discourses” (7%), “Geopolitical and macroeconomic dynamics” (6%), “Food system safety and security” (4%), and “Religious and spiritual interpretations” (4%). Smaller but recurrent categories addressed “Public health prevention debates” (3%), “Institutional trust and scientific legitimacy” (2%), “Diagnostic validity and PCR methodology” (2%), and “Biosafety and virus origin hypotheses,” “Alternative therapeutic practices,” and “Legal, criminal, and international affairs” (1%).

Figure 2.

Hierarchical clustering used to group topics into thematic categories. Hierarchical clustering was performed on the topics extracted from the corpus to organize them into nested clusters based on pairwise semantic similarity. The dendrogram represents the hierarchical relationships among topics, illustrating how they are grouped into broader thematic categories.

Table 1.

Manually defined thematic categories (meta-topics) summarizing groups of topics, ranked from top to bottom by decreasing number of documents.

Theme	Topics included	Representative keywords	Definition	Number of documents
Avian influenza epidemiology	0, 5, 22	Virus, bird, flu, human, influenza, egg, California, state, Florida, emergency, Texas, coop, run, wire, chicken, floor	Content addressing avian influenza transmission dynamics, clinical symptoms, host populations, and detected human and animal cases, as well as global epidemic episodes. This category also encompasses surveillance challenges, farm biosecurity measures, and the ecological interfaces shaping the spread of the virus.	60,238 (66%)
Political and media discourses	1, 12	Lie, Nate, Royce, Tim, Adam, play, game, trust, Kennedy, Trump, charge, worry, vaccine, Robert, fix, worm	Public discussions involving political actors, media figures, or mediated controversies, reflecting how health-related issues are framed, debated, or instrumentalized in political communication.	6,173 (7%)
Geopolitical and macroeconomic dynamics	3, 8, 9, 16	China, Chinese, CCP, world, Russia, Ukraine, war, Putin, price, gounging, inflation, raise, Biden, Joe, Trump, administration	Narratives linking global economic instability, interstate power relations, and armed conflicts, highlighting how geopolitical tensions shape perceptions of risk and influence global health and biosecurity concerns.	5,617 (6%)
Food system safety and security	4, 6	Cat, raw, food, pet, dog, feed, indoor, bird, milk, raw, drink, cow, pasteurize, dairy, pasteurization, pasteurized	Discourses addressing food production, animal feeding practices, and the sanitary management of animal-derived products, including virus transmission via the food chain (e.g. milk) and infections in companion animals through consumption of infected food products.	3,953 (4%)
Religious and spiritual interpretations	2	God, lord, Jesus, man, shall, Christ, bible, right, verse, pray, earth, word, bless, world, pestilence, day	References to religious beliefs, spiritual frameworks, and moral narratives used to interpret health crises, uncertainty, and societal challenges.	3,606 (4%)
Public health prevention debates	11, 15, 24, 31, 32	Mask, wear, face, masking, lockdown, lock, another, comply, Cuomo, Chris, Dave, Smith, bullshit, nonsense, bunch, worry, doom, worried, scary	Debates surrounding the rationale, legitimacy, and societal acceptance of preventive interventions, such as masking and lockdowns, highlighting tensions in risk communication and behavioral compliance.	2,357 (3%)
Institutional trust and scientific legitimacy	13, 19, 21, 23	Science, scientist, trust, scientific, pharma, Pfizer, big, FDA, CDC, trust, credibility, lie, MRNA, vaccine, technology, therapy	Discourses regarding the credibility of scientific institutions, regulatory agencies, and pharmaceutical industries, reflecting broader concerns regarding expertise, accountability, and transparency, particularly regarding vaccine development.	1,883 (2%)
Biosafety and virus origin hypotheses	14, 18, 28, 29	Gate, Bill, Gates, Fauci, gain, function, research, Fauci, weapon, bio, biological, warfare, lab, leak, release, intentional	Arguments and conjectures concerning laboratory biosafety, research oversight, and competing hypotheses about viral emergence and pathogen origin, including theories of laboratory escape and gain-of-function studies.	1,739 (2%)
Environmental and demographic risk narratives	17, 26, 27, 34	Population, depopulation, billion, million, plandemic, plan, another, next, radiation, poison, climate, warming, evil, world, planet, humanity	Discourses expressing concerns about environmental hazards, demographic decline, and societal destabilization, often framed through broader narratives of systemic or existential risk, including alarmist claims and theories of planned pandemics aimed at reducing the global population.	1,547 (2%)
Legal, criminal, and international affairs	20, 25, 30	Siti, Fahri, Fadilah, Indonesia, Supari, Indonesian, free, release, Fauci, pardon, Anthony, Biden, prison, jail, arrest, Trump	Discussions concerning legal processes, criminal allegations, governance issues, and international political controversies intersecting with public health narratives, particularly involving international public health officials and policymakers.	1,176 (1%)
Diagnostic validity and PCR methodology	10, 33	PCR, test, testing, positive, false, inventor, Mullis, method, testing, copy, asymptomatic, trust, fake, unreliable, case, accurate	Discussions on diagnostic tools, particularly PCR-based methods, focusing on their accuracy, interpretability, and role in public health surveillance, including debates on test reliability, detection capacity, and potential falsification of results.	1,124 (1%)
Alternative therapeutic practices	7	Silver, oil, ivermectin, vitamin, eye, sunlight, oregano, water, help, bleach, pink, garlic, immune, zinc, colloidal, leaf	Discussions of nonconventional health practices, natural remedies, and informal therapeutic strategies outside mainstream biomedical frameworks, specifically focusing on approaches claimed or used to prevent, alleviate, or treat infections associated with the virus.	971 (1%)

Thematic distribution in comments and videos

Thematic distributions differed substantially between videos and comments (Fig. 3 and Tables S3 and S4). Videos predominantly emphasized epidemiological, geopolitical, and food safety themes, whereas comments more frequently introduced issues related to institutional trust, scientific legitimacy, and virus origin hypotheses.

Figure 3.

Distribution of defined thematic categories in comments and videos. The distribution is based on the proportion of texts assigned to each category in the videos and comments.

To assess thematic correspondence between videos and user comments, we analyzed cosine similarity indices (see Materials and methods for details) presented in matrix form (Fig. 4). Several categories exhibited strong self-correspondence between videos and comments, most notably “Political and media discourses,” which showed a high cosine similarity (0.88). Likewise, “Religious and spiritual interpretations” category remained largely stable across videos and comments (0.87), as did “Alternative and therapeutic practices” (0.80) and “Institutional trust and scientific legitimacy” (0.79). Conversely, some thematic categories appear to be substantially reinterpreted in the comments section, as reflected by comparatively low cosine similarity scores. This pattern is particularly evident for “Biosafety and virus origin hypotheses” category (0.31) and “Legal, criminal, and international affairs” (0.55). Beyond self-correspondence, several high cross-category cosine similarity scores point to systematic thematic reinterpretations between videos and comments. For instance, the category “Diagnostic validity and PCR methodology” shows notable overlap with “Institutional trust and scientific legitimacy” (0.61). Similarly, “Legal, criminal, and international affairs” is closely associated with “Public health prevention debates” (0.66) and “Religious and spiritual interpretations” (0.58). The category “Avian influenza epidemiology” also appears to be reappropriated in an expanded sense within the comments, exhibiting links to “Institutional trust and scientific legitimacy” (0.56) and serving as a point of convergence for reinterpretations related to “Environmental and demographic risk narratives” (0.54).

Figure 4.

Cosine similarity matrix between videos and comments by thematic category. Each value in the matrix represents the cosine similarity between the mean semantic embedding of a video category and that of a comment category, reflecting the degree of thematic overlap. Higher values indicate stronger correspondence between the content of videos and the associated user comments, whereas lower values suggest potential thematic divergence.

Thematic evolution across time

We analyzed the temporal dynamics of the thematic categories in the videos (Fig. 5A) and in the comments (Fig. 5B). “Avian influenza epidemiology” category remained consistently present throughout the study period in both videos and comments. Food system safety and security displayed episodic peaks in comments and videos, notably in mid-2023 and early 2025, corresponding to reports of H5N1 infections in dairy cattle and associated food safety concerns (13, 14). The theme “Geopolitical and macroeconomic dynamics” exhibited several waves of attention in video content, notably in 2022–2024, corresponding to major geopolitical disruptions, including the war in Ukraine (16), the escalation of the Israel–Hamas conflict (17), and the US presidential election cycle (18). In comments, this theme emerged earlier and more irregularly, particularly during the economic disruptions of the COVID-19 pandemic (19). “Political and media discourses” showed intermittent prominence in videos but sustained and amplified engagement in comments, especially during periods of heightened political polarization. Peaks observed in 2020–2021 aligned with US electoral dynamics (20) and public interventions by political figures, such as Donal Trump (21), while renewed activity in 2023–2024 reflected both electoral developments and media-driven controversies. Several themes displayed temporally concentrated patterns. “Diagnostic validity and PCR methodology” was largely confined to the early pandemic period (mid-2020 to early-2021), coinciding with intense public debate over testing strategies (22). Finally, some themes that remained marginal in videos emerged more distinctly in comments. “Legal, criminal, and international affairs” peaked sharply in mid-2020, reflecting potential controversies surrounding the management of the COVID-19 response, particularly the role of public health officials, such as Anthony Fauci (23). Detailed temporal distributions for each thematic category are presented in Figs. S1 and S2.

Figure 5.

Evolution of themes by year (2020–2025). A) Based on the proportion of videos on avian influenza attributed to each topic created by thematic modeling. B) Based on the proportion of comments on avian influenza attributed to each topic created by thematic modeling. Values are expressed as monthly proportions within each source. Temporal patterns are interpreted descriptively and are not based on formal statistical trend testing.

Topic-based sentiment

Sentiment overview

Sentiment analysis revealed marked contrasts between videos and comments. Videos were predominantly neutral (76%), whereas comments were largely negative (67%). Temporal sentiment dynamics (Fig. S3) showed a consistent antagonistic pattern between negative and neutral sentiment in both videos and comments, with shifts between phases of heightened criticism and periods of more factual or descriptive discourse. Sentiment distributions differed markedly between videos and comments (χ² = 14,006.76, P < 0.001), with a substantial effect size (Cramér's V = 0.39), indicating substantial differences in sentiment composition.

Topic and sentiment

Marked contrasts emerged when sentiment was examined across thematic categories (Fig. 6). In user comments, most themes were predominantly negative, including “Geopolitical and macroeconomic dynamics” (86% negative) and “Avian influenza epidemiology” (68% negative), reflecting anxious or critical receptions of these issues. In contrast, “Religious and spiritual interpretations” exhibited the highest proportion of positive sentiment (40%), while “Alternative therapeutic practices” were largely neutral (71%). Subjectivity scores across comment themes remained consistently below 0.5, indicating that negative polarity was primarily associated with argumentative or interpretative engagement rather than emotional expressiveness. “Food system safety and security” showed slightly higher subjectivity (0.4), suggesting a more personalized or experiential framing. Video content, by comparison, remained predominantly neutral across all themes. For example, 79% of segments related to “Avian influenza epidemiology” were neutral, reinforcing their informational function. Nevertheless, certain categories displayed a comparatively more negative tone, notably “Geopolitical and macroeconomic dynamics” (70% negative) and “Religious and spiritual interpretations” (63% negative), indicating that controversial or value-laden topics may be framed less neutrally even in producer-generated content. Average subjectivity scores were generally higher in videos, particularly for “Legal, criminal, and international affairs” and “Environmental and demographic risk narratives,” indicating a more interpretative or speculative tone when addressing sensitive sociopolitical or environmental issues.

Figure 6.

Temporal distribution of sentiments according to thematic categories. A) Distribution of sentiments across thematic categories in videos. B) Distribution of sentiments across thematic categories in comments. Circle colors indicate sentiment polarity (red: negative, yellow: neutral, and green: positive), and circle sizes reflect the number of texts per category and time period. Temporal patterns are interpreted descriptively and are not based on formal statistical trend testing.

Temporal sentiment patterns were largely stable across themes in both videos and comments, with a limited number of notable deviations. In videos, “Political and media discourses” became progressively more negative in 2023–2024, suggesting increasingly critical or distrust-oriented framings. In comments, “Religious and spiritual interpretations” showed persistent oscillations between positive and negative sentiment, reflecting heterogeneous and fluctuating audience reactions. Similarly, “Food system safety and security” alternated between neutral and negative sentiment over time, consistent with episodic concern triggered by specific reporting or epidemiological events. Detailed values for polarity and subjectivity scores, as well as sentiment distributions by thematic category for both comments and videos, are presented in Figs. S4 and S5.

Discussion

Social media platforms have become central spaces for the circulation, interpretation, and contestation of health information, particularly during the emergence of health threats. In this context, YouTube has emerged as an important platform for seeking health information, reaching a wide audience and facilitating primarily horizontal, user-driven information flows (8, 9). This environment can amplify both accurate and misleading content, raising significant concerns about misinformation and its implications for public health (10). In the context of avian influenza, renewed global attention, driven by outbreaks in livestock, increasing reports of infections in mammals, sporadic human cases, and associated economic consequences, has emerged amid significant uncertainty regarding the evolution of the virus and interspecies transmission (13). By examining both video content and user comments on YouTube, this study provides insight into how an evolving zoonotic risk is presented, reinterpreted, and amplified within a widely accessible digital platform, highlighting the potential consequences of online information dynamics on public health preparedness and communication. Our findings show that the production and reception of avian influenza–related content are closely aligned with sanitary, socioeconomic, political, and media contexts. Temporal trends in videos and comments followed major epidemiological and economic events, particularly from 2023 onward, including outbreaks in US dairy farms, infections in companion animals linked to contaminated products, and the increase in human cases in 2025 among agricultural workers (14). As observed during the COVID-19 pandemic, this synchrony between real-world crises and online activity creates conditions conducive to uncertainty, questioning, and the emergence of alternative narratives (24).

Although epidemiological aspects dominate both videos and comments, our analysis reveals that user comments extend beyond informational content and integrate politicized, speculative, and interpretative framings. The recurrence of themes related to epidemiology, diagnostic testing, and political or media figures across both layers suggests that users partially adopt the framing proposed by content creators, consistent with previous research on narrative circulation in audiovisual media (25). However, users also actively reinterpret information by embedding it within broader frameworks involving institutional trust, scientific legitimacy, sociopolitical conflict, environmental risks, and hypotheses regarding viral origin. These dynamics closely mirror those observed during COVID-19, when uncertainty and scientific complexity facilitated conspiracy narratives and alternative explanations (26).

The temporal distribution of themes further highlights the role of external events in shaping online discourse. Geopolitical and macroeconomic narratives closely follow major developments such as the war in Ukraine (16), postpandemic economic instability (19), and US electoral cycles (18), which users mobilize as interpretive lenses for understanding avian influenza. References to political and media figures, including Donald Trump, Robert F. Kennedy Jr, and Anthony Fauci, illustrate how controversies from previous health crises persist and are repurposed as explanatory templates (27–29). Historical figures, such as Siti Fadilah Supari, associated with earlier H5N1 disputes also reemerge, indicating that contemporary narratives are often integrated into preexisting belief systems (30, 31).

Sentiment analysis reinforces these findings by showing that user comments are predominantly negative, particularly for controversial themes such as biosafety and viral origin. Narratives involving intentional laboratory leaks, viral manipulation, or influential actors, such as Bill Gates, frequently associated with conspiracy discourses are marked by high negative polarity (32). While video content remains largely neutral, the argumentative negativity observed in comments, combined with low subjectivity scores, suggests a discourse that emphasizes critique and suspicion rather than emotional reaction. Such patterns are known to promote engagement and amplification on social platforms (33). These dynamics have potential implications for public health. Previous research shows that perceived controversy and institutional distrust reduce adherence to recommendations and increase confusion (34). In the context of avian influenza, recurring challenges to the credibility of institutions such as the Centers for Disease Control and Prevention (CDC, United States of America), the Food and Drug Administration (FDA, USA), and the pharmaceutical industry may undermine acceptance of future preventive measures, including vaccination, as previously observed during the 2009 H1N1 pandemic (35). At the same time, the episodic prominence of religious and spiritual interpretations highlights the activation of nonscientific explanatory frameworks during health crises. While these narratives may offer meaning and support, they can also shape risk perception and behavioral responses, as documented during Ebola outbreaks (36) and COVID-19 (37).

Taken together, this study shows that avian influenza H5N1 is not discussed solely as a public health or veterinary issue on YouTube but functions as a focal point for political, scientific, economic, and ideological controversies. These controversies, largely expressed through negative and argumentative discourse, contribute to shaping collective risk perception and may indirectly influence health behaviors (37). They also create fertile ground for the emergence and amplification of misinformation and disinformation, which has been shown to distort understanding of health risks, reduce trust in public health authorities, and undermine adherence to evidence-based health guidance in epidemic contexts (38). Misleading and emotionally charged narratives on social platforms can delay appropriate risk mitigation actions, encourage inappropriate self-directed responses, and erode confidence in official messages, as observed across multiple infectious disease settings (38). These dynamics are of particular concern in the context of H5N1, where uncertainty about viral evolution, sporadic human cases, and the potential for zoonotic transmission can be exploited to propagate inaccurate or strategically manipulated narratives that hinder effective public health communication. Although social network analysis has proven useful for identifying user clusters, influential actors, and interaction patterns on social media platforms, its application was not appropriate in this study. User engagement was highly dispersed (39). The additional user-contribution analysis showed that engagement was highly dispersed: most users commented only once, recurrent participation across videos was limited, and no single user dominated comment activity at either the video or thematic level. These structural characteristics prevented the construction of a meaningful interaction network and suggest that the observed thematic patterns were not primarily driven by a small group of highly active identifiable accounts. However, this does not rule out coordinated inauthentic or malicious activity distributed across multiple accounts. A single actor or organization could technically use several YouTube accounts to post comments, which would not be detectable through repeated activity by the same user ID alone. The available comment-level metadata therefore allowed us to describe repeated participation by identifiable user IDs but not to infer whether multiple accounts may have been operated by the same actor or coordinated group.

Several limitations should be considered when interpreting these findings. Video identification relied exclusively on the YouTube Data API v3, which captures only publicly available content and does not ensure exhaustive coverage, potentially omitting relevant discussions. In addition, transcripts could not be extracted for all videos, limiting the completeness of the production layer and potentially underrepresenting certain topics or nuances. Furthermore, the available YouTube limited our ability to examine whether geographic origin or source type shaped the observed discussions. Commenter geolocation was unavailable, while channel-level location information, when present, was optional, self-declared, and inconsistently populated. It may therefore not reflect the actual geographic origin of the channel operator, production team, or intended audience, especially in misinformation contexts where channels may hide, omit, or misrepresent their origins. In addition, channels were not formally classified by source type, which prevented systematic comparisons between institutional, media, industry-related, individual, or influencer-produced content and their associated audience responses. A substantial proportion of text segments was assigned to the “−1” category in the MPNet-based topic modeling, reflecting unclustered content, a common issue in social media analyses that may obscure less frequent or subtle themes. Although model parameters were optimized, topic modeling inherently involves trade-offs between interpretability and coverage. Finally, while sentiment classification using XLM-RoBERTa showed strong performance, reliance on a single annotator for training data may introduce subjectivity; larger, multi-annotator datasets would improve robustness and reproducibility.

In conclusion, discourse on avian influenza H5N1 on YouTube unfolds within a complex and fragmented social media ecosystem, where scientific knowledge, institutional communication, and individual interpretations intersect. YouTube functions both as a conduit for authoritative information and as a vector for misinformation and disinformation, which can distort public understanding of H5N1, amplify perceived risks or complacency, and undermine adherence to preventive measures. By integrating misleading content into preexisting beliefs and polarized debates, social media platforms directly shape risk perception and challenge the effectiveness of institutional communication. Understanding these dynamics is critical for anticipating public responses to future zoonotic threats, designing targeted interventions to counter misinformation, and sustaining trust in health authorities within increasingly contested digital information environments.

Materials and methods

Data extraction

YouTube videos related to avian influenza were identified using the YouTube Data API v3 (40), accessible via the Google Cloud platform, between 2020 January 1 and 2025 June 30. A keyword-based search query was designed to identify relevant videos that reference avian influenza in their titles or descriptions. The search included the following terms: “avian influenza,” “bird flu,” “H5N1,” and “avian flu.” For each retrieved video, a set of metadata was systematically collected, comprising the unique video identifier (ID), title, description, channel name, publication date, number of views, and number of likes. Public comments associated with each video were extracted through the same API. Video transcripts were obtained through the Innertube API (41). All data extraction procedures were implemented using Python (version 3.11).

Data analysis

A total of 11,465 videos were initially collected, along with 177,945 comments between 2020 and 2025. The analytical workflow followed a multistage process (Fig. 7).

Figure 7.

Overview of the methodological framework used in this study, illustrating the main steps from data collection to analysis.

Data preprocessing

This study focuses on two interrelated textual datasets, video transcripts and viewer comments, which together constitute complementary layers of the same discursive ecosystem. Both datasets were merged using the video ID as the unique key, with each record containing the textual content and a source label indicating its origin. All textual data underwent a systematic preprocessing pipeline to ensure consistency across sources. First, all texts were translated into English using Google Translate (42). Next, using the Natural Language Toolkit library in Python (43), we performed a series of text-cleaning operations. These included removing timestamps, HTML tags, URLs, emojis, and nonalphabetic characters; converting all text to lowercase; normalizing whitespace; and eliminating duplicate entries (44). Given the substantial variation in text length across transcripts and comments, several documents exceeded the token limits imposed by the embedding model in the next section. To address this issue, all texts were tokenized using the model-specific tokenizer to obtain precise token counts. When a document exceeded the maximum length, it was divided into sequential segments, each within the token limit and cut only at complete-token boundaries, with all segments retained for subsequent analyses. Furthermore, to assess whether comment patterns were driven by a small number of highly active users, we conducted a descriptive analysis of user contribution patterns. Using unique commenter IDs, we computed the distribution of comments per user, the concentration of contributions among the most active users, repeated participation across videos and thematic categories, and the maximum contribution of a single user within individual videos and macro-categories. We also screened for exact duplicate comments posted by the same user as a rudimentary indicator of repetitive behavior.

Topic modeling and topic similarity

Latent thematic structures within the corpus were identified using topic modeling implemented via the BERTopic framework, a robust method for short, noisy social media texts (45, 46). The adapted pipeline, termed MPNet modeling, employed the MPNet-base sentence transformer to generate dense semantic embeddings. The procedure comprised three principal computational phases: (i) semantic embedding of text segments; (ii) dimensionality reduction via Uniform Manifold Approximation and Projection (UMAP), which preserves topological structure; and (iii) clustering using hierarchical density-based spatial clustering of applications with noise (HDBSCAN), a density-based algorithm that isolates noise. Representative terms for each topic were derived using a class-based term frequency-inverse document frequency (c-TF-IDF) weighting scheme. In the HDBSCAN clustering step, segments that did not belong to a dense and stable cluster were assigned to topic −1. This category was therefore treated as a residual category rather than as a substantive topic equivalent to the retained topics. To support the interpretation of residual and broad low-specificity topics, we conducted a post hoc lexical diagnostic using the most frequent unigrams and bigrams and examined representative segments. This diagnostic analysis did not modify the topic assignments but was used to clarify the distinction between residual unclustered discourse and retained thematic clusters and to identify recurring but overlapping content dimensions within broad topics.

Given the sensitivity of topic interpretability to algorithmic parameters, we systematically optimized hyperparameters for UMAP (including n_neighbors and n_components) and HDBSCAN (min_cluster_size and min_samples) to maximize topic coherence and topic diversity. Topic coherence quantifies the semantic consistency of the top-ranked words within each topic. In contrast, topic diversity measures the proportion of unique lexical items relative to the total lexical items across topics (47). The selected configuration yielded coherence and diversity scores of 0.48 and 0.72, respectively. The complete set of tested hyperparameter values for MPNet modeling is provided in Table S7. Following this optimization, we performed an iterative refinement cycle. Topics identified as irrelevant or outside the scope of avian influenza were manually removed, with their associated documents, based on the authors' subject-matter expertise, and the model was retrained on the filtered corpus until a stable and domain-relevant thematic architecture was obtained (48). Subsequently, we manually synthesized the fine-grained topics into higher-order thematic categories, informed by their most probable terms and hierarchical clustering dendrograms. To evaluate thematic coherence between video content and comments, we computed a semantic signature for each macro-category by averaging the mpnet model embeddings of its constituent texts. Cosine similarity (49) between these category-level vectors was then calculated to quantify alignment; lower similarity values were interpreted as indicative of thematic divergence or reinterpretation. Cosine similarity was used as a descriptive measure of semantic proximity between category-level embedding centroids. These values reflect geometric relationships in the embedding space and were not interpreted as inferential correlations between variables. Finally, we conducted a longitudinal analysis to trace the temporal dynamics, emergence, persistence, and decline, of identified themes.

Sentiment analysis

Sentiment analysis was conducted on the text corpus following the topic modeling phase. We approached sentiment classification as a supervised learning task, utilizing a manually annotated dataset to categorize textual segments into positive, negative, or neutral polarities. This methodological step aimed to delineate prevalent affective orientations and evaluative stances within the discourse (50). Several pretrained transformer architectures, namely RoBERTa, MPNet, BERT-base, and XLM-RoBERTa (51), were evaluated for the classification task. Model selection was based on performance metrics obtained via fine-tuning on a domain-specific annotated corpus of 2,500 texts. The XLM-RoBERTa model achieved the highest accuracy (0.81) and was consequently selected for its robust contextual representation and efficacy in processing informal, multilingual user-generated content (52). For each text segment, two sentiment-derived indices were computed. A continuous polarity score (53) was calculated, defined as the difference between the predicted probabilities of positive and negative sentiment classes, yielding values ranging from −1 (entirely negative) to +1 (entirely positive). In addition, a subjectivity score (53) (range 0–1) was computed, generated using the TextBlob library, to quantify the extent to which each text expressed opinion rather than factual statement. Temporal dynamics of sentiment were analyzed to identify fluctuations in affective valence across the observation period. Furthermore, differences in sentiment distributions between videos and comments were assessed using χ² tests of independence. Given the large sample size, Cramér's V was calculated as an effect size measure to quantify the magnitude of the association and to avoid relying solely on P values (54).

Supplementary material

Supplementary material is available at PNAS Nexus online.

Competing interest

The authors declare no competing interests.

Funding

The authors declare no funding.

Author contributions

Lisa Vors (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Validation, Visualization, Writing—original draft), Israa Farouk (Conceptualization, Data curation, Methodology, Resources, Software, Validation, Writing—review & editing), Paul Delon (Methodology, Resources, Software, Validation), Didier Raboisson (Project administration, Supervision, Validation, Writing—review & editing), Jean-Christophe Boucher (Conceptualization, Methodology, Supervision, Validation, Writing—review & editing), and Guillaume Lhermie (Conceptualization, Methodology, Supervision, Validation, Writing—review & editing)

Data availability

The data analyzed in this study were retrieved from publicly available YouTube videos and associated comments using the YouTube Data API. Due to platform policies and the dynamic nature of online content, the raw extracted YouTube data are not redistributed. This includes raw text segments, original text, commenter identifiers, channel names, and other metadata that could facilitate re-identification of users or content. The Python scripts used for data collection, preprocessing, topic modeling, sentiment analysis, statistical analyses, and figure generation are provided in the Supplementary material, together with documentation of the analytical workflow and model parameters. These scripts allow the data collection and analytical pipeline to be replicated from publicly available YouTube content, subject to availability through the YouTube Data API (https://developers.google.com/youtube). Aggregate outputs needed to support the reported results are included in the article and Supplementary material.