Data for this study were collected through X’s application programming interface (API), a tool that allows developers to access and retrieve publicly available data from the platform [22]. We used the platform’s Basic tier access that permits retrieval of up to 10,000 tweets per month. In addition, the majority of our dataset consists of manually extracted public tweets collected over 5 months, which formed the primary source of our data. To ensure comprehensive coverage, 2 researchers conducted literature reviews and expert consultation methods [23]; the following keywords were used for inclusion: “LivingWithHIV,” “HIVPositive,” “HIVStories,” “HIVAware,” “EndHIVStigma,” “HIVDisclosure,” “HIVSupport,” “UequalsU,” “PrEP,” “AIDSAwareness,” “HIVAdvocate,” “HIVCommunity,” “HIVJourney,” “HIVTesting,” “FightHIV,” “HIVStigma,” “HIVCare,” and “HIVUndetectable.”

We collected tweets between June 1, 2023, and August 31, 2024, yielding an initial dataset of 229,340 tweets. The study team implemented comprehensive data preprocessing to optimize text quality. This process included: (1) text normalization through conversion to lowercase to ensure consistent term recognition [24]; (2) noise reduction by removing special characters, URLs, “@” mentions, hashtag symbols (while preserving the hashtag text content for thematic analysis), punctuation marks, and emojis to enhance topic modeling accuracy [25]; (3) linguistic filtering to exclude non-English character and non-ASCII (American Standard Code for Information Interchange) symbols to ensure compatibility with our algorithm; and (4) tokenization to segment text into individual terms for computational analysis [26]. In addition, we performed text normalization using lemmatization to reduce inflectional forms to their base forms (eg, “testing” to “test”), which helps consolidate semantically similar terms [27]. To address data redundancy, we applied deduplication algorithms that identified and removed identical and near-identical content, including retweets [28]. After completing all preprocessing steps, the final study sample consisted of 191,972 unique tweets, representing approximately 83.7% of the initial dataset.

This study involved secondary analysis of publicly available data and was granted exempt status through Arizona State University’s Institutional Review Board (IRB; STUDY00020345). To protect privacy, we removed all personally identifiable information during preprocessing. To address identifiability concerns, all user-generated content in our findings was either removed or paraphrased, ensuring that users cannot be traced through search engines. Since we analyzed only publicly available tweets, no informed consent was required. All data were stored in anonymized form on the research institution’s infrastructure to prevent unauthorized access. No compensation was provided to participants as this was an analysis of publicly available social media data.

To understand emotional content in tweets, we conducted sentiment analysis using 2 well-known sentiment classifiers: the Valence Aware Dictionary and sEntiment Reasoner (VADER) and the National Research Council Canada Emotion Lexicon (NRC Emotion Lexicon) [42]. We selected both tools because VADER offers overall sentiment polarity optimized for social media language, while NRC offers granular emotional classification that VADER cannot capture. The VADER analyzer provided polarity scores (positive, negative, neutral, and compound—a unified score that combines the other 3 scores into a single sentiment metric ranging from −1 to +1) for each tweet. We then used the NRC Emotion Lexicon to map these tweets to 8 basic emotions (anger, fear, anticipation, trust, surprise, sadness, joy, and disgust) [43]. This mapping process integrated both the VADER sentiment scores and the presence of emotion-indicating terms from the NRC lexicon. The resulting emotion scores were minimum-maximum normalized to a 0‐1 scale to ensure comparability across the dataset.

We used time-series visualizations to examine temporal patterns in HIV-related X engagement from June 2023 to August 2024. We analyzed both absolute metrics (total tweet volume) and normalized per-tweet engagement metrics (comments, transfers, favorites, bookmarks, and views) to account for varying posting frequencies. The visualizations were structured as multi-axis time-series plots to facilitate comparative analysis between related metrics while accommodating their different scales. We constructed four complementary visualizations: (1) average comments and transfers per tweet, (2) average favorites and bookmarks per tweet, (3) average views per tweet, and (4) monthly tweet volume.

Analysis of 191,972 tweets identified different term frequency patterns in HIV-related social media discourse. “Testing” and “support” emerged as the most prominent terms, reflecting the dual emphasis on clinical services and assistance in HIV conversations. Secondary high-frequency terms included “community” and “treatment.” Health-focused terms like “care” and “prevention” appeared with substantial frequency, while social impact terms such as “stigma” and “awareness” maintained a strong presence throughout the dataset. Action-oriented language including “fighting” and “need” consistently appeared, suggesting an advocacy dimension on the X platform.

Our theory-driven classification of HIV-related discourse revealed 5 distinct categories of content. The information and education theme represented the largest category (120,985/191,972, 63.02%), followed by opinions and commentary (23,863/191,972, 12.43%), personal experiences and stories (19,672/191,972, 10.25%), stigma and social impact (14,252/191,972, 7.42%), and support and resources (13,200/191,972, 6.88%). The supervised classification models achieved high performance metrics, with an accuracy of 0.91 and a macro F₁-score of 0.90. Representative tweets for each category are presented in Table 1.

Following our theory-driven LDA approach, we further analyzed each of the 5 predefined theoretical themes to extract 3 subtopics within each theme. This resulted in 15 total subtopics that provided additional granularity to our predetermined theoretical framework. Each theme—information and education, personal experiences and stories, opinions and commentary, stigma and social impact, and support and resources—contained 3 subtopics with distinctive top word distributions as presented in Table 2.

Based on our data-driven LDA topic model analysis, we identified 15 as the optimal number of topics. As shown in Table 3, perplexity decreased from the topic number 5 to 15 and then increased for higher number of topics. Similarly, log-likelihood reached its peak at 15 topics. GridSearch determined that a learning rate of 0.9 provides the best performance for this configuration, where both evaluation metrics achieve their optimal values simultaneously. This parameter selection represents the optimal tradeoff between model complexity and interpretability. Fewer topics result in oversimplified representations that miss important thematic nuances, while more topics lead to oversegmentation that reduces interpretative clarity. See Multimedia Appendix 2 for visualization.

After analyzing the top word distributions, 3 researchers constructed topic labels and reached consensus. They identified patterns of thematic similarity and finally organized the topics into 8 broader thematic themes: community experiences, medical and epidemiological information, global perspectives, social commentary, stigma and discrimination, support and services, policy and governance, and prevention (see Table 4).

The sentiment analysis showed the temporal distribution of the 8 emotions across our dataset from June 2023 to August 2024 (see Figure 1). As shown in the stacked area chart, where the y-axis represents normalized emotional intensity, fear, anger, and trust appear as dominant emotions throughout the study period. While the overall distribution of emotions remained relatively stable, temporal analysis demonstrated subtle fluctuations in emotional intensity, particularly during public health announcements and awareness activities.

Our analysis of tweet engagement metrics (n=191,972) revealed distinct temporal patterns around HIV awareness events (see Figure 2). The tweet volume reached its highest point during World AIDS Day in December 2023 (approximately 20,000 tweets), with subsequent peak engagement metrics in January 2024, including comments (1.3 per tweet), transfers (8 per tweet), and views (4000 per tweet). A second period of high activity occurred during National HIV Testing Day and Zero HIV Stigma Day in June-July 2024, when both posting volume (14,000 tweets) and engagement metrics increased simultaneously, with views reaching 3000 per tweet and favorites peaking at 32 per tweet.

This study presents a comprehensive analysis of HIV-related discourse on 191,972 tweets over 14 months from June 2023 to August 2024. Our primary contribution is the innovative combination of theory- and data-driven LDA approaches, offering complementary perspectives on HIV discourse that neither method could achieve alone. Our approach integrated 4 complementary methods: artificial intelligence methods for text classification and topic modeling, qualitative thematic analysis for contextual interpretation, sentiment analysis for emotional content understanding, and traditional observational methods for temporal trend analysis. Our dual theory and data-driven topic modeling approaches help bridge the gap between theoretical frameworks and real-world public discourse, revealing both established knowledge and emerging public concerns in HIV communication.

The theory-driven LDA approach organizes the discourse within 5 comprehensive themes: information and education, personal experiences and stories, opinions and commentary, stigma and social impact, and support and resources. We further developed these themes into 15 subtopics to allow for more detailed analyses. These categories are shaped by previous theoretical framework expectations and reflect academic concerns with structured, future-oriented responses to the HIV epidemic. This process not only strengthens the existing theoretical framework but also supports their continuity and evolution in academic research [44]. For instance, the prominence of youth health care and prevention programs may reveal an investment in long-term health education and intervention. Similarly, discussions about government policy and medical ethics may mirror the scholarly emphasis on institutional responsibility and systemic solutions [45].

In this sense, the theory-driven approach plays a critical role in reinforcing, elaborating, and updating existing frameworks. The “Information and education” theme includes subtopics like health care in African contexts, youth-specific prevention, and public health messaging. “Personal experiences and stories” offers a more emotional layer, capturing the realities of living with HIV, survival narratives, and personal reflections. “Opinions and commentary” points to structured discussions around medical perspectives, social issues, and government policies. The themes of “Stigma and social impact” and “Support and resources” cover interpersonal and racial stigma, as well as community-based and mental health support. Altogether, the 15 subcategories reflect a well-rounded academic framing that integrates institutional, medical, emotional, and social dimensions of the HIV conversation.

The data-driven LDA method captured 15 topics that emerged organically from public discussions, which we further organized into 8 broader themes: Community experiences, medical and epidemiological information, global perspectives, social commentary, stigma and discrimination, support and services, policy and governance, and prevention. While it may lack the clear structure of theory-led models, it uncovered discursive themes that are overlooked in research. Examples include digital commentary on the intersection of HIV and COVID-19, grassroots prevention campaigns, and anxieties around funding and service provision.

In this sense, the data-driven approach reveals what matters most to users in the real world on discussion platforms, from LGBTQ+ and people of color advocacy to personal treatment narratives and critiques of health care access. Community experiences include LGBTQ+advocacy, personal treatment experiences, and POC-led discussions, while medical and epidemiological information focuses on disease impact, diagnosis statistics, and public screening efforts. Global perspectives reintroduce the African context but with a broader emphasis on global disparities. Social commentary captures how HIV is discussed alongside other major issues like COVID-19, particularly through digital media. The prevention cluster extends from traditional messaging to community-led initiatives and biomedical approaches like PrEP [46]. Overall, these themes reveal a bottom-up perspective, highlighting lived experience, advocacy, and the urgency of practical solutions—dimensions that are often underrepresented in traditional academic models.

Despite different methodological approaches, both theory- and data-driven LDAs converge on several key themes, highlighting a shared prioritization across theory and discourse. (1) Prevention is a critical area in both analyses, with the theory-driven model focusing on biomedical messaging (eg, antiretroviral use), and the data-driven model elaborating on multiple approaches, including PrEP and community-led efforts. This suggests that prevention-centered interventions should integrate both biomedical messaging (eg, antiretroviral use) from theory-driven approaches with the community-led and PrEP-focused strategies highlighted in data-driven analysis. (2) Lived Experiences with HIV are central across both models. The theoretical model dissects structured narratives (eg, survival stories), while the data-driven results foreground advocacy (eg, LGBTQ+ experiences) and treatment-sharing communities [47]. (3) Stigma is addressed comprehensively: racial, community, and interpersonal impact in the theory-driven model parallel anti-stigma campaigns and discrimination awareness in the data-driven analysis [48]. It points to intervention needs that address racial, community, and interpersonal dimensions through coordinated antistigma campaigns. (4) Support systems, though framed differently, appear in both. The theory-driven approach emphasizes community and mental health support, while the data-driven model foregrounds access issues, service delivery, and biomedical innovation [49]. It suggests that interventions should bridge institutional mental health resources with community-based service delivery innovations for comprehensive care. (5) Global health and Africa feature prominently in both, underscoring shared recognition of HIV’s global dimensions—though the data-driven themes offer a more explicitly labeled Global HIV Context.

These areas of convergence not only validate the robustness of certain themes across both analytical approaches, but also illustrate how the 2 mindsets behind them—top-down and bottom-up—can complement each other and be integrated in public health discourse analysis.

The most important divergence lies in what each approach reveals that the other does not. The data-driven approach identifies 3 themes absent from the theory-guided structure: Digital social commentary, policy and funding initiatives, and community-led prevention. These themes reflect real-time societal reactions, participatory politics, and community-driven change—critical aspects often missing in literature-based modeling [50]. In contrast, the theory-driven model includes nuanced psychosocial categories like personal feelings or medical perspectives that reflect deeper affective or clinical framing not prominent in spontaneous X discourse.

These differences are not merely methodological—they highlight actionable gaps between public discourse and academic frameworks. Data-driven themes capture lived urgency and sociopolitical energy, while theory-driven themes offer structure, historical continuity, and future-oriented [51]. Effective HIV communication requires bridging this gap by integrating grassroots narratives with evidence-based frameworks to ensure messages resonate with target communities while maintaining scientific accuracy [52].

The word cloud analysis reveals “testing” and “support” as dominant terms in HIV-related social media discourse, reflecting emphasis on both clinical services and community assistance. Secondary terms like “treatment,” “care,” “prevention,” and “community” demonstrate the focus on health interventions [53], while terms like “stigma” and “awareness” highlight social dimensions [54]. The relative absence of policy-related terminology suggests public discourse may emphasize individual experiences over structural solutions, indicating an opportunity for health communicators to expand discussions around systemic approaches to HIV response [55].

Sentiment analysis of HIV-related discourse revealed a complex emotional landscape where trust appears alongside negative emotions like fear and sadness throughout the study period. While emotional distribution remained relatively stable over time, subtle fluctuations occurred during health announcements and awareness campaigns. These findings suggest that HIV discourse encompasses both concern (represented by fear and sadness) and confidence (represented by trust), likely reflecting public responses to risk information alongside messages about treatment advances and support resources [56]. The persistent presence of trust may indicate growing public confidence in HIV interventions, while the concurrent fear highlights ongoing concerns about transmission and stigma [57]. This emotional duality underscores the importance of balanced communication strategies that acknowledge concerns and fears while reinforcing efficacy and support messages. This strategy—pairing the experience of fear with solutions that can help alleviate the fear—has been established as a highly effective persuasion technique in public health campaigns for decades [58].

Interestingly, consistent with a multitude of past research specific to social media [59,60], emotional fluctuations often coincided with periods of heightened social media activity around major awareness events, suggesting a possible link between emotional dynamics and patterns of public engagement. Temporal analysis revealed 2 distinct patterns of social media interaction around major HIV awareness events: a delayed engagement pattern following institutional communications and an immediate engagement pattern driven by community-driven messages. The first pattern emerged during World AIDS Day in December 2023, where a sequential engagement emerged; in this case, peak tweet volume (20,000 tweets) occurred before peak engagement metrics in January 2024. This delayed pattern may reflect the nature of peer influence in institutional health communication: official agencies and health organizations initially disseminated formal health messages and campaign information, creating an information-rich environment that requires time for peer networks to process. The lag in engagement illustrates how individuals process health information in stages: first absorbing official messages, then contextualizing them locally, and finally transforming them into relevant peer-to-peer discussions [61]. In contrast, the second pattern during National HIV Testing Day and Zero HIV Stigma Day in June-July 2024 showed simultaneous increases in both tweet postings and engagement. This surge coincided with high participation from HIV-affected communities, including people living with HIV, their support networks, health care providers, and advocacy groups. When members of these interconnected communities share personal experiences related to testing or stigma, they strengthen existing social bonds and prompt immediate engagement from others who share similar experiences or support roles [62]. These contrasting patterns highlight how peer influence operates differently in different networks: institutional communications lead to gradual, top-down information sharing, while community-based messages trigger immediate peer engagement through horizontal social network connections [63].

For health care practitioners, our findings reveal that HIV discourse contains a complex emotional blend of trust, fear, and sadness. This suggests that, consistent with other public health campaigns [58], communication strategies should acknowledge these diverse emotions. Clinicians should integrate supportive messaging that addresses emotions while providing factual information. Our temporal analysis identifies strategic timing opportunities during awareness events when both posting volume and user interaction peak. Policymakers can use these insights to design interventions that address the disconnect we identified between individual experiences and structural solutions [64], particularly in addressing stigma themes that appeared prominently in both analytical approaches. Future research should build on our multimethod framework to further explore underrepresented topics in HIV social media discourse, examine how digital platforms continue to shape public health communication, and leverage these data with key behavioral outcomes like attendance at local health events, clinic visits, and engagement with public health resources [65].

This study has several limitations. First, data collection was limited to the X platform, which may not represent the full spectrum of social media HIV discourse. Second, our 14-month analysis period captures only recent trends, potentially missing longer-term patterns in HIV-related communication. Third, our sentiment analysis was conducted at the overall dataset level rather than within specific topics, which limits our understanding of how emotional response may vary across different topics. Fourth, our analysis focused solely on text content without incorporating visual media such as images and videos, which could provide additional context. While our topic modeling approach using LDA is appropriate for text data, future research would benefit from multimodal analysis that integrates both textual and visual elements for a more comprehensive understanding of HIV communication on social media.

This study demonstrated the value of a multimethod approach in analyzing HIV-related social media discourse. Our parallel analysis through theory-driven and data-driven LDA revealed complementary perspectives on HIV communication on social media. Word frequency analysis identified “testing” and “support” as dominant themes, while sentiment analysis showed a complex emotional landscape with trust appearing prominently alongside fear and sadness throughout the study period. Temporal analysis demonstrated strategic opportunities during HIV awareness events when both posting volume and engagement metrics peak significantly. These findings underscore the importance of developing HIV messaging that integrates medical information with community perspectives, addresses the emotional complexity revealed in our sentiment analysis, and leverages the timing patterns identified in our temporal data to maximize reach and impact. As social media platforms increasingly serve as accessible sources of information when traditional health resources may present barriers, these communication strategies become even more essential for effective public health outreach.