The study protocol for the parent review [20] was registered in PROSPERO (CRD42023458870) [21]. Our report followed the Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRSIMA) Statement [22].

Of the 142 identified citations, 53 were duplications and 32 were excluded by title screening (Figure 1). Of the remaining 57 citations, 49 were excluded by full-report screening because they were modeling studies (n=21), interventional studies (n=13), non-AI studies (n=8), reviews (n=6), or non-PrEP–relevant studies (n=1). The remaining 8 studies were eligible for this review.

The study years included were from 2010 through 2022 (Table 1). Because social media posts could not be validated and may not be reliable, all studies scored low on our study quality assessment for measurement, classification, and other biases. Additionally, because social media posts were voluntary and may not represent the voices of those who did not use social media, the study’s validity was diminished; thus, all studies had moderate risk of bias.

Studies used ML in various social media platforms to identify HIV or PrEP-related social media posts (e.g., text, still images, video footage) with the hashtag #HIV or #PrEP on Instagram [28–31], Reddit [30, 31], Tumblr [31], X (formerly Twitter) [30–35], or YouTube [31]. Three studies used NLP [29, 33, 34], which mimicked human decision-making, while one study [28] used convolutional neural networks (CNNs) for analyzing visual imagery.

The total number of social media posts screened by the included studies was more than 58.9 million. The highest number screened for one study was by Adrover [36], who purchased from an official X data reseller almost 40 million tweets that represented the full and unbiased stream of data posted between September 2010 and August 2013. The elements in social media posts that AI analyzed were categorized as either language content (e.g., tweets, texts) or visual content (e.g., still images, video footage).

This systematic review identified 8 infodemiology studies that used AI to screen more than 58.9 million social media posts. Analyzing language elements (e.g., tweets, texts) in social media enabled us to analyze PrEP-related experiences or concerns (e.g., barriers to uptake [30, 31, 35], adverse effects of PrEP [36]), and rumors (e.g., PrEP can prevent COVID-19 infection [30]). Analysis also allowed for the geolocation of MSM who may be PrEP candidates [34] and to predict trends of HIV incidence [33]. Meanwhile, analyzing visual elements (e.g., still images, video footage) on video platforms (e.g., Instagram) revealed demographic profiles of pictures of human faces posted [28] and images of #HIV posts [29]. Here, we discuss how AI can enhance infodemiology to process social media posts, the advantages and disadvantages of using AI in the field, and future uses of AI for analysis.

First, infodemiology studies using AI can help gather mass data. This includes real-time, health-related information and how people consume such information (e.g., adverse effects of medications [36] or accessibility of care during public health emergencies [30, 31]), from a wide variety of communities; specifically from those who may not be included in traditional surveys. One of the biggest questions is the reliability and accuracy of capturing the actual communities’ experiences and conversations.

The Infodemiology studies in this review found that the reasons for changes in PrEP use during the COVID-19 pandemic were the inaccessibility of PrEP care [31] and reduced sexual behaviors [30]. These were the 2 most common impacts of COVID-19 found by a recent systematic review examining the influences of the pandemic on PrEP care from 46 studies [38]. Additionally, renal toxicity, nausea, and vomiting reported on X as experiences of the communities [36] are well-recognized potential side effects of PrEP [39]. These findings indicated that scanning social media posts with AI can capture information typically collected by using traditional time-consuming and labor-intensive surveys, but with the advantage of collecting mass data quickly in a timely manner.

Second, social media helps spread public health information quickly and widely. Relative to traditional media (e.g., newspapers, TV, radio broadcasts), social media accelerates the recirculating of information [36]. As an example, tweeting about the first antiretroviral treatment approved for PrEP (i.e., Truvada) hit a peak when the FDA approval spurred coverage by the traditional media [36]. 512 unique users who tweeted about their HIV experiences on X had an average of 2,300 followers between 2010-2013 [36]. This suggests that a post by one person can spread broadly and influence thousands of followers and communities.

Meanwhile, one of the major PrEP-related tweet topics was to advocate better access, stating that insurance coverage and pricing were the major concerns in accessing care and using PrEP [35]. As a matter of fact, financial support, including changes in medication options (e.g., availability of generic drugs), Medicaid expansion, insurance coverage policies and regulations (e.g., Affordable Care Act or ACA), and PrEP assistance programs from manufacturers, are available [40]. A significant barrier may be low PrEP awareness, including a lack of information on the availability of PrEP financial support. Thus, when we introduce or promote preventive or public health strategies, we can take advantage of the influential factors of social networks and social media posts to promote awareness of the strategies [36].

On the other hand, information posts on social media may not always be appropriate or accurate. Negative sentiment may increase and influence communities to raise concerns [36]. For example, tweets associating undesirable medical experiences, such as adverse effects of medications, may negatively impact persons interested in the medications. False information can also be widely shared. During the COVID-19 pandemic, the rumor that PrEP can be used as a prevention method or precaution against COVID-19 infection was widely spread on X [30]. Evidence that PrEP can prevent COVID-19 infection, or lead to a less severe course of COVID-19 disease, has not been presented or verified [41].

Our current analyses do not explicitly differentiate between the sources of social media posts. However, distinguishing between unsolicited or unregulated posts and those posted by authoritative sources (e.g., government agencies, community organizations) or professionals and experts could provide deeper insights into the sources of information shared on social media. Future studies in this area could explore this differentiation in greater detail to better understand the impact of different sources on public perceptions and behaviors.

Social media can serve as both a powerful tool for promoting awareness and a platform that may contribute to the spread of misinformation. By targeting key influencers, addressing misinformation, or providing additional information to communities on social media platforms, we can enhance PrEP awareness and uptake.

Next, geolocated social media posts can identify populations with concerns and where those issues are across the country. Our review found that HIV-related tweets were more likely to be posted from the counties that had high HIV incidence rates or the priority jurisdictions of the Ending the HIV Epidemic in the U.S. initiative [33, 34]. Meanwhile, the HIV-related tweets were negatively associated with the county-level HIV incidence rate in the following year [33]. These findings could suggest that higher numbers of HIV-related tweets may be indicators of increased risk awareness, support, and community involvement [33]. Additionally, faces appearing in social media posts with the #HIV hashtag are more likely to belong to individuals who have a lower risk of HIV infection, especially older persons, women, and White persons. This could suggest that HIV-related social media posts may not resonate with persons at high risk for HIV and they might underestimate their own risk [28].

Finally, noise problems in ML could be tricky and impact the performance of scanning and analyzing social media posts. “Noise” means tweets about unrelated topics. By dropping the search terms that could be noise [36], analysis problems can be reduced but the ability to identify relevant data may also be reduced. The strategies of noise reduction, such as data cleaning, algorithm selection, and validation, should be well planned and executed. Beyond the use of AI, researchers still have to go through several validation steps, crowdsourcing, and human interventions to declare the validity of filters [32]. Additionally, further refining AI algorithms to identify relevant, high-quality, and influential social media posts related to PrEP is necessary. The integration of such factors as context, source credibility, and legitimacy into AI models is crucial to ensure that analyses capture not only the volume of posts but also evaluate their quality and influence.

One of the review limitations was the limited number of studies identified and reviewed. This is likely due to the evolving nature of research in this field, as well as the relatively novel application of AI in PrEP and infodemiology studies. Some of the evidence discussed in this review was concluded from single studies rather than from a synthesis of multiple studies. It is important to avoid making overly broad conclusions based on the limited number of studies. A synthesis of evidence across multiple studies is needed to draw more definitive conclusions and guide policy and practice effectively. Additionally, the variability in platforms used, study years, and data collection methods presented challenges in generalizing the findings across studies. Our included studies did not provide detailed information on several aspects, such as content type, engagement metrics, tone, and target audiences, which are crucial for understanding the reach and impact of these posts. To strengthen the evidence of PrEP use and behavioral trends, we need additional high-quality studies that explore these research areas in greater depth and consistency.

In conclusion, this review found that infodemiology studies provide real-time, public health trends and behaviors. AI, including ML and DL, helps scan mass, real-time, user-generated data on social media platforms quickly and accurately to capture simultaneous public health concerns and the locations/communities of those discussing the concerns. Social media data may serve as a proxy of people’s behaviors in communities, and analyzing big data has the potential to provide the information needed to improve patient outcomes and gain new insights into various human behaviors [42–44]. Our review explored whether AI can be used to identify populations who may face public health concerns and the areas/locations that may need public health interventions the most, as well as messages that the public seeks. However, we cannot ignore the issue of infodemic, which refers to an excessive amount of information, including false information in online or offline settings [45]. AI may further help identify specific public health topics that are unnecessary in certain locations. We should take advantage of developing technology and use it appropriately and wisely to improve and promote good health for all.