VaxPulse: AI Vaccine Infodemic Surveillance
- VaxPulse is an AI-enabled platform designed to detect and respond to vaccine information disorder through real-time surveillance and data-driven analysis.
- It employs deep learning, active learning, and data augmentation to monitor online vaccine discourse, sentiment, and AEFI narratives across multiple digital platforms.
- The system supports public health decision-making with modular dashboards and retrieval-augmented querying for evidence-grounded analysis and rapid response.
Searching arXiv for the most relevant VaxPulse papers and closely related vaccine-infodemic monitoring work. VaxPulse is a family of AI-enabled vaccine intelligence systems designed to monitor vaccine-related information disorder, public concerns, and post-licensure safety narratives in near real time. In the current arXiv literature, the name denotes at least three closely related implementations: VaxPulse Vaccine Infodemic Risk Assessment Lifecycle (VIRAL), an active global infodemic surveillance and response platform; an AEFI-centric framework developed to extend Victoria’s vaccine safety service SAEFVIC with online concern monitoring; and VaxPulse Query Corner, a retrieval-augmented interface for evidence-grounded querying of vaccine-related online discussion (Dimaguila et al., 6 Jul 2025, Javed et al., 7 Jul 2025, Javed et al., 17 Jul 2025). Across these formulations, VaxPulse is positioned as a decision-support environment for immunisation programs, infodemic managers, government agencies, health communicators, and vaccine safety teams.
1. Scope, nomenclature, and system variants
VaxPulse emerged from the practical need to detect and respond to vaccine infodemics during and after the COVID-19 pandemic. The 2025 papers describe vaccine infodemics as involving misinformation, disinformation, gaps in information, mistrust, and inauthentic amplification that can reduce vaccine confidence and undermine immunisation programs. They also frame the problem as a public health threat rather than merely a communication problem, because online narratives can reduce vaccination coverage, increase hesitancy, and contribute to outbreaks of vaccine-preventable disease (Dimaguila et al., 6 Jul 2025).
The name does not refer to a single immutable software artifact. Rather, it denotes a platform lineage with overlapping goals and partially different operational emphases: global infodemic surveillance, AEFI-contextualised post-licensure monitoring, and retrieval-grounded analyst interaction. This suggests a modular research program in which data ingestion, classification, topic analysis, dashboards, and interactive querying can be recombined for different public health workflows.
| VaxPulse component | Primary function | Distinctive elements |
|---|---|---|
| VaxPulse VIRAL | Global vaccine infodemic surveillance and response | Deep learning, active learning, data augmentation, PowerBI dashboards |
| VaxPulse for SAEFVIC | Post-licensure surveillance enhancement | AEFI-centric sentiment and topic analysis |
| VaxPulse Query Corner | Query answering and summarization | RAG, vector database, reranking, grounded answers |
A recurrent description of VaxPulse VIRAL is that it provides a “near real-time pulse” of vaccine-related information disorder. The platform is also described as an “information disorder mission control” environment, indicating that its outputs are intended to support action rather than passive observation (Dimaguila et al., 6 Jul 2025).
2. Analytical targets and monitored signals
The core analytical target of VaxPulse is vaccine-related online discourse at operational scale. The system continuously monitors public sentiment around priority vaccines and immunisation programs, vaccine concerns and self-reported adverse events, social bot influence, misinformation and disinformation in posts and media, and language complexity as it relates to sentiment and mis/disinformation (Dimaguila et al., 6 Jul 2025).
In the AEFI-centric formulation, these signals are explicitly tied to adverse events following immunisation rather than treated as generic vaccine sentiment. That distinction is central to the framework. A negative post may reflect a real adverse event experience, a request for reassurance, uncertainty about side effects, or concern about a specific vaccine, dose, or demographic issue. VaxPulse therefore attempts to contextualise negative, neutral, and positive discourse rather than collapsing all critical language into anti-vaccine sentiment (Javed et al., 7 Jul 2025).
The data sources reported across the VaxPulse papers include X via Tweepy, Reddit via PRAW, YouTube via the Google API Client Library, Facebook, and Google Trends (Javed et al., 7 Jul 2025). For one reported AEFI-oriented analysis, the framework retrieved 871,596 social media comments related to adverse effects of COVID-19, Shingrix®, and RSV vaccines from August 2023 to July 2024, with 819,388 comments remaining after cleaning and a filtered subset of 705,145 COVID-19 vaccine-related comments (Javed et al., 7 Jul 2025).
This scope places VaxPulse within digital epidemiology and infodemic surveillance, but with a stronger emphasis on operational immunisation intelligence than on descriptive social media analytics alone.
3. Machine learning architecture and workflow
VaxPulse VIRAL is described as an AI-powered social listening platform that leverages deep learning, active learning, data augmentation, fine-tuned classifiers, and Generative AI for collecting, augmenting, and processing data (Dimaguila et al., 6 Jul 2025). Deep learning serves as the backbone for text classification, sentiment analysis, topic detection, and large-scale pattern recognition in noisy online discourse. Active learning is used to prioritise samples for expert review in settings where narratives evolve rapidly and labeled data become outdated. Data augmentation is used to diversify training data across vocabularies, languages, vaccine topics, and emerging narrative styles (Dimaguila et al., 6 Jul 2025).
The VIRAL development process is presented as an iterative learning health system rather than a static product. The reported workflow consists of six stages: observed online vaccine concerns, monitoring sentiments and personally experienced vaccine reactions, tracking social bot influence, identifying misinformation/disinformation, building an international expert network, and refining the platform and dashboards based on feedback (Dimaguila et al., 6 Jul 2025). That workflow formalises continuous adaptation as part of system design.
The AEFI-centric framework specifies a seven-stage pipeline: data scraping, preprocessing, vaccine-specific comment selection, sentiment analysis, concerns identification, topic modelling, and future work on multilingual data analysis (Javed et al., 7 Jul 2025). Preprocessing removes links, mentions, special characters, and short sentences. Vaccine-specific comment selection uses a two-step procedure: an ensemble of fine-tuned RoBERTa-large-mnli and GPT-4o for broad categorisation into general vaccine discussions, personal experiences, and unrelated topics, followed by LLM-based prompt engineering with GPT-4o to identify comments that are specifically vaccine-related (Javed et al., 7 Jul 2025).
For sentiment classification in the SAEFVIC-oriented work, CT-BERT V2 was fine-tuned on 13,715 annotated posts to classify vaccine-AEFI comments into negative, neutral, and positive, achieving 93% accuracy. A hybrid approach then used GPT-4o for comments with low BERT confidence (Javed et al., 7 Jul 2025). Topic extraction is performed with BERTopic, chosen because it uses BERT-based embeddings, captures semantic similarity better than classical topic models, and can be refined using LLMs for better topic representations (Javed et al., 7 Jul 2025).
4. Dashboards, governance, and decision-support design
VaxPulse is not presented as a classifier in isolation; it is presented as a dashboard-mediated operational system. The VIRAL papers report PowerBI dashboards built as a “deck of graphs” that can be updated regularly and tailored to specific audiences. Dashboard priorities and landing-page content were selected through review by experts in informatics, immunology, paediatrics, machine learning, data science, epidemiology, vaccine safety, infectious diseases, and consumer engagement (Dimaguila et al., 6 Jul 2025).
The example dashboard functions include infodemic risk assessment, country and data source filters, time-window selection, geographic distribution, sentiment analysis, sentiment trend, misinformation trend, misinformation labels, topic-level concern categories, and vaccine-specific discussion examples such as shingles vaccine concerns (Dimaguila et al., 6 Jul 2025). The papers repeatedly stress that these visualizations are intended to make complex social data immediately interpretable and actionable.
The system’s governance model is explicitly interdisciplinary and international. The reported expert network involves Indonesia, Philippines, India, Canada, Switzerland, and the World Health Organization. Feedback loops also include the Asia-Pacific Vaccine Research Network, National Immunisation Technical Advisory Group members, vaccinologists, vaccine researchers, clinicians, Ministry of Health personnel, and domain experts in vaccine safety, pediatrics, immunology, infectious disease, data science, and machine learning (Dimaguila et al., 6 Jul 2025). This matters because local epidemiological and sociocultural priorities are treated as first-class design constraints rather than downstream deployment issues.
A formal Infodemic Risk Index is described as a future direction rather than a completed component. The literature also mentions an attempt to track the significance of topic changes over time using a “velocity change” formulation, although the exact formula is not provided in the available text (Dimaguila et al., 6 Jul 2025). Multilingual expansion is similarly framed as ongoing, with support planned for Filipino, Urdu, Hindi, Bangla, Spanish, and Farsi to enable ethno-lingual stratification (Dimaguila et al., 6 Jul 2025).
5. AEFI contextualisation and post-licensure surveillance
The SAEFVIC-oriented VaxPulse paper extends conventional vaccine safety surveillance by integrating AEFI reporting with sentiment analysis, concern extraction, and topic modelling. The central claim is that AEFI data alone are insufficient for contemporary vaccine surveillance because public interpretation of adverse events now unfolds continuously in online media, where personal experiences, rumors, misinformation, and distrust co-circulate (Javed et al., 7 Jul 2025).
This framework therefore treats contextualisation as analytically indispensable. A real concern may require transparent safety communication; misinformation may require correction; neutral uncertainty may require education and reassurance. The same AEFI topic can have different operational meanings depending on whether surrounding discourse is fearful, angry, skeptical, or merely information-seeking. This directly addresses a common misconception in vaccine social listening: that negative sentiment is equivalent to anti-vaccine opposition.
For the COVID-19 vaccine subset in the reported analysis, VaxPulse identified 9.87% negative, 27.23% positive, and 62.90% neutral comments (Javed et al., 7 Jul 2025). The authors interpret this as a shift from more negative sentiment toward more neutral sentiment, indicating increased public interest in vaccine information and side effects rather than purely hostile opposition. Persistent concern themes included safety, side effects, misinformation, trust in authorities, and previous negative experiences, with mis/disinformation remaining a major issue, especially around co-administered vaccines (Javed et al., 7 Jul 2025).
Illustrative AEFI-related topics included Vaccine-Related Deaths, Blood Clot Concerns and Experiences, and Myocarditis and Pericarditis Following COVID-19 Vaccination (Javed et al., 7 Jul 2025). A focused case study on women’s vaccine hesitancy identified irregular menstrual cycles, experiences and recommendations for vaccines during pregnancy, and COVID vaccine and miscarriage concerns as prominent themes. The most common female-related AEFI concern was menstrual cycle abnormalities, peaking around November 2023 with under 800 comments and again in May 2024 with about 1,200 comments (Javed et al., 7 Jul 2025). The paper further reports that SAEFVIC investigated menstrual changes following COVID-19 vaccination and found an association, and that some people who menstruate felt dismissed and distressed because clinicians were skeptical about the potential vaccine link. SAEFVIC responded with TikTok videos and Instagram videos intended to reassure women, explain that the effects were temporary, and educate the public (Javed et al., 7 Jul 2025).
The same study also reports that social bots contributed 23.72% of tweets and amplified AEFI-related discussion in human posts (Javed et al., 7 Jul 2025). That finding links infodemic monitoring to vaccine safety surveillance: amplification dynamics can distort perceived prevalence and urgency of safety concerns even when the underlying topic is real.
6. Retrieval-augmented querying: VaxPulse Query Corner
VaxPulse Query Corner extends the platform from monitoring and dashboarding into evidence-grounded analyst interaction. Its rationale is twofold. First, topic modelling and sentiment analysis can be too coarse for operational questions such as whether people believe vaccines cause long-term side effects or what people are saying about the second dose of a specific vaccine. Second, standalone LLMs may miss current events and community-specific concerns, and may hallucinate in ways that are particularly problematic in public health communication (Javed et al., 17 Jul 2025).
The Query Corner pipeline has four reported stages: dataset creation, first iteration retrieval, second iteration refinement, and answer formulation (Javed et al., 17 Jul 2025). Online media posts are scraped from X, Reddit, YouTube, and Facebook using authorized APIs, cleaned, and then filtered through a two-step process consisting of a fine-tuned BERT classifier and GPT-4o with prompt engineering to isolate comments about a specific vaccine such as Shingrix® (Javed et al., 17 Jul 2025).
Selected comments are embedded with OpenAI text-embedding-ada-002, producing 1536-dimensional embeddings stored in a vector database using LangChain (Javed et al., 17 Jul 2025). Query processing then retrieves the top- semantically similar comments, applies LangChain longcontextreorder to mitigate the “lost in the middle” problem, and uses FlashRank as a second-stage reranker, with early exit when there are fewer than 10 comments (Javed et al., 17 Jul 2025). A second retrieval pass applies contextual compression with an 80% similarity threshold and selects 50% of the top-ranked comments from the first iteration before ranking and reranking again. Final responses are generated with GPT-4o in one of four output modes: Answer the Question, Topic of discussion, Summarise, or Public Concerns (Javed et al., 17 Jul 2025).
The principal case study uses a Shingrix social media corpus comprising 60,935 total comments from January 2018 to October 2024, of which 35,103 vaccine-related comments were segregated (Javed et al., 17 Jul 2025). Evaluation used RAGAs with 276 synthetic test cases: 103 for question answering, 56 for public concerns, 45 for summarization, and 72 for discussion topics, with 50% simple questions, 25% reasoning, and 25% multi-context questions (Javed et al., 17 Jul 2025).
Retrieval quality improved from an average context precision of 0.56 and context recall of 0.85 in the first iteration to 0.66 and 0.91 in the second iteration (Javed et al., 17 Jul 2025). For answer generation, the reported average metrics were 0.90 faithfulness and 0.89 answer relevancy for question answering, 0.90 and 0.91 for topics of discussion, 0.96 and 0.94 for summarization, and 0.88 and 0.82 for public concerns (Javed et al., 17 Jul 2025). These results support the claim that VaxPulse can convert large volumes of vaccine-related online discussion into queryable evidence rather than only into labels or topic clusters.
7. Relation to adjacent methods, misconceptions, and boundaries
VaxPulse sits within a broader methodological ecosystem of vaccine-related digital signals, but it is broader in scope than any single precursor. Internet search data can serve as a near-real-time “pulse” of vaccination interest: Google’s Vaccination Search Insights index was positively associated with first-dose vaccinations in the subsequent three weeks, and early monthly VSI was strongly correlated with later vaccination rates, reaching in state-level analyses (Malahy et al., 2021). Social-media stance extraction can also track hesitancy dynamics: a heterophily-aware H2GAT+PE framework for Twitter achieved about 0.803 accuracy versus about 0.652 for a text-only RoBERTa baseline, with up to 23% improvement in accuracy, and was used to track event-driven shifts in vaccination attitudes in near real time (Chen et al., 2022). Survey-based vaccine estimates, however, remain vulnerable to severe nonresponse bias: proxy pattern-mixture models showed that two large U.S. surveys had substantially overestimated vaccine uptake, and that sensitivity analysis could recover the direction of the bias and provide meaningful bounds (Andridge, 2023).
These related studies clarify both the promise and the limits of VaxPulse-style systems. Search behavior can lead vaccination uptake, social discourse can reveal fast-moving stance changes, and survey estimates can be biased even at very large scale. VaxPulse differs by combining several of these logics into an operational platform: large-scale online collection, machine learning classification, bot monitoring, expert feedback loops, dashboarding, AEFI contextualisation, and retrieval-grounded querying. This suggests that VaxPulse is best understood as an integrated vaccine infodemic observatory rather than as a single model or a single metric.
A naming boundary is also necessary. The paper "ALIVE: A Low-Cost Interactive Vaccine Storage Environment Module ensuring easy portability and remote tracking of operational logistics to the last mile" describes a portable, actively cooled, IoT-connected vaccine storage and transport module for last-mile cold-chain logistics, but it does not explicitly use, define, or mention the name VaxPulse anywhere in the provided text (Datta et al., 2024). VaxPulse therefore should not be conflated with vaccine storage hardware, remote cold-chain tracking modules, or logistics systems unless a separate documented linkage is established.
A further misconception addressed by the VaxPulse literature is that online concern monitoring is reducible to generic sentiment analysis. The system architecture, especially in the SAEFVIC and Query Corner variants, explicitly rejects that reduction. It instead treats vaccine discourse as a structured mixture of AEFI narratives, information seeking, misinformation, trust dynamics, bot amplification, and vaccine-specific hesitancy themes that must be interpreted in context (Javed et al., 7 Jul 2025, Javed et al., 17 Jul 2025).