Large Language Models at Population Scale: A Survey and Taxonomy of Public Health Applications

1. Introduction

Public health is concerned with protecting and improving the health of entire populations, and it shapes decisions with far-reaching societal consequences (Acosta 2025,Hattab et al. 2025). Authorities must decide when to declare an outbreak, how to allocate scarce vaccines, and whether a new intervention will mitigate or exacerbate health disparities. These decisions are often made in real time, under substantial uncertainty, and across heterogeneous populations with different risks, resources, and needs. As global health threats become more frequent and complex, from pandemic preparedness to climate-driven disease emergence, the need for scalable, data-driven tools that can support population-level decision-making has become increasingly urgent (Belova et al. 2025,Dixit 2025,Liu et al. 2024,Rizzo et al. 2024).

Recent advances position large language models (LLMs) as promising public health tools (De Angelis et al. 2023,Grattafiori et al. 2024,Guo et al. 2025,Xu et al. 2026). These foundation models efficiently process unstructured text, extract signals from noisy data, and generate actionable summaries (Bommasani et al. 2021). These capabilities naturally support core functions, including social media-based outbreak detection, cross-lingual evidence synthesis, tailored health communication, and equity-focused policy auditing.

However, the current evidence base for LLMs in health is centered primarily on clinical medicine rather than public health. A growing ecosystem of specialized foundation models, clinical benchmarks, and systematic reviews has documented LLM applications in diagnosis, treatment recommendation, medical documentation, and biomedical text mining. More recently, specialized reviews have examined narrower health sub-domains, including social determinants of health extraction (Keloth et al. 2025), mental health screening (Xu et al. 2024), and pharmacovigilance (Hakim et al. 2025). Yet these efforts remain isolated: each addresses a single task at the individual level without connecting it to the broader public health pipeline or examining how LLM roles shift across surveillance, forecasting, communication, and governance. No existing work provides a comprehensive, cross-functional synthesis of LLMs across the entire public health spectrum.

Unlike clinical medicine, which focuses primarily on individual patients, public health addresses questions such as how diseases spread through communities, how to counteract vaccine misinformation at scale, and whether interventions exacerbate systemic inequities. These differences create three structural challenges for LLMs. (1) Data heterogeneity: public health requires reasoning over heterogeneous, population-scale signals, ranging from social media streams and epidemiological time series to verbal autopsy narratives and wastewater metagenomics (Deiner et al. 2024,Liu et al. 2025), with no clear parallel in clinical AI. (2) Cascading failure modes: whereas a clinical AI hallucination may affect a single patient, a public health AI error, such as a false epidemic alert, culturally insensitive messaging, or algorithmically amplified misinformation, can cascade across entire populations. (3) Equity as a first-class objective: public health models must perform equitably across languages, cultures, and healthcare systems, rather than treating bias mitigation as an afterthought.

To bridge this gap, we present the first comprehensive survey of LLMs across diverse public health applications, synthesizing 192 papers published from 2023 to early 2026. To ensure a strict population-health focus, we employed a three-step boundary test (institutional attribution, core function, and substitution; Appendix A.2) to exclude individual clinical diagnostics and drug discovery. The resulting corpus is then organized through a two-dimensional taxonomy(Table 1). The task dimension follows the operational pipeline of public health systems, Sense → Predict → Communicate → Protect → Intervene → Govern, yielding six categories (T1–T6). The role dimension captures how an LLM functions within each task: as a Sensor, Predictor, Communicator, Simulator, or Auditor. Our main contributions are:

• We conduct the first cross-functional survey of LLMs in public health, covering the full pipeline from surveillance to governance with a population-health boundary test.
• We introduce a task × role taxonomy that reveals systematic gaps: the field is heavily concentrated in extraction-oriented sensing (Sensor dominates T1 and T3, which together account for over half the corpus), while anticipatory roles (Predictor, Simulator) and equity-focused auditing remain underexplored.
• We quantify structural imbalances, including that 80.7% of papers use English-only data and most depend on proprietary models, and distill four gap-driven research directions grounded in the taxonomy.

To bridge this gap, we present the first comprehensive survey of LLMs across the full spectrum of public health, synthesizing 192 papers published from 2023 to early 2026. To maintain a strict population-health focus, we applied a three-step boundary test (institutional attribution, core function, and substitution; Appendix A.2) during screening to exclude individual clinical diagnostics and drug discovery. We organize the resulting corpus through a two-dimensional taxonomy, i.e., six core public health tasks mapped against five functional LLM roles, and conduct systematic extraction to surface the methodological adaptations, evaluation gaps, and systemic risks that distinguish population-scale deployment from clinical AI.

The task dimension follows the operational pipeline of public health systems—Sense → Predict → Communicate → Protect → Intervene → Govern—yielding six categories (T1–T6; Table 1). The role dimension captures how an LLM functions within each task—as a Sensor extracting structured signals, a Predictor forecasting trajectories, a Communicator generating actionable messages, a Simulator powering agent-based environments, or an Auditor evaluating fairness and deployment readiness. Across the corpus, Sensor is the dominant role, reflecting the field’s current emphasis on evidence extraction; forecasting, simulation, and intervention-oriented generation remain comparatively underexplored.

2. Background & Preliminaries

Public health as population-level AI. Unlike clinical medicine’s focus on individual patients, public health targets populations, communities, and jurisdictions. Its core actions are therefore collective rather than individual: surveillance, prevention, preparedness, risk communication, screening, vaccination, and resource allocation. This population-level scope expands both the data regime and the evaluation target. Public health systems must reason over social media, policy documents, news, surveillance counts, mobility, environmental signals, scientific evidence, and clinical records, often under reporting delays and distribution shift. Success is not accuracy alone: models must be timely, robust, useful for intervention planning, and sensitive to disparities across populations. These demands shape the surveyed literature across epidemic forecasting (Du et al. 2025,Gong et al. 2025,Liu et al. 2025), misinformation (Song et al. 2025), pharmacovigilance (Carpenter and Altman 2023,Sidorov et al. 2025), health equity (Pfohl et al. 2024,Pierson et al. 2025), and frontline support (Al Ghadban et al. 2023,Ramjee et al. 2025).

From public health functions to survey tasks. The task dimension in this survey follows the operational pipeline of public health systems. Surveillance and early warning motivate T1 Population Health Surveillance; predictive preparedness motivates T2 Epidemic Forecasting & Modeling; health promotion and information management motivate T3 Infodemiology & Health Communication; population assessment and disparities motivate T4 Health Equity, SDOH & Global Health; health protection and prevention motivate T5 Population Intervention & Practice; and preparedness, coordination, and responsible deployment motivate T6 Governance, Ethics & Policy. This framing explains why LLMs in public health are used not only for medical question answering, but also for event extraction, epidemic intelligence, misinformation analysis, evidence review, intervention design, and policy support. Table 2 summarizes the contrast with clinical medicine.

LLM adaptation strategies in public health. Large language models (LLMs) are foundation models trained on large text corpora to predict and generate language (Bommasani et al. 2021,Vaswani et al. 2017). Rather than review architecture in detail, we focus on the adaptation patterns that recur across public health tasks; additional technical background is provided in Appendix A.1. Prompting enables rapid classification, extraction, and communication without parameter updates, and is common in misinformation detection and public health discourse analysis (Brown et al. 2020,Wei et al. 2022). Fine-tuning, including parameter-efficient methods such as LoRA (Hu et al. 2022), is used when tasks require domain-specific labels or numerical adaptation, such as influenza forecasting, SDOH extraction, or adverse-event mining. Retrieval-augmented generation grounds outputs in external evidence and is especially important for fact-checking, guideline interpretation, and policy support (Lewis et al. 2020). Finally, agentic systems connect LLMs to tools, simulators, databases, or multi-agent workflows, enabling applications such as epidemic simulator construction and emergency response planning (Aoki and Ghaffarzadegan 2026,Datta et al. 2026,Shi et al. 2026).

Why public health poses unique challenges for LLMs. Public health introduces challenges that go beyond individual-level clinical AI. First, many tasks require spatiotemporal and mechanistic reasoning: disease spread depends on contact networks, geography, seasonality, reporting delays, interventions, and changing pathogen dynamics, whereas current LLMs have no native representation of transmission parameters or compartmental models (Du et al. 2025,Gong et al. 2025,Liu et al. 2025). Second, public health operates inside an information ecosystem. The same generative capability that supports misinformation detection, fact-checking, and counterspeech can also generate persuasive health misinformation at scale, creating a dual-role paradox with no direct clinical analog (De Angelis et al. 2023,Hussain et al. 2025,Modi et al. 2025). Third, public health treats equity as a first-class objective rather than a secondary bias metric: models must work across languages, cultures, infrastructures, and marginalized communities, while current evidence remains heavily English-centered and high-resource (Chen et al. 2025,Ji et al. 2025,Pfohl et al. 2024,Zhou et al. 2025). These challenges motivate the task-specific taxonomy and recur throughout the survey.

3. LLM Role Dimension

We define LLM roles by the model’s primary functional output in a public health workflow rather than by architecture, training strategy, or data modality. This distinction matters because the same decoder-only model, prompting method, or social-media dataset can support very different public health functions: extracting a signal, forecasting a trend, generating a message, auditing a system, or simulating an intervention. The role dimension therefore asks what the model produces, who consumes the output, and what downstream public health action it supports.

The five roles correspond to recurring questions in population health workflows: Sensor asks what is happening, Predictor asks what will happen next, Simulator asks what could happen under alternative conditions, Auditor asks whether a system is reliable, fair, and safe, and Communicator asks what should be conveyed to people. These roles are not strictly sequential, but together they span the core stages of public health AI from evidence extraction to anticipation, planning, evaluation, and response. This role-based view also previews a structural imbalance in the literature: current work remains strongest in retrospective sensing, while forecasting, simulation, and deployment-oriented evaluation are still emerging.

Sensor. Sensor models convert heterogeneous public health data into structured labels, entities, or signals for downstream analysis. The output is usually consumed by analysts, surveillance systems, or statistical models rather than directly by the public. Examples include event extraction from outbreak reports (Bhadelia et al. 2026,Consoli et al. 2024,Parekh et al. 2024), adverse-event extraction in pharmacovigilance (Li et al. 2024), verbal-autopsy and mortality coding (Coutinho et al. 2026,Wen et al. 2026), and SDOH extraction from clinical text (Gu et al. 2025,Keloth et al. 2025). Across these settings, LLMs function as an evidence-extraction layer that transforms noisy population-level inputs into structured public health intelligence.

Predictor. Predictor models produce forecasts of population-health trajectories, risks, or future states. This role is concentrated in epidemic forecasting and selected surveillance settings. In T2, LLM-based and foundation-model forecasters integrate epidemic time series with policy, mobility, genomic, or spatial context for infectious disease prediction (Du et al. 2025,Dudley et al. 2025,Gong et al. 2025,Kalahasti et al. 2025,Li et al. 2025,Liu et al. 2025,Moon et al. 2025,Panja et al. 2025). Outside classical forecasting, related predictive uses appear in genomic surveillance, such as SARS-CoV-2 variant fitness prediction (Ito et al. 2025). The key distinction from Sensor is temporal: Predictor systems estimate what is likely to happen next rather than structuring evidence about what has already happened.

Simulator. Simulator models generate or control synthetic epidemic dynamics, policy scenarios, or agent behavior in closed-loop environments. This role appears in work on generative epidemic agents (Williams et al. 2023), simulator construction from natural-language specifications (Datta et al. 2026), policymaking agents in structured epidemic settings (Aoki and Ghaffarzadegan 2026), and coordinated intervention planning across interacting regions (Shi et al. 2026). The key distinction from Predictor is counterfactual: Simulator systems explore possible trajectories and decision sequences under alternative assumptions, making them especially relevant for stress testing and planning under uncertainty.

Auditor. Auditor models evaluate whether AI systems, public health decisions, or generated outputs are reliable, fair, and deployment-ready. Unlike the other roles, the auditor’s object is often the model or system itself rather than a public health phenomenon. Applications include equity toolkits and bias mitigation (Ji et al. 2025,Pfohl et al. 2024), anti-LGBTQIA+ and mental-health bias audits (Chang et al. 2025,Haider et al. 2025), adversarial disinformation-risk analyses (Hussain et al. 2025,Modi et al. 2025), and epidemiological QA or policy benchmarking (Espinosa et al. 2025,Harris et al. 2025,Wei et al. 2026). In T6, auditing encompasses institutional risks, role creep, and geographic bias (Chu et al. 2025,Wilson et al. 2024,Zhou et al. 2025). Auditors therefore serve as a primary mechanism for measuring system bias, robustness, and safety.

Communicator. Communicator models translate public health knowledge into human-facing dialogue, guidance, or decision support for communities, frontline workers, patients, and policymakers. The boundary with Sensor is the output audience: Communicator outputs are intended to be read and acted on by people, not merely processed by downstream analytic systems. Central to T3 and T5, they generate evidence-grounded counterspeech, corrective notes, and vaccine messaging (Anik et al. 2025,Hou et al. 2025,Song et al. 2025,Wu et al. 2025). In intervention settings, they power chatbots and assistants for HIV prevention, smoking cessation, and frontline support (Abroms et al. 2025,Govathson et al. 2026,Humphries et al. 2026,Narayan et al. 2026,Ramjee et al. 2025). While LLMs increasingly serve as public- and worker-facing interfaces, fluency alone is insufficient; communicators must remain rigorously evidence-grounded, audience-aware, and operationally safe (Abroms et al. 2025,Liu et al. 2025).

4. A taxonomy of LLMs in Public Health

Public health tasks progress from sensing threats to governing systems. Accordingly, the Task dimension (Table 1) categorizes the literature into six areas: T1 Population Health Surveillance (53), T2 Epidemic Forecasting & Modeling (18), T3 Infodemiology & Health Communication (52), T4 Health Equity, SDOH & Global Health (22), T5 Population Intervention & Practice (18), and T6 Governance, Ethics & Policy (25). T1 and T3 comprise over half the corpus, skewing research toward signal extraction and discourse monitoring, whereas T4 and T6 tackle systemic fairness and deployment infrastructure. Across T1–T6, we identify shared methodologies, evaluation gaps, and a shift in LLM roles from early-stage sensing to later-stage evaluation, communication, and simulation.

4.1. Population Health Surveillance (T1)

This section reviews LLMs for population health surveillance, that is, the detection and monitoring of health threats from heterogeneous population-level data. Compared with standard clinical diagnostics, surveillance must operate over much broader signals, ranging from social media and informal news to verbal autopsy narratives and wastewater metagenomics, and must do so under conditions of incomplete reporting, multilingual coverage, and rapidly evolving threats. The defining computational challenge is therefore not a single prediction problem, but the conversion of noisy, unstructured, and cross-modal inputs into reliable, timely, and structured signals that can drive public health response.

Event and syndromic surveillance. The first and largest strand treats surveillance as an information-extraction problem over public-facing text, where LLMs are used to convert news articles, official outbreak bulletins, and social-media posts into structured event records. For early outbreak detection, few-shot prompting with frontier models and domain-adapted ensembles outperform rule-based baselines on WHO Disease Outbreak News and global news streams, with several systems framing extraction as multilingual or schema-driven event extraction (Bhadelia et al. 2026,Consoli et al. 2025,Consoli et al. 2024,Hong et al. 2023,Parekh et al. 2024). Social media platforms serve as complementary real-time biosensors: fine-tuned systems have predicted COVID-19 surges 7.63 days ahead of official reports (Xie et al. 2025) and generated influenza alerts in Wales weeks before primary care consultations (Sheridan et al. 2025), while PH-LLM scales infoveillance across dozens of multilingual classification and extraction tasks (Zhou et al. 2025) and multilingual symptom detection extends surveillance coverage into low-resource languages (Jannah et al. 2025). Beyond infectious diseases, the same extraction-centric pattern has been adapted to injury and behavioral surveillance, from agricultural injury narratives (Muller et al. 2025) to child maltreatment case reports analyzed with local, privacy-preserving models (Stoll et al. 2025) and large-scale Reddit analysis of opioid-use-disorder discussions (Testagrose et al. 2025). A notable recent trend is the explicit push toward deployment-ready architectures, including Health Sentinel for national-scale monitoring (Pant et al. 2025) and multi-agent swarms designed for autonomous, continuous intelligence gathering (Wattamwar and Akwafuo 2026), signaling a shift from isolated extraction studies toward operational early-warning pipelines.

Clinical and mortality surveillance. A second strand moves inside the health system itself, extracting infection and mortality signals from electronic health records, emergency department notes, and civil registration narratives. On the infection side, zero-shot GPT-4 achieves strong symptom identification across multi-site emergency department notes (McMurry et al. 2025), and combining LLM-based case extraction with spatiotemporal scan statistics detects localized outbreaks earlier than case-count-only baselines (Goncalves et al. 2025); complementary studies show that integrating LLMs with clinical expertise can support detection of catheter-associated and broader hospital-acquired infections (Pan et al. 2025,Perret and Schmid 2024), while in-hospital deployments are beginning to move from pilot to routine infection surveillance (Wu et al. 2025). On the mortality side, automating verbal autopsy is a flagship application in settings where civil registration is weak: recent work shows that LLMs substantially outperform traditional probabilistic algorithms for cause-of-death assignment in sub-Saharan Africa, both in general adult mortality (Carshon-Marsh et al. 2026,Wen et al. 2026) and in dedicated systems such as LAVA for language-model-assisted verbal autopsy (Chen et al. 2025). Other work targets structured mortality coding and high-stakes injury data, including constrained decoding for ICD-10 assignment (Coutinho et al. 2026), supervised classification of violent-death narratives with compact LLMs (Parker 2025), and improved extraction of overdose-related drugs from death investigation reports (Funnell et al. 2026,Omaki et al. 2025). Across this strand, LLMs are increasingly positioned as a way to convert legacy free-text mortality and clinical archives into machine-readable surveillance data, especially for diseases and populations that current structured systems fail to capture well.

Biological and safety surveillance. A third strand extends surveillance from text and records into biological and safety data streams. Sequence language models track pathogen evolution at population scale: CoVFit predicts SARS-CoV-2 variant fitness (Ito et al. 2025), ARNLE links autoregressive next-token prediction to variant prevalence (Liu et al. 2025), and METAGENE-1 demonstrates that a metagenomic foundation model can detect pathogens from wastewater without species-specific references (Liu et al. 2025). Beyond viruses, related models have been applied to predicting antibiotic resistance from genomic sequences (Kim and Kim 2025) and to broader biological-sequence analysis for infectious-disease research (Kaur and Butt 2025). For post-market drug and vaccine safety, a parallel literature uses fine-tuned LLMs to extract adverse events and vaccine-reaction relationships from clinical text and triage notes, outperforming zero-shot prompting and prior NLP baselines (Daluwatte et al. 2024,Khademi et al. 2025,Khademi et al. 2024,Li et al. 2024,Li et al. 2025), while LLMs are also being evaluated for upstream pharmacovigilance tasks such as literature screening and structured-query generation (Laurence et al. 2025,Painter et al. 2025). More complex reasoning tasks, however, remain challenging: off-the-shelf LLMs struggle on AEFI causality assessment and case-level adjudication (Abate et al. 2025). This strand also places stronger emphasis on safeguards, including semantic guardrails for multilingual pharmacovigilance (Hakim et al. 2025) and audits of sociodemographic bias and prompt-induced suggestibility in drug-safety decisions (Liu et al. 2025).

Across these three strands, T1 is dominated by the Sensor role, with LLMs primarily used to convert heterogeneous, unstructured inputs into structured surveillance signals. Adaptation strategies cluster along a clear axis: prompting and few-shot use of frontier models are typically sufficient for broad extraction over public text and outbreak news, whereas fine-tuning of open or compact models becomes critical for highly structured, safety-critical, or privacy-sensitive tasks such as pharmacovigilance, mortality coding, and child-protection classification. Relative to traditional rule-based or supervised pipelines, LLM-based surveillance systems consistently improve flexibility, cross-source and cross-lingual coverage, and timeliness, with some studies reporting detection leads of days to weeks over official statistics. At the same time, several recurring weaknesses limit operational uptake: causal and protocol-based reasoning (e.g., AEFI adjudication) remains brittle, multilingual and LMIC coverage is still uneven, and most evaluations are retrospective rather than prospective. The main open problem is therefore not whether LLMs can extract surveillance signals at all, but how to build surveillance systems that are simultaneously accurate, timely, calibrated, multilingual, privacy-aware, and integrable into the workflows of real public health agencies.

4.2. Epidemic Forecasting & Modeling (T2)

This section reviews LLMs, time-series foundation models, and agentic systems for epidemic forecasting, mechanistic modeling, and public health decision support. Compared with standard clinical prediction, epidemic modeling must operate under sparse, revised, and spatially coupled surveillance data, together with shifting interventions and behavior. These challenges make T2 distinct from conventional time-series forecasting: beyond point accuracy, useful systems must remain robust to reporting delays and distribution shift, represent uncertainty, and support decisions whose consequences propagate across populations and regions.

Time-series epidemic prediction. Most T2 work extends standard time-series forecasting with either multimodal context or epidemiological structure. A first line of work augments epidemic trajectories with exogenous public health signals. PandemicLLM combines epidemic time series with policy, genomic, and spatial signals (Du et al. 2025); EpiLLM and MIFlu integrate mobility or contextual text into autoregressive forecasting (Gong et al. 2025,Moon et al. 2025); and fine-tuned LLMs improve longer horizon influenza prediction (Li et al. 2025). A second line introduces more explicit epidemiological inductive bias. Structurally informed approaches include compartment pretraining for epidemics(CAPE) (Liu et al. 2025), simulation-trained foundation models (Mantis) (Dudley et al. 2025), and zero-shot and general-purpose forecasters (Kalahasti et al. 2025,Panja et al. 2025). A third, lighter-weight use treats LLMs less as standalone epidemic models and more as forecasting assistants or off-the-shelf baselines (Huang et al. 2025,Liang et al. 2025). Taken together, these studies suggest that current progress comes less from replacing epidemic forecasting with generic language modeling than from enriching forecasts with broader context, transfer, or mechanistic structure. At the same time, much of the literature still emphasizes forecast performance on benchmark datasets, while leaving more operational questions, such as robustness to revisions, calibration under uncertainty, and generalization across outbreaks, regions, and diseases, less fully examined.

Epidemic simulation & agent-based modeling. Fewer papers use LLMs for simulation, and most do not replace classical epidemic simulators directly. Instead, one line of work uses LLMs to support simulator construction, specification, or iterative model design. ChatGPT supports iterative design of transmission models (Kwok et al. 2024), while EpiAgent formulates simulator construction as program synthesis with epidemiological flow-graph constraints (Datta et al. 2026). A complementary line embeds LLM agents directly in epidemic environments for policy reasoning (Aoki and Ghaffarzadegan 2026). These two directions serve different purposes: the former treats LLMs as interfaces for building or modifying epidemiological models, whereas the latter uses them as decision-making agents inside structured epidemic worlds. This distinction is important because simulation in public health is not only a prediction problem, but also a counterfactual and planning problem. The main challenge is therefore not simply generating plausible trajectories, but ensuring that natural-language flexibility remains aligned with epidemiological validity, intervention logic, and interpretable assumptions about transmission dynamics.

Emergency response & decision support. A third strand targets preparedness and operational decision-making. EPIWATCH provides open-source epidemic intelligence for early warning under incomplete surveillance (Quigley et al. 2025), while foundation model forecasting is being extended to counterfactual policy analysis (Kalahasti et al. 2025). LLM agents have been applied to coordinated intervention design across mobility-coupled regions (Shi et al. 2026) and adaptive policymaking in structured epidemic environments (Aoki and Ghaffarzadegan 2026). Perspective papers highlight both the promise and deployment risks of these systems (Dixit 2025,Lau et al. 2026,Rizzo et al. 2024). Relative to retrospective forecasting, this strand places stronger emphasis on actionability: systems are expected not only to estimate future epidemic states, but also to support scenario comparison, intervention timing, and policy coordination under uncertainty. This raises a higher bar for evaluation, since strong predictive performance alone does not guarantee reliable decision support when recommendations may alter behavior, resource allocation, or downstream disease dynamics.

Across these three strands, T2 is driven mainly by the Predictor and Simulator roles rather than the extraction-oriented Sensor role that dominates earlier parts of the public health pipeline. Current work develops along four broad directions: multimodal epidemic forecasting, mechanistically informed foundation models, LLM-assisted simulator construction, and agentic decision-making in structured epidemic settings. Compared with traditional forecasting pipelines, these systems broaden the range of inputs and planning functions that can be incorporated, but evaluation remains relatively narrow: many studies still prioritize forecast accuracy or proof-of-concept simulation over revision robustness, calibrated uncertainty, counterfactual validity, and real operational usefulness. The main open problem is therefore not merely forecasting epidemic curves more accurately, but building systems that couple flexible multimodal reasoning with epidemiological realism, reliable uncertainty quantification, and decision-ready support for public health practice.

4.3. Infodemiology & Health Communication (T3)

This section reviews how LLMs and related generative AI systems are used to monitor, interpret, and shape public health information ecosystems. T3 is broader and more socially embedded than classical epidemic forecasting: online discourse can function as an early population signal, the content of that discourse can itself become a target for intervention, and generative models can both repair and pollute the information environment. The literature therefore follows a maturity gradient from monitoring and classification, to evidence-grounded verification, to audience-specific intervention, and finally to governance of model-generated risk. This dual-use structure makes T3 one of the clearest settings in which LLMs are simultaneously public health tools and public health threats (De Angelis et al. 2023,Joseph et al. 2025,Morita et al. 2024,Saeidnia et al. 2026).

Misinformation detection & fact-checking. The largest T3 strand treats LLMs as sensors and auditors for health misinformation. Early work uses prompting or lightweight adaptation to classify claims, explain decisions, and combine machine judgments with human or crowd signals (Rostami and Hawamdeh 2025,Zong et al. 2025). A second wave moves beyond classification toward evidence-grounded verification: retrieval-augmented systems check COVID-19 and vaccine claims against scientific literature, while long-form fact-checking decomposes complex narratives into claims that can be individually verified (Huang et al. 2025,Li et al. 2025). A third wave turns the lens back on the models themselves. New benchmarks include multimodal and LLM-generated vaccine misinformation (Ahmad et al. 2025,Zhang et al. 2025), and adversarial audits show that health chatbots can be induced to generate or endorse disinformation under jailbreak, role-playing, or system-instruction attacks (Hussain et al. 2025,Modi et al. 2025). Thus, progress in T3.1 is not only a story of better misinformation detection, but also a warning that detection systems must be evaluated against the generative threats they help create.

Vaccine attitudes & hesitancy monitoring. A second strand focuses on vaccine confidence, where the technical trajectory runs from stance detection to behaviorally relevant communication. LLMs classify sentiment, hesitancy, and fine-grained anti-vaccine concerns across COVID-19, HPV, and cross-national mRNA vaccine discourse (Annan et al. 2025,Boatman et al. 2024,Kim et al. 2024,Liu et al. 2025,Mu et al. 2024,Xu et al. 2024). This line is beginning to connect digital traces to real-world outcomes: online vaccination behavior signals have been linked to population vaccination patterns (Oh et al. 2025), suggesting that LLM-coded discourse may complement official uptake surveillance. The field is also moving from measurement to intervention. Tailored pro-vaccine messages, counter-arguments, common-ground-based framing, and chatbot systems explore whether LLMs can translate population attitudes into audience-aware communication (Dhanuka et al. 2025,Hou et al. 2025,Stureborg et al. 2024,Xia et al. 2025). These studies make vaccine communication the most mature T3 area for testing whether offline language-model performance can produce measurable public health behavior change.

Public health discourse analysis. Beyond misinformation and vaccination, LLMs are increasingly used as qualitative-analysis engines for large public health corpora. Rather than predicting disease trajectories, these systems summarize themes, identify concerns, and structure discourse at a scale that would be infeasible for manual coding. Human-validation studies suggest that LLMs can support inductive thematic analysis and broad discourse classification at substantially lower cost than expert-only workflows (Deiner et al. 2024,Espinosa and Salathé 2024). Applications include climate-related public health discourse, Alzheimer’s disease infodemiology, cytomegalovirus awareness, and campaign evaluation (Belova et al. 2025,Dai et al. 2025,Rosebrock et al. 2026). The main methodological value of this strand is not simply automation: it turns otherwise diffuse public conversation into analyzable public health intelligence, while raising unresolved questions about representativeness, platform bias, and whether discourse changes translate into policy or health outcomes.

Substance use and tobacco surveillance. Substance-use and tobacco studies form a distinct infodemiology strand because they combine high-stigma topics, rapidly evolving language, and strong links to population harm. LLMs have been used to monitor opioid-related chatter, identify self-disclosures of use and recovery, expand slang lexicons, classify tobacco and e-cigarette sentiment, and audit adherence to cessation guidance (Carpenter and Altman 2023,Elmitwalli et al. 2024,Han et al. 2025,Mittal et al. 2025,Sidorov et al. 2025,Yang et al. 2024). This area provides unusually direct evidence that social-media signals can track real-world burden, with opioid discourse trends compared against external mortality patterns (Mittal et al. 2025). It also shows how T3 can move from passive monitoring to language intervention: stigma-detection and rewriting systems attempt to reduce harmful framing around substance use rather than merely describe it (Bouzoubaa et al. 2024). The limitations are equally clear: most studies remain U.S.-centric, text-only, retrospective, and weakly connected to downstream prevention or treatment programs.

Risk communication & health messaging. The most intervention-oriented part of T3 studies whether LLMs can generate useful, credible, and audience-appropriate responses. Evidence-grounded counterspeech uses retrieval-augmented and multi-agent frameworks to rebut misinformation while maintaining factuality and tone (Anik et al. 2025,Song et al. 2025). Broader corrective mechanisms include community notes, inoculation, myth correction, pro-vaccine message generation, common-ground tailoring, and chatbot-based information support (Gullison and Fu 2025,Hou et al. 2025,Laily et al. 2026,Malek et al. 2026,Reis et al. 2026,Stureborg et al. 2024,Wu et al. 2025). More generally, work on health-news quality evaluation, belief change in human–AI interaction, and adherence to public health guidance emphasizes that fluent communication is insufficient: public health communicators must be evidence-grounded, culturally and literacy appropriate, and operationally safe (Abroms et al. 2025,Liu et al. 2025,Lu et al. 2026).

Across these strands, prompting is the dominant adaptation strategy, fine-tuning is most useful for stance classification and domain-specific behavior labels, and RAG is emerging as the default architecture for fact-checking and counterspeech. The central gap is the transition from classification to intervention. Many systems report accuracy or F1, yet few assess downstream outcomes such as belief change, vaccination uptake, reduced stigma, or improved adherence to guidance. Evaluation is also fragmented: misinformation detection often relies on automated metrics, while substance-use and communication studies more often require human or expert judgment. Finally, the field remains narrow in language, geography, platform, and modality. Moving from strong offline classification to robust, multilingual, multimodal, and prospectively evaluated communication systems remains the defining open challenge for T3.

4.4. Health Equity, SDOH & Global Health (T4)

This section reviews how LLMs are used to identify social determinants of health (SDOH), audit bias and fairness, and extend health tools to low-resource settings. Health equity is a foundational public health principle, yet the social and structural factors driving disparities are among the hardest signals to measure at population scale: they are unevenly documented, embedded in free text, often invisible in structured fields, and shaped by language, geography, and the design of health systems themselves. LLMs are increasingly framed as a possible response to this measurement problem, both because they can read unstructured records at scale and because they can extend clinical and public health tools into linguistic and infrastructural environments that conventional NLP pipelines have largely ignored. At the same time, the same flexibility that makes LLMs attractive for equity research also makes them a new source of equity risk, motivating a parallel literature on bias auditing and equitable deployment.

SDOH information extraction. A first line of work uses LLMs as scalable extractors of social and structural risk factors that are otherwise locked inside free-text clinical notes (Ong et al. 2024). Instruction-tuned open models such as LLaMA have been used to extract a broad set of SDOH factors across multiple institutional EHR datasets, generalizing across documentation styles better than encoder-based clinical baselines, especially for class-imbalanced or under-represented factors (Keloth et al. 2025), while SBDH-Reader shows that prompt-engineered general-purpose LLMs can match this performance under independent cross-institutional validation without fine-tuning (Gu et al. 2025). To handle infrequent or sensitive categories, recent work proposes staged zero-shot pipelines that decompose extraction into context retrieval, relevance verification, and factor extraction, with the explicit goal of improving recall on rare SDOH factors that supervised baselines tend to miss (Wang et al. 2025), and synthetic note generation, in which LLM-generated clinical notes are combined with authentic data to train downstream classifiers (Gabriel et al. 2024). A complementary line analyzes the equity orientation of the broader research landscape itself, using LLMs to extract health-equity covariates from registered research projects and to characterize how often demographic and structural factors are studied at all (Nananukul and Kejriwal 2024). Together, these studies suggest that SDOH extraction has moved from feasibility studies toward multi-institution and increasingly deployment-oriented use, but most evidence is still drawn from US English-language EHRs, leaving generalization across health systems, languages, and documentation cultures as the central open question.

Bias detection and fairness auditing. A second strand explicitly audits whether LLMs reproduce or amplify existing health disparities, treating LLMs themselves as objects of equity evaluation. EquityMedQA introduces a 4,619-example adversarial benchmark and a multifactorial human assessment framework covering six dimensions of bias. It shows that adversarial prompts and diverse rater pools reveal substantially more bias than standard medical QA evaluations (Pfohl et al. 2024). EquityGuard takes a mitigation-oriented approach, using contrastive learning to reduce inequities that arise when clinically irrelevant socio-demographic factors are added to prompts for clinical trial matching and medical question answering (Ji et al. 2025). Domain-specific audits further identify concrete failure modes, including persistent anti-LGBTQIA+ bias across four mainstream LLMs, with inappropriate response rates of 43–65% (Chang et al. 2025), and intersectional bias in mental health reasoning that is amplified during multi-hop question answering (Haider et al. 2025). A complementary line of work uses LLMs to audit human rather than algorithmic bias: counterfactual analyses of emergency department triage notes show that female patients receive systematically lower severity ratings than male patients with otherwise identical clinical content (Adames et al. 2025). Perspective papers broaden the discussion from model development to deployment-stage risks and proactive uses, arguing that clinical documentation, equity datasets, and equitable access must be redesigned as LLMs become routine intermediaries in health care (Pierson et al. 2025,Singh et al. 2023). Overall, this strand marks a shift from generic capability benchmarks toward adversarial and explicitly equity-focused evaluation. However, most existing audits remain English-only, focus on a narrow set of clinical scenarios, and rarely connect observed biases to downstream clinical or population health outcomes.

Global health and low-resource settings. A third strand asks whether LLMs can meaningfully serve populations and health systems beyond high-income, English-speaking contexts. Reviews and roadmaps document persistent inequities at the research level, with the bulk of LLM-in-health work originating in HICs and disproportionately optimized for English-language clinical contexts, and propose multidimensional frameworks for equitable adoption in low- and middle-income countries (LMICs) (Chen et al. 2025,Ong et al. 2026). Benchmark efforts make this gap concrete: AfriMed-QA spans 60+ medical schools across 16 African countries and 32 specialties and shows that even strong general-purpose LLMs lag on USMLE-style questions in this setting, with biomedical-specialized models often underperforming general ones (Olatunji et al. 2024), while simulated-patient evaluations of ChatGPT for diseases disproportionately affecting LMICs report high diagnostic accuracy alongside frequent inclusion of unnecessary or harmful medications (Si et al. 2024). Linguistic adaptation is a second major theme: Vietnamese-specific fine-tuning with LoRA/QLoRA on 337K health prompt-response pairs improves communication quality while enabling on-premise deployment (Bui et al. 2025), and modular architectures such as L2M3 combine LLMs with machine translation to support multilingual medical knowledge for community health workers (Gangavarapu 2024). Applied evaluations move closer to real LMIC workflows: silent trials and prospective protocols evaluate LLM assistance to community health workers in Rwanda, including ambient listening of CHW–patient encounters in Kinyarwanda, where strong frontier models can approach CHW-level referral accuracy while smaller models fall sharply behind (Menon et al. 2025,Shimelash et al. 2026), and LLMs have been used to analyze open-ended survey responses about why children in the DRC fail to access vaccination services, achieving up to 96% accuracy with as few as 20–100 labeled examples (Burstein et al. 2025). Taken together, this work shifts the equity question from whether LLMs are biased on standard benchmarks toward whether they can be deployed in linguistically, infrastructurally, and clinically realistic LMIC settings without amplifying existing global health inequities.

Across these three strands, T4 is shaped by two distinct LLM roles: a Sensor role used to extract SDOH and equity-relevant signals from unstructured records, and an Auditor role used to evaluate, mitigate, or surface bias in models, datasets, and human decisions. Adaptation strategies follow this division: instruction tuning and prompt engineering dominate SDOH extraction, while bias auditing relies more heavily on adversarial benchmarks, counterfactual prompts, and contrastive learning, and global-health work increasingly couples lightweight fine-tuning with translation and CHW-facing deployment. Compared to standard NLP equity research, this literature is distinctive in its move toward intersectional, deployment-stage, and multilingual evaluation rather than purely demographic-attribute accuracy gaps. Yet several gaps persist: most extraction work remains tied to US EHRs and English text, most bias audits do not connect to measurable health outcomes, and most LMIC studies are still pilot- or protocol-stage with limited prospective evidence. The central open problem for T4 is therefore to develop equity-aware LLM systems whose accuracy, bias, and downstream impact are jointly evaluated under realistic linguistic, geographic, and clinical conditions, rather than at the level of individual benchmarks or single-site studies.

4.5. Population Intervention & Practice (T5)

This section reviews how LLMs are used for direct public health action: deploying interventions, supporting frontline workers, allocating scarce resources, screening for mental and behavioral health conditions, and shaping everyday health practice. Compared to surveillance and forecasting, T5 is the most user-facing part of the public health pipeline, and the people on the other side of the system are rarely clinicians: they are community health workers, smokers trying to quit, individuals navigating HIV-prevention services, socially isolated older adults, or recovery seekers turning to online forums for support. This shifts the technical problem from “can the model produce a correct answer?” to “can the model produce a credible, safe, and contextually appropriate response in an ongoing interaction with a non-expert user?”, and it makes deployment design, human oversight, and evaluation under real-world use central rather than peripheral. Most systems in T5 therefore take a Communicator-oriented form and pair generative capabilities with domain knowledge, clinical guidelines, retrieval augmentation, or expert-in-the-loop validation.

Digital health interventions. A first strand studies LLM-powered chatbots, conversational agents, and decision-support tools that deliver counseling, screening, or intervention content directly to end users. HIV prevention is one of the most active testbeds: the AHF retrospective cohort of more than 155,000 adults shows that AI-augmented communication is associated with significantly higher PrEP initiation, follow-up attendance, and appointment adherence, with the largest gains among younger and historically underserved patients (Narayan et al. 2026); “Your Choice” achieves a System Usability Scale of 92/100 in stigma-free PrEP eligibility assessment among South African adults (Govathson et al. 2026); CHIA, a GPT-4o-based bilingual PrEP chatbot grounded in motivational interviewing, scores high on accuracy and trustworthiness in internal testing while exposing significant Spanish-language performance gaps (Tao et al. 2026); and qualitative work with men who have sex with men and transgender women in KwaZulu-Natal shows that multi-agent HIV/mental-health chatbots are valued for privacy, anonymity, and human-like interaction, but penalized for slow responses, repetitive replies, and limited emotional rapport (Humphries et al. 2026). A second cluster of systems supports community health workers (CHWs) and frontline staff in LMIC settings: ASHABot, a WhatsApp-based RAG chatbot deployed with India’s ASHA workers, is treated as both a private channel for sensitive questions and a near-authoritative reference, with supervisors expanding the knowledge base in response to usage patterns (Ramjee et al. 2025); the SMART health GPT pilot uses RAG to provide ASHAs with locally relevant maternal-care information and reports that retrieval-grounded designs are preferred over fine-tuning for this use case (Al Ghadban et al. 2023); and PRIORITY2REWARD takes the same population further by translating health-worker preferences into reward functions for restless multi-armed-bandit allocation in maternal mHealth programs, allowing CHWs to specify equity priorities and immediately inspect the resulting policies (Verma et al. 2025). Closely related work uses LLM agents to simulate maternal-health beneficiary listening behavior with explicit uncertainty estimation and decision-focused evaluation, demonstrating how LLM predictions can shape allocation decisions rather than just describe them (Martinson et al. 2025). Behavioral health and substance use form a third cluster: BeFreeBot, a ChatGPT-based SMS smoking-cessation chatbot, achieves 70% engagement, 100% USPSTF guideline alignment, and 30% self-reported 7-day abstinence in a 23-person pilot (Abroms et al. 2025), while LLM fine-tuning with data downsampling raises message-intent F1 from 0.41 off-the-shelf to 0.91 in a deployed smoking-cessation mHealth program (Rahman et al. 2026); complementary studies use LLMs to mine Reddit for evolving barriers to opioid recovery, surfacing pandemic-era shifts from stigma-related obstacles to systemic barriers like treatment discontinuity (Ekanayake et al. 2025); and CHI 2023’s CareCall study of a deployed LLM chatbot for socially isolated individuals provides one of the most explicit accounts of the deployment trade-offs at this layer, including gains in holistic understanding and emotional support against challenges in unpredictability and reliability for vulnerable users (Jo et al. 2023).

Population mental health screening. A second, narrower but important strand uses LLMs as scalable screeners for mental and behavioral health signals in non-clinical text. Mental-LLM is representative of the technical pattern: by instruction-fine-tuning on mental-health data, Mental-Alpaca and Mental-FLAN-T5 outperform GPT-3.5 by 10.9% and GPT-4 by 4.8% in balanced accuracy on population mental-health prediction from online text, performing on par with task-specific state-of-the-art models while remaining substantially smaller (Xu et al. 2024). This style of work is appealing for public health precisely because clinical screening cannot reach the populations whose distress is most visible online, but it raises distinct concerns that do not arise in clinical NLP: known racial and gender biases in mental-health classification, the ethical status of repurposing non-clinical social-media content, and the absence of meaningful follow-up infrastructure when screening runs at scale. As a result, T5.2 currently sits at a tension point: technical performance is approaching a level where deployment is conceivable, but the governance and equity questions surrounding population-level mental-health screening from social media remain effectively unresolved.

Health promotion and literacy. A third strand examines broader health-promotion and public health practice applications, where LLMs are typically embedded in routine workflows rather than user-facing chat. This includes LLM-based identification of workplace violence and communication failures from healthcare safety reports, where physical-violence and verbal-abuse classification reaches F1 above 0.80 and 0.94 respectively and outperforms prior NLP techniques (Becker et al. 2025,Sridi and Brigui 2023); evaluation of LLMs as natural-language-to-code translators for aggregated public health data, in which 11 systems are benchmarked on Czech public-health analyses and shown to vary widely in reliability, with executable but incorrect code as a recurring failure mode (Klempir et al. 2025); and domain-specific reviews and questionnaires assessing LLM use in public health dentistry and disease prevention, where the literature still emphasizes opportunity-and-threat framings and curriculum readiness rather than empirical effectiveness (Angyal et al. 2025,Tiwari et al. 2023). A growing thread within this strand explicitly studies occupational and population-level practice settings, asking how LLMs reshape the role of public health practitioners themselves rather than only the content they produce (Vos et al. 2025).

Across these three strands, T5 is the part of the survey where “does this system work?” depends most directly on real users, real workflows, and real time. A few patterns are consistent. First, evaluation has moved beyond accuracy: the most informative T5 studies report engagement, behavioral outcomes (e.g., PrEP initiation, abstinence, referral accuracy), or qualitative acceptance from end users, and they increasingly draw participants from the populations targeted by the intervention rather than from convenience samples. Second, retrieval grounding, expert-in-the-loop validation, and modest fine-tuning on program-specific data (rather than larger generic models) tend to be the design choices that move pilots toward deployment, especially for CHWs and underserved populations. Third, equity and safety concerns are not a separate workstream: they appear inside almost every deployment study, from Spanish-language gaps in HIV chatbots to bias in population mental-health prediction. The main open problem for T5 is therefore to move from short, single-site pilots toward sustained, multi-site, outcome-evaluated deployments in which intervention effectiveness, equity, and operational safety are measured jointly under everyday conditions of use.

4.6. Governance, Ethics & Policy (T6)

This section reviews how LLM-based public health systems are governed, evaluated, and deployed responsibly. As these tools move from research prototypes toward operational use in surveillance, decision support, and policy work, governance questions become unavoidable: who should benchmark these systems, what classes of risk should be measured, how should hallucinations and biases be surfaced before they shape resource allocation, and what institutional arrangements are needed before deployment is responsible? Compared with preceding task categories, T6 focuses less on what LLMs can do in principle and more on how they should be built, benchmarked, and integrated in practice. We organize this literature into three strands: governance-oriented system design, evaluation and policy-readiness benchmarking, and deployment and implementation science.

Ethics, safety, and risk frameworks. A first strand studies governance-oriented system design, treating the system architecture itself as the locus of accountability. One sub-direction democratizes mechanistic and modeling tools that previously required expert programming: an LLM-based assistant grounded in the open-source CMS framework lets users generate, simulate, and analyze compartmental disease models from natural-language descriptions, with few-shot EMODL examples dramatically reducing zero-shot syntactic errors (Proctor and Chabot-Couture 2024). A closely related sub-direction builds multi-agent research and response systems with explicit roles and validation steps. EpidemIQs uses a dual-agent (scientist + task-expert) architecture to autonomously conduct epidemic research from literature review through simulation to manuscript generation, achieving 79% task success across diverse scenarios at roughly $1.57 per study (Samaei et al. 2026), while EpiPlanAgent integrates task decomposition, knowledge grounding, and plan simulation/validation to automate digital emergency response planning with expert-in-the-loop checks on guideline alignment (Mao et al. 2025). Generative agent-based modeling extends this paradigm into population-level behavioral simulation: agents in epidemic environments spontaneously quarantine when sick or self-isolate when cases rise, producing emergent multi-wave dynamics that begin to resemble real pandemics (Williams et al. 2023), and “Infected Smallville” uses the same machinery to test behavioral immune system hypotheses by manipulating disease threat in agent settings (Choi et al. 2025). A fourth sub-direction targets disaster and outbreak response: DisasterResponseGPT generates plans of action grounded in domain guidelines that are comparable to human-authored ones while reducing turnaround from hours to seconds (Goecks and Waytowich 2023), and broader review and commentary work argues that the field underinvests in integrating LLM-based agents into outbreak analytics pipelines and disaster-management workflows more generally (van Hoek et al. 2024,Xu et al. 2025). Finally, several proposals reframe public health infrastructure itself around retrieval and architectural grounding: MEGA-RAG combines multi-source evidence retrieval, cross-encoder reranking, and discrepancy-aware refinement to cut hallucination rates by more than 40% on public health QA (Xu et al. 2025), and a position paper argues that public health systems should embrace a multi-layered epidemic early-warning architecture coupling distributed LLM agents, centralized analytics, and regionally enriched knowledge bases (DENG and Jin 2025). Across these systems, evidence grounding, expert checkpoints, and modular validation are recurring design commitments, even though most evaluations remain proof-of-concept and real-world emergency deployment is still rare.

Policy analysis and evaluation. A second strand asks the more direct question of whether current LLMs are fit to participate in public health reasoning and policy support, and answers it by building benchmarks that probe progressively harder tasks. EpiQAL, the first diagnostic benchmark for epidemiological question answering, partitions tasks into factual recall, multi-step inference, and conclusion reconstruction, and shows that current model rankings shift across subsets, that scale alone does not predict success, and that strong general LLMs can still fail systematically on evidence-grounded epidemiological reasoning (Wei et al. 2026). PubHealthBench, built from 8,000+ questions derived from 687 UK government public-health guidance documents, finds that frontier proprietary models exceed 90% accuracy on multiple-choice items, outperforming humans with cursory search-engine use, but drop below 75% on free-form responses where deployment risk is highest (Harris et al. 2025), and an open-source benchmarking platform combining a Ruby/Rails prediction prototype with an R Shiny evaluation app shows high task- and dataset-level variability across LLMs even when overall accuracy looks competitive (Espinosa et al. 2025). Application-oriented evaluations make these limitations concrete in policy-relevant settings: LLMs evaluated on childhood lead-testing resource allocation average only 0.46 accuracy and frequently overlook the highest-risk neighborhoods, citing outdated data and stereotypical narratives in place of current epidemiological evidence (Afane et al. 2025); bespoke RAG pipelines for public-health evidence reviews achieve 90%+ acceptable accuracy on objective study-design fields but fall to 54% or below on the outcome data that matters most for synthesis (Simmons et al. 2025); retrieval-augmented text-to-SQL pipelines materially improve epidemiological querying over EHR/claims data but remain insufficient for unsupervised use (Ziletti and DAmbrosi 2024); and SatHealth demonstrates that satellite-derived environmental factors, when fused with claims data and SDOH indices, substantially improve regional public-health modeling, illustrating the more general point that benchmarks confined to text leave critical multimodal signals out of view (Wang et al. 2025). Taken together, this literature pushes back against unsupervised deployment of LLMs in high-stakes policy settings, but it also helps locate the boundary: structured factual recall is increasingly tractable, while free-form reasoning, resource allocation, and outcome-level evidence synthesis still require careful human supervision and grounded retrieval.

Deployment and implementation science. A third strand turns from system design and benchmarking to the institutional and regulatory conditions under which LLMs can be safely operationalized in public health. A risk taxonomy and reflection tool for LLM adoption, derived from focus groups with public health professionals and individuals with lived experience across vaccines, opioid use disorder, and intimate partner violence, identifies four distinct risk dimensions, individual behaviors, human-centered care, information ecosystems, and technology accountability, and translates them into reflection questions that public health programs can use before adopting LLMs (Zhou et al. 2025). A complementary cross-sectional audit of the OpenAI GPT Store finds 1,055 health-related custom GPTs serving over 360,000 cumulative conversations, with 624 (59%) using healthcare-professional titles in their names despite none being approved by FDA, EU MDR, or TGA: an empirical case study of unregulated “role creep” in consumer-facing health AI (Chu et al. 2025). Geographic and global-health blind spots are also documented systematically: an evaluation of generative AI for health-policy identification across all UN Member States reports 78% concordance with expert datasets for vaccination policy but markedly worse performance in African, Southeast Asian, and Eastern Mediterranean regions, raising the prospect that uncritical deployment could entrench global inequities (Wilson et al. 2024). Against this backdrop, AMR-Policy GPT provides one of the clearest worked examples of a governance-aware, retrieval-grounded deployment: built on a multilingual database of antimicrobial-resistance national action plans from 146 countries, it answers AMR policy queries with traceable references and explicit One Health framing (Chen et al. 2025). Position pieces tie these threads together by arguing for validation-board frameworks for public health AI (Hattab et al. 2025), systematic AI literacy in public health curricula (Acosta 2025), and clearer institutional accountability for generative AI in scientific communication and policy work (Yoga Ratnam 2025), alongside more skeptical perspectives on whether epidemiology can in fact be automated given current hallucination, dataset-access, and reliability constraints (Bann et al. 2026).

Across these three strands, T6 is dominated by the Auditor and Simulator roles: LLMs are evaluated as candidate components of public health infrastructure, audited for hallucination, bias, accountability, and privacy risks, and embedded in agentic systems whose architectures are themselves the unit of governance. Several patterns are consistent. First, retrieval grounding, evidence traceability, and multi-agent validation are emerging as default design choices for any system intended to support real public health reasoning. Second, frontier models perform well on structured benchmarks but degrade sharply on free-form reasoning, resource allocation, and LMIC settings, exposing a gap between benchmark accuracy and decision-readiness. Third, governance-oriented work increasingly couples technical evaluation with institutional and regulatory analysis, including risk taxonomies, role-creep audits, validation boards, and AI literacy in public health training. The main open problem is therefore not whether LLMs can produce useful public health artifacts, but how to build the evaluation, oversight, and deployment infrastructure that allows their use in high-stakes settings to be responsibly accountable rather than only technically impressive.

5. Discussion and Conclusions

Our analysis highlights a central tension: current LLMs are designed mainly for document-level language processing, whereas public health reasoning is population-level, spatiotemporal, and socially embedded. This mismatch is not merely technical; it shapes what current systems do well, what they still miss, and where the field is structurally imbalanced. Across our taxonomy, the literature is strongest in extraction-oriented sensing and communication-heavy tasks, while forecasting, simulation, and deployment-oriented evaluation remain comparatively underdeveloped. More broadly, public health tasks form an operational pipeline rather than a collection of isolated benchmarks: surveillance signals feed forecasts, forecasts influence communication and intervention, and all of these ultimately affect governance and resource allocation. As a result, weaknesses at one stage can propagate downstream. We summarize the resulting research agenda in four directions.

Cross-modal integration of epidemiological data. Key non-textual signals—epidemic time series, genomics, mobility, and wastewater—remain largely siloed. Initial alignment efforts combine time series with policy (Du et al. 2025), integrate compartmental priors (Liu et al. 2025), or map infection dynamics to language (Gong et al. 2025). Future progress requires multimodal epidemiological foundation models with mechanistic grounding to handle distribution shifts and emerging outbreaks (Kwok et al. 2024). More fundamentally, public health forecasting and preparedness cannot be reduced to text understanding with auxiliary covariates appended at the end. Epidemic trajectories evolve under feedback from interventions, behavior, seasonality, and spatial coupling, while genomic and environmental signals often operate on different temporal and biological scales. A central challenge is therefore representation: future systems must jointly encode textual guidance, temporal dynamics, spatial interaction, and mechanistic constraints, rather than treating these modalities as loosely coupled inputs. This also changes what counts as success. It is not enough to improve retrospective benchmark performance; models must remain robust under reporting delays, revision, regime shifts, and emerging pathogens. In this sense, the next generation of public health LLM systems will likely look less like standalone chat models and more like multimodal reasoning systems grounded in epidemiological structure.

The dual-role paradox: LLMs as both tool and threat. The same models used to detect misinformation can also generate it at scale (De Angelis et al. 2023,Hussain et al. 2025). This risk is already visible in system-instruction attacks on health chatbots (Modi et al. 2025) and in the difficulty of distinguishing AI-generated from human-authored health misinformation (Zhang et al. 2025). In public health, such failures can propagate through information networks and undermine trust. This makes safety a population-level problem, motivating retrieval-grounded systems tied to evolving guidance (Xu et al. 2025) and literacy-aware counterspeech generation (Song et al. 2025). The key implication is that safety in public health cannot be framed only as single-response harmlessness. A fluent but subtly misleading model may distort risk perception, reduce adherence to guidance, or amplify uncertainty during outbreaks, even when each individual interaction appears plausible in isolation. This is especially important because T3 and T5 systems are often designed to operate at scale, where seemingly small per-instance errors can accumulate into large population-level effects. The field therefore needs a broader safety lens that includes provenance, susceptibility to adversarial prompting, consistency with evolving guidance, and downstream effects on trust and behavior. More generally, the same generative flexibility that makes LLMs attractive for surveillance summarization, public communication, and triage support also creates new attack surfaces. Public health applications should therefore be evaluated not only for utility, but for their potential to accelerate the very harms they are intended to mitigate.

Population-scale alignment and global health equity. Public health demands equitable population-level performance. However, current systems often reproduce disparities, showing higher adverse-event risks for disadvantaged patients (Liu et al. 2025) and racial gaps in mortality classification (Parker 2025). Globally, an English-centric reliance on proprietary models (Chen et al. 2025) creates a "multilingual penalty," degrading performance across Southeast Asian languages (Zhou et al. 2025), Rwanda (Shimelash et al. 2026), and other African settings (Wilson et al. 2024). Mitigating these challenges requires equity-aware alignment data, subgroup-sensitive rewards, improved cross-lingual transfer, and smaller on-premise models tailored for low-resource deployment (Ji et al. 2025,Pfohl et al. 2024). Yet equity in public health should be understood as more than a post hoc fairness audit. In this domain, inequity can emerge at multiple levels: in what data are collected, which populations are represented in benchmarks, what languages are supported, which institutions can afford deployment, and how model outputs are translated into policy. The current literature still leans heavily toward English-centric and resource-rich settings, even though many of the most consequential public health use cases involve multilingual communication, weak surveillance infrastructure, and low-resource operational environments. This means that alignment for public health must be population-aware from the outset. Models must be evaluated across regions, subgroups, and health systems, and they must support local adaptation rather than assuming that a single frontier model can generalize universally. In practice, this also elevates questions of reproducibility, accessibility, and deployment sovereignty: open and smaller models may be less capable in some settings, but they can be essential for privacy-preserving, locally governed, and infrastructure-constrained public health use.

Evaluation beyond static benchmarks: synthetic societies. Static QA benchmarks are insufficient for tasks involving sequential decisions, delayed feedback, and collective behavior. Generative agent-based modeling offers a promising alternative: LLM agents can quarantine in simulated epidemics (Williams et al. 2023), reduce social contact under disease threat (Choi et al. 2025), and improve coordinated intervention outcomes (Shi et al. 2026). A natural next step is synthetic societies that can serve as population-level stress tests for scenarios that are difficult to evaluate in the real world. More broadly, public health evaluation must move beyond single-step correctness toward decision relevance. A system that performs well on extraction, classification, or QA may still fail when deployed in settings with delayed surveillance, conflicting guidance, heterogeneous populations, or intervention feedback. This suggests a need for evaluation protocols that are longitudinal, subgroup-aware, and operationally grounded. Useful benchmarks should test not only whether a model can answer correctly, but whether it remains calibrated under uncertainty, robust to temporal drift, consistent across linked tasks, and safe under iterative human–AI interaction. Synthetic societies are promising precisely because they provide a bridge between offline benchmarking and real-world deployment: they can expose emergent failure modes, coordination breakdowns, and policy side effects that are invisible in static test sets. At the same time, they should be viewed as complements rather than substitutes for real deployment studies, prospective trials, and implementation science. The long-term goal is not merely more realistic simulation, but evaluation frameworks that reflect how public health decisions unfold over time and across populations.

Conclusion. This survey organizes 192 papers through a two-dimensional taxonomy spanning six public health tasks and five LLM roles. Overall, the field is strongest in extraction-oriented sensing but remains early in forecasting, simulation, alignment, and multilingual deployment. Our synthesis suggests that the main bottleneck is not simply the lack of larger or more capable language models, but the mismatch between document-centered architectures and the requirements of population health reasoning. Public health systems must integrate heterogeneous modalities, support action under uncertainty, remain equitable across languages and settings, and be evaluated in ways that reflect real downstream consequences. Public health should therefore be viewed not only as an application domain, but also as a demanding testbed for multimodal learning, equitable alignment, and realistic evaluation. The accompanying repository at https://github.com/publichealthllm/llms-in-public-health-survey is maintained to track this rapidly evolving literature.

Appendix A.1. Additional Background on Large Language Models

Large language models: Additional background for PH professionals. A large language model (LLM) is a neural network with up to hundreds of billions of parameters, trained on large text corpora to predict and generate language. Modern LLMs are built on the transformer architecture (Vaswani et al. 2017), whose self-attention mechanism captures long-range dependencies and is well suited to public health tasks involving lengthy guidelines, surveillance reports, and heterogeneous evidence. Most systems in this survey begin with a foundation model (Bommasani et al. 2021) and adapt it to a specific public health task. Two main pre-training paradigms dominate: autoregressive models, such as the GPT family, predict the next token and are especially effective for generation and dialogue (Achiam et al. 2023,Brown et al. 2020), while masked language models such as BERT reconstruct missing tokens and are often used for classification and extraction (Devlin et al. 2019); encoder–decoder models such as T5 combine both directions (Raffel et al. 2020). Open-weight families such as LLaMA also matter for reproducibility and deployment in resource-constrained settings (Touvron et al. 2023). Since pre-training alone does not produce models that reliably follow instructions or avoid harmful outputs, alignment is typically added through supervised fine-tuning and reinforcement learning from human feedback (Ouyang et al. 2022). In practice, public health applications mainly rely on four adaptation strategies: prompting, including few-shot and chain-of-thought prompting (Brown et al. 2020,Wei et al. 2022); fine-tuning, including parameter-efficient methods such as LoRA (Hu et al. 2022); retrieval-augmented generation, which grounds responses in external documents (Lewis et al. 2020); and agentic systems, which connect LLMs to tools to execute more complex workflows.

Appendix A.2. Survey Methodology and Literature Selection

Search strategy. We conducted a systematic literature search following PRISMA guidelines. We queried five major academic databases—PubMed, IEEE Xplore, ACM Digital Library, Google Scholar, and Semantic Scholar—as well as preprint servers (arXiv, medRxiv) for papers published between January 2023 and early 2026. The search strategy combined two categories of keywords: model-centric terms (Large Language Model, LLM, Foundation Model, GPT, Generative AI) and domain-centric terms (Public Health, Epidemiology, Surveillance, Pharmacovigilance, Infodemiology, Health Equity, Verbal Autopsy, Wastewater Surveillance). To ensure comprehensive coverage, we supplemented the database queries with targeted venue-specific searches across both AI conferences (ICML, ICLR, NeurIPS, ACL, EMNLP, NAACL, AAAI, CHIL, ML4H, FAccT) and public health journals (Nature Medicine, The Lancet, AJPH, IJE, JMIR, BMC Public Health, Frontiers in Public Health). Forward and backward citation tracking was performed on key seed papers to capture additional relevant work. Our initial search yielded 281 candidate papers after deduplication.

Screening and boundary test. Because the boundary between individual-level clinical AI and population-level public health is often blurred, we applied a three-step boundary test to every candidate paper:

1.

Institutional attribution test: Does the work’s primary application reside within the public health system (e.g., CDC, WHO, health departments, schools of public health, epidemiological registries)?

2.

Core function test: Does the core task fall within one of the recognized core public health functions—surveillance, disease prevention, health communication, health equity, population intervention, or governance—as defined by established frameworks such as the CDC’s 10 Essential Public Health Services?

3.

Substitution test: If removed from the public health context, would the work retain independent significance in another domain? If yes, its primary identity is likely outside public health.

A paper must satisfy at least the first two criteria. We explicitly exclude individual-level clinical applications (diagnosis, treatment, prognosis), drug discovery, consumer-level health assistants, and medical education unless targeting public health training or community health worker capacity building. Screening was performed independently by two reviewers, with disagreements resolved through discussion. This process led to the exclusion of 35 candidate papers and the reclassification of 10 borderline cases (including individual-level mental health prediction systems and personalized ADR prediction tools). A subsequent scope audit removed an additional 10 papers, yielding a final corpus of 192 papers. The six-category task taxonomy (T1–T6) and five LLM roles were then inductively developed from this final corpus, as described in Section 4. All paper URLs and metadata were manually verified for correctness.

Data extraction. For the 192 included papers, we performed systematic data extraction from full text. For each paper, we recorded structured metadata including: primary PH task (T1–T6), LLM functional role (Sensor, Predictor, Communicator, Simulator, Auditor), technique tags (Prompting, Fine-tuning, RAG, Agentic, etc.), base LLM(s), model access (Open/Proprietary), data source type, geographic focus, language(s), evaluation design, and deployment stage. We additionally recorded narrative fields—including a one-line summary, key results, and the specific public health challenges addressed—to support the qualitative synthesis in each task section. Task-specific fields (e.g., surveillance modality for T1, prediction horizon for T2, platform analyzed for T3) were recorded where applicable.

Appendix A.3. The Public Health Data Ecosystem

Table A1 summarizes the data modalities, language distribution, and evaluation benchmarks identified across the 192 surveyed papers.

Table A1. Distribution of data modalities and language coverage across 192 surveyed papers. Percentages for data sources sum to more than 100% because some papers use multiple modalities.

Data source category	% of papers
Social media (Twitter/X, Reddit, Weibo)	25.5
Scientific literature & guidelines	16.1
Simulated / agent-based data	11.5
Clinical text (EHR notes, discharge summaries)	9.9
Official reports (WHO DONs, CDC bulletins)	9.4
Epidemiological time-series	6.8
Death certificates & verbal autopsy narratives	4.2
Policy documents (e.g., OxCGRT)	3.6
Genomic / environmental (wastewater, sequences)	3.1
Consumer reviews	2.6
Surveys & questionnaires	7.3
Language coverage
English-only	80.7
Monolingual non-English	14.1
Genuinely multilingual (≥3 languages)	5.2

Table A1. Distribution of data modalities and language coverage across 192 surveyed papers. Percentages for data sources sum to more than 100% because some papers use multiple modalities.

Data source category	% of papers
Social media (Twitter/X, Reddit, Weibo)	25.5
Scientific literature & guidelines	16.1
Simulated / agent-based data	11.5
Clinical text (EHR notes, discharge summaries)	9.9
Official reports (WHO DONs, CDC bulletins)	9.4
Epidemiological time-series	6.8
Death certificates & verbal autopsy narratives	4.2
Policy documents (e.g., OxCGRT)	3.6
Genomic / environmental (wastewater, sequences)	3.1
Consumer reviews	2.6
Surveys & questionnaires	7.3
Language coverage
English-only	80.7
Monolingual non-English	14.1
Genuinely multilingual (≥3 languages)	5.2

Unlike clinical AI, which primarily operates on structured EHRs and medical imaging, public health applications draw on a remarkably heterogeneous data landscape. Three patterns stand out. First, social media dominates (25.5%), reflecting the field’s emphasis on real-time population-level signal extraction—yet nearly all studies use Twitter/X or Reddit, leaving platforms with large global user bases (TikTok, WhatsApp, WeChat) almost entirely unexplored. Second, several modalities have no clinical parallel: epidemiological time-series, verbal autopsy narratives, wastewater metagenomics, and government policy documents are unique to public health and require specialized tokenization, alignment, or constrained decoding strategies that standard clinical NLP does not address. Third, the linguistic imbalance is severe: 80.7% of papers operate exclusively on English data despite public health being an inherently global endeavor, and only PH-LLM systematically evaluates across 29 languages, revealing significant performance degradation for Southeast Asian languages.