What AI Surveillance Actually Delivers for Public Health

The most revealing number in AI for public health surveillance is not a model score. It is the distance between what the tools can now demonstrate and what health agencies report using. In ASTHO’s 2025 Profile data, 14% of U.S. health agencies reported using AI for surveillance, while 30% reported using AI for administrative tasks; the survey sample for these AI-use questions was n=32–44, so the figures should be read as a clear signal from responding agencies rather than a census of the entire public health system.[1]

That split is not hard to understand. Administrative work can tolerate more review, more correction, and more distance from immediate public health action. Surveillance is less forgiving. A signal may be early, but someone still has to decide whether it is real, whether it matches laboratory data or syndromic feeds, whether it reaches the right jurisdiction, and whether the receiving team has the staff and authority to act before the window closes.

The evidence is still worth taking seriously. A 2025 narrative review by Mendes and colleagues describes machine learning applications across outbreak detection, disease forecasting, anomaly detection, and automated processing of health and non-health data streams, with studies using methods including long short-term memory models, random forests, and deep neural networks.[2] The better question is no longer whether AI can find patterns in public health data. It is where those patterns are strong enough, timely enough, and operationally connected enough to change what agencies do.

What Counts as AI Surveillance

AI surveillance is often discussed as if it were one tool. In practice, the label covers several different jobs. Some systems scan media, official reports, and open-source signals for early outbreak indications. Others forecast disease trends from historical surveillance data. Some extract signals from emergency department visits, laboratory feeds, or clinical text. Others help integrate fragmented data streams so that public health staff can see a coherent pattern sooner.

Those distinctions matter because a good result in one task does not automatically transfer to another. A model that improves a short-term forecast for a specific infectious disease is not the same as a platform that can reliably route outbreak alerts to a local health department. A system that extracts more signals from open-source data may still increase workload if it cannot distinguish actionable events from noise. The field’s strongest evidence often sits at the task level, while the public health promise depends on the whole chain.

Earlier Signals Are Real, but They Are Only the First Hand-Off

BlueDot remains the example people reach for because it is concrete. The system reportedly identified a signal related to COVID-19 in late 2019, nine days before the official announcement, using automated analysis of sources such as news reports and airline ticketing data.[3] That is the kind of interval public health people notice. Nine days can matter when an outbreak is moving through travel networks and official case definitions are still settling.

But the BlueDot case also shows why detection should not be treated as the whole outcome. An early signal has to be interpreted, verified, communicated, and acted on. It may arrive before laboratory confirmation is available. It may point to a location that lacks resources for investigation. It may be one of many alerts competing for attention. The public health value begins when an agency can turn the alert into a decision, not when a platform records that the alert existed.

The broader pandemic-preparedness literature describes similar ambitions for platforms such as the Epidemic Intelligence from Open Sources initiative, or EIOS, which supports event-based surveillance by helping users monitor and assess public health threats from open information sources.[3] The same literature points to the potential value of integrating nontraditional streams, including genomic and mobility data, into surveillance workflows.[3] That integration is promising, but it also raises the operational difficulty. Every additional stream brings questions about timeliness, provenance, governance, and whether the signal means the same thing across jurisdictions.

Prediction Gains Are Encouraging When the Setting Is Specific

Forecasting studies provide some of the clearest measured improvements, because the task can be framed against observed disease activity. Mendes and colleagues describe work in Korea where long short-term memory models outperformed ARIMA models for chickenpox and scarlet fever prediction, with reported prediction accuracy improvements of 20%–26%.[2] That is a meaningful difference for a defined model comparison, disease pair, and setting.

It is also a narrower finding than many general claims about AI surveillance. The result does not mean LSTM models will outperform conventional methods for every disease, region, data feed, or forecast horizon. It means that, in this case, a recurrent neural network approach captured temporal patterns better than the comparator model. For public health agencies, that distinction is not academic. Forecasting is useful only when the prediction horizon matches a decision horizon: staffing clinics, preparing communications, prioritizing inspections, or targeting investigation resources.

The same caution applies to other model classes in the review. Random forests, deep neural networks, and other machine learning approaches can improve classification or prediction in selected surveillance tasks, but the review literature is still weighted toward task-specific and retrospective evaluation rather than prospective evidence that AI-supported surveillance changes outbreak control or population health outcomes.[2] Retrospective accuracy is useful. It is not the finish line.

Operational Platforms Sit Closer to the Mess

The most important surveillance systems are not always the ones with the cleanest model diagram. They are the ones forced to live with imperfect feeds, reporting delays, changing clinical behavior, and the daily work of public health interpretation. Recent reporting has profiled systems including SENTINEL, EpiWatch, and BioSense as examples of AI-enabled or AI-adjacent infectious disease surveillance work moving beyond a single laboratory demonstration.[4]

BioSense is especially useful as a reminder that surveillance modernization is not just model selection. It sits in the world of syndromic surveillance, where emergency department and other near-real-time data can help agencies identify patterns before confirmed diagnoses arrive. The promise is speed, but the work depends on whether data are complete enough, coded consistently enough, and available to the right analysts in time to support judgment.

Event-based systems such as EpiWatch and related open-source intelligence approaches occupy a different place in the surveillance stack. They can widen the field of view by scanning public information, but the output still has to be curated. More signals can mean more awareness; it can also mean more triage. If the staffing model assumes that AI eliminates review, the system is likely to disappoint the people asked to use it.

This is where the adoption number starts to look less like hesitancy and more like a capacity indicator. Agencies may be willing to experiment with AI, but surveillance adoption requires a stable technical and organizational layer: data sharing agreements, feed monitoring, quality checks, alert governance, analyst training, escalation pathways, and documentation that survives staff turnover.

Why Administrative AI Is Moving Faster

The 30% administrative-use figure in the ASTHO data deserves attention because it shows that public health agencies are not simply refusing AI.[1] They appear more willing to use it where the workflow is easier to supervise and the consequences of a missed nuance are less immediate than in outbreak detection. Drafting, summarization, grant analysis, and internal knowledge management can be reviewed before they affect the public.

CDC’s generative AI materials show why that use case is attractive. CDC reported saving more than 5,500 hours on grant analysis and estimated 41,000 staff hours saved through generative AI use, while also emphasizing enterprise principles around responsible implementation.[5] For an understaffed agency, those hours are not trivial. They can represent fewer evenings lost to document review and more time for analysis, follow-up, or coordination.

Still, administrative efficiency should not be quietly converted into surveillance effectiveness. Saving staff hours may free capacity that supports better surveillance work, but it does not prove earlier outbreak control, better case ascertainment, or reduced disease burden. It is adjacent evidence: important, practical, and not a substitute for outcome evaluation.

The Adoption Barrier Is Not One Barrier

The ASTHO findings point to several reasons surveillance AI has not become routine infrastructure. Among responding agencies, 64% identified lack of guidance as a barrier, 55% identified workforce skills, and data interoperability remained a major obstacle to adoption.[1] Those are not soft concerns. They are the conditions that determine whether a model can leave a pilot and enter daily surveillance practice.

Guidance matters because agencies need to know what level of validation is adequate, who owns an AI-generated alert, how false positives and false negatives are documented, and when human review is mandatory. Without that, each agency is left to invent its own rules or avoid higher-risk uses. That produces exactly the kind of uneven adoption that already characterizes much of U.S. public health technology.

Workforce readiness is equally concrete. Surveillance teams need people who can understand model outputs without treating them as either magic or noise. They need epidemiologists who can ask whether a signal is plausible, informaticians who can diagnose feed problems, and leaders who can decide when an alert warrants action. A platform cannot compensate for a hollowed-out analytic workforce.

Interoperability is the oldest problem in the room, and AI does not make it disappear. A model trained or tuned on one jurisdiction’s data may struggle when definitions, reporting practices, or data completeness differ elsewhere. If laboratory data, syndromic feeds, case reports, and genomic signals arrive at different speeds and in different formats, the AI layer inherits the disorder. It may even make the disorder harder to see if outputs look cleaner than the data underneath.

Bias and equity concerns sit inside these same operational questions, not outside them. Surveillance tools trained on uneven data can reproduce uneven visibility: communities with better access to care, more complete testing, or more consistent digital reporting may be easier for systems to see. The regulatory and governance problem is discussed more fully in ClinicalMind’s analysis of algorithmic bias in public health AI, but for surveillance teams the practical question is immediate: whose outbreaks become visible early, and whose remain late signals in an incomplete dataset?

Evidence Needs to Move Closer to Decisions

The current evidence base supports a measured conclusion. AI can detect relevant signals earlier in some contexts, improve prediction accuracy in specific disease-model comparisons, and help process information at a scale that manual review cannot easily match.[2][3] It also shows why a low adoption rate is not just a communications problem. The missing layer is implementation evidence: whether the alert reached the right unit, whether staff trusted it enough to investigate, whether the investigation changed timing or allocation, and whether that change improved public health outcomes.

A broader review of AI in public health describes the same tension between promise and challenge, including the need for governance, validation, privacy protections, and equity safeguards.[6] In surveillance, those safeguards are not add-ons after deployment. They shape whether agencies can use the system at all. The legal and policy environment is also fragmented, as ClinicalMind’s review of federal and state AI regulation in public health shows, and fragmented rules make multi-jurisdiction surveillance harder to standardize.

The next useful studies will look less like isolated leaderboards and more like surveillance evaluations. They will report alert volume, false-positive burden, analyst review time, escalation rates, integration with existing systems, and what changed after an alert. They will distinguish a model that retrospectively identified a pattern from a workflow that prospectively improved response. They will be honest when an AI system saves time but does not yet show disease-control impact.

This is not a case for waiting until every uncertainty is resolved. Public health surveillance has always worked with incomplete information. But AI systems need to be judged by the realities of the agencies expected to use them. A nine-day early signal is valuable if it enters a prepared system. A stronger forecast matters if it lines up with decisions that can be made in time. Administrative savings matter if they strengthen, rather than merely decorate, the surveillance workforce.

AI surveillance is therefore neither a speculative promise nor a mature public health layer. The evidence already shows real capability. The harder proof is now operational: agencies must be able to integrate, govern, interpret, and act on AI-generated signals reliably enough for those signals to matter outside early-adopter settings.

References

1. ASTHO 2025 Profile data, Association of State and Territorial Health Officials, 2025
2. Artificial intelligence in public health: a narrative review of applications, challenges, and future directions, Frontiers in Public Health, 2025
3. Artificial Intelligence for Pandemic Preparedness and Response: A Call for Equitable and Ethical Public Health Surveillance, PMC
4. Artificial intelligence infectious disease outbreaks, HealthBeat, July 1, 2025
5. Considerations for GenAI in Public Health, Centers for Disease Control and Prevention
6. Artificial intelligence in public health: promises and challenges, PMC