Last updated: October 21, 2026
Every outbreak starts small: a cluster of fevers in one hospital, a few unusual pneumonia cases, a spike in pharmacy sales nobody can explain. The question public health agencies have asked for decades is simple to state and hard to answer: can we see the wave coming before it breaks?
Machine learning has changed how that question is approached. Algorithms now read news reports in dozens of languages, scan flight data, combine weather records with case counts and blend the output of many competing models. Some of these tools have a genuinely useful track record. Others have failed in public and taught the field hard lessons. This guide looks at both, and shows you how to judge the accuracy claims you will read about them.
How we researched this article
We prioritised peer-reviewed studies, systematic reviews and official publications from the World Health Organization (WHO) and the U.S. Centers for Disease Control and Prevention (CDC). Company claims are labelled as such. Where a figure comes from a preprint (a study not yet peer-reviewed), we say so. We did not test any of the systems ourselves.
Two different jobs: detection and forecasting
«Predicting outbreaks» usually means one of two things, and mixing them up is the source of most hype.
Early detection (often called event-based surveillance or epidemic intelligence) means finding the first signs that something unusual is happening. Natural language processing sifts through large volumes of online text and flags reports that look like an emerging health threat.
Forecasting means estimating how a known disease will develop: how many cases, hospital admissions or deaths to expect next week or next month. Models learn from historical case counts and from drivers such as temperature, rainfall, mobility or school calendars.
From signal to action: how outbreak intelligence works
- Digital signals: news, bulletins, travel and case data are collected continuously.
- Automated detection: algorithms filter, translate and flag reports that look unusual.
- Epidemiological review: specialists check whether the signal makes scientific sense.
- Confirmation: national authorities verify it with laboratory and clinical data.
- Public health response: alerts, testing, travel advice or control measures follow.
Simplified overview. The automated step speeds up the process; the human and laboratory steps decide whether a signal is real.
The case that made headlines: Wuhan, December 2019
The best-known example comes from the start of COVID-19. According to a 2021 Hoover Institution review, the automated HealthMap system issued a public alert about unidentified pneumonia cases in Wuhan on December 30, 2019, the same day the human-moderated ProMED network circulated a report. The next day, the Toronto company BlueDot sent its clients a warning.
BlueDot says its platform uses natural language processing and machine learning to scan news and official reports in 65 languages, combined with global airline ticketing data. Using travel data, its team anticipated that the first exported cases were likely to appear in cities such as Bangkok, Seoul, Taipei and Tokyo, and published that analysis in the Journal of Travel Medicine in January 2020.
| Date | Event | Source |
|---|---|---|
| Dec 30, 2019 | HealthMap automated public alert; ProMED report on unexplained pneumonia in Wuhan | Hoover Institution |
| Dec 31, 2019 | BlueDot alerts its clients; WHO’s China office is informed of the cases | BlueDot / J Travel Med |
| Jan 4, 2020 | WHO reports the cluster publicly on social media | WHO |
| Jan 5, 2020 | WHO publishes its first Disease Outbreak News report with a risk assessment | CDC advisory citing WHO |
| Jan 8, 2020 | CDC issues a Health Alert Network advisory to US clinicians | CDC |
| Jan 9, 2020 | WHO reports that the cause is a novel coronavirus | WHO |
Two further points are easy to miss. These systems detected an outbreak that had already started; none predicted it before the first patients fell ill. And BlueDot’s reports are not produced by the algorithm alone: epidemiologists check each signal before it is sent to clients. An early alert only matters if someone verifies it and acts on it.
The cautionary tale: Google Flu Trends
The authors called this «big data hubris»: the assumption that huge datasets can replace, rather than supplement, traditional surveillance. The model latched onto search terms that happened to correlate with flu season, and it missed an off-season outbreak entirely. Changes to Google’s own search engine also shifted the data underneath it. The lesson still shapes the field: digital signals are an extra layer that must be anchored to clinical and laboratory data.
What forecasting looks like when it works
The strongest evidence comes from large, transparent forecasting collaborations. From April 2020, the US COVID-19 Forecast Hub, set up with CDC funding, collected weekly predictions from more than 90 academic, industry and independent groups. A 2022 evaluation in PNAS found that the combined «ensemble» forecast was the most consistently accurate for COVID-19 deaths at state and national level between April 2020 and October 2021.
The same study is also a reality check. About two-thirds of the 27 individual models it evaluated beat a simple baseline, and probabilistic error at a 20-week horizon was three to five times larger than at one week. A European hub covering 32 countries reached a similar conclusion: across 29 evaluated models, the ensemble outperformed most of its components and often all of them.
Dengue, a mosquito-borne disease closely tied to weather, is another active area, and the evidence is more mixed than headlines suggest. A 2025 systematic review reported that machine learning methods, particularly random forests, often outperformed classic statistical models such as ARIMA in the studies it included, with temperature and rainfall among the most important inputs. But an earlier review in PLOS Neglected Tropical Diseases, covering 99 models from 64 studies, found that around 20% reported no validation at all and only about 5% were tested on external data. A 2025 preprint from Bangladesh illustrates the gap: its hybrid model captured outbreak timing well but overestimated outbreak size.
How accurate are AI outbreak forecasts? Four metrics that matter
A single headline number rarely tells you whether a system is useful. These four measures give a fuller picture:
| Metric | What it tells you | Question to ask |
|---|---|---|
| Sensitivity | The share of real outbreaks the system detects | How many outbreaks did it miss? |
| False-positive rate | How often it raises alerts that are never confirmed | How many false alarms came with each true one? |
| Calibration | Whether stated probabilities match reality: events given a 70% chance should happen about 70% of the time | Were its probabilities tested against outcomes? |
| Forecast horizon | How far ahead it predicts and how error grows with time | How accurate was it at one week versus one month? |
The main systems compared
| System | Job | Access | What it is designed for | Public validation | Known limitations |
|---|---|---|---|---|---|
| BlueDot | Detection and spread estimates | Commercial | Scanning news and official reports, adding air travel data | Peer-reviewed travel-spread analyses; no independent audit of overall alert accuracy found | Proprietary methods |
| HealthMap | Detection | Public | Automated mapping of online outbreak reports | Methods described in academic publications | Depends on what gets reported online |
| WHO EIOS | Detection | Free for countries and eligible organizations | Shared open-source intelligence for national teams | Country-level assessments reported by WHO | Signals still need national verification |
| Forecast hubs (US and Europe) | Forecasting | Public | Combining many models into one ensemble | Peer-reviewed evaluations (PNAS 2022, eLife 2023) | Accuracy drops beyond a few weeks |
| Google Flu Trends | Nowcasting | Discontinued | Estimating flu activity from search queries | Independent evaluation in Science (2014) | Large overestimates in 2011–13 |
Where governments are investing
In October 2025, WHO launched version 2.0 of its Epidemic Intelligence from Open Sources (EIOS) system at the WHO Hub for Pandemic and Epidemic Intelligence in Berlin. The upgrade adds AI tools for automated analysis and signal detection and can process sources such as radio broadcasts, which are automatically transcribed and translated. First developed in 2017, EIOS was used by 120 countries and 30 organizations and networks by the end of 2025, according to WHO.
In the United States, the CDC launched its Center for Forecasting and Outbreak Analytics in April 2022, with initial funding from the American Rescue Plan, to turn modeling and outbreak analytics into faster decisions for federal, state and local health officials.
How to judge a claim that «AI predicted an outbreak»
- Did it predict the outbreak, or detect one that had already started?
- Was the warning timestamped and public at the time, or claimed afterwards?
- Which reference date is it being compared against?
- How many false alarms did the same system raise?
- Was the model tested on external data and compared with a simple baseline?
- How far ahead was the forecast, and how quickly did accuracy fall?
- Did a human expert review the signal before it was acted on?
Limitations
Outbreak prediction faces problems that better algorithms alone cannot fix. Reporting is uneven: regions with fewer journalists, laboratories or internet users generate weaker signals, so the places at highest risk can be the hardest to monitor. Disease dynamics change when people change their behavior, so a forecast can be «wrong» precisely because it prompted action. Truly new pathogens have no historical data to learn from. And many published accuracy figures come from retrospective tests, which tend to look better than real-time performance.
Privacy is another concern. Mobility, search and social media data can improve models, but they raise questions about consent and surveillance that public health agencies must handle carefully.
Frequently asked questions
Did AI predict COVID-19 before it happened?
No. Systems such as HealthMap and BlueDot detected reports of an outbreak that had already begun in Wuhan and flagged them on December 30–31, 2019. That is early detection, not prediction of emergence.
Which diseases are easiest to forecast?
Diseases with long, reliable records and clear seasonal patterns, such as influenza and dengue, are the most forecastable. Climate-sensitive diseases benefit from weather data. New or rare pathogens are the hardest.
Why do ensemble models work better?
Combining many independent models averages out the biases and blind spots of each one. In both the US and European COVID-19 forecast hubs, the ensemble was more consistently accurate than most individual models.
What is the most important number in an accuracy claim?
There is no single one. Sensitivity means little without the false-positive rate, and a forecast’s accuracy only makes sense alongside its horizon and how it was validated.
Can I use these tools myself?
Partly. HealthMap offers a public outbreak map, and public health agencies publish forecasts for flu and other respiratory illnesses. EIOS is reserved for national authorities and eligible organizations, and BlueDot is a commercial service. For personal health decisions, follow guidance from your local health authority.
Related reading on Helathora
Official sources
- World Health Organization. Epidemic Intelligence from Open Sources (EIOS).
- World Health Organization. WHO upgrades its public health intelligence system to boost global health security (October 13, 2025).
- Centers for Disease Control and Prevention. HAN Advisory: Outbreak of Pneumonia of Unknown Etiology in Wuhan, China (January 8, 2020).
- Cramer EY et al. Evaluation of individual and ensemble probabilistic forecasts of COVID-19 mortality in the United States. PNAS, 2022.
- Sherratt K et al. Predictive performance of multi-model ensemble forecasts of COVID-19 across European nations. eLife, 2023.
- Lazer D, Kennedy R, King G, Vespignani A. The Parable of Google Flu: Traps in Big Data Analysis. Science, 2014.
- Bogoch II et al. Pneumonia of unknown aetiology in Wuhan, China: potential for international spread via commercial air travel. Journal of Travel Medicine, 2020.
- Hoover Institution. Open-Source Early Warning of the COVID-19 Outbreak (2021).
- Leung XY et al. A systematic review of dengue outbreak prediction models: Current scenario and future directions. PLOS Neglected Tropical Diseases, 2023.
- Sutriyawan A et al. Global Forecasting Models for Dengue Outbreaks in Endemic Regions: A Systematic Review. 2025.
- Bayesian hybrid statistical and machine learning models for dengue forecasting in Bangladesh. medRxiv preprint, 2025.
- Centers for Disease Control and Prevention. Center for Forecasting and Outbreak Analytics.
Medical disclaimer: This article is for general information only and does not constitute medical or public health advice. The outbreak tools and forecasts described here are used by professionals and do not replace official guidance. For advice about a current outbreak or your own health, consult your local health authority or a qualified healthcare professional.