What problem does it solve?
Formal surveillance depends on laboratories, hospitals and official reports, which reach a national agency once cases are confirmed and reported. News, social media and other public sources can carry the first sign of an outbreak, in the local language, before a report arrives. Canada's GPHIN review panel cites a case in which smartphone and Twitter data helped identify an Ebola outbreak in West Africa three days before an official WHO announcement. A small team cannot read that flow by hand: the panel reports that GPHIN's queries pulled over 7,000 articles a day into its platform and that about half were filtered out before analysts manually reviewed the rest.
Open source signals are noisy, and the panel notes that false positives are an accepted outcome of event based surveillance, because missing a real event costs more than checking one that turns out to be nothing. The system has to raise the likely events quickly, work across languages and leave verification and escalation to people.
- Canada's GPHIN review panel reports that smartphones and Twitter data helped public health professionals identify an Ebola outbreak in West Africa three days before an official WHO announcement.Global Public Health Intelligence Network (GPHIN) Independent Review Panel Final Report (2021)
How does it work?
- Define sources and queries. The service licenses a news feed and adds RSS feeds from health bodies and media, and selected social media accounts. Analysts write and refine the queries. GPHIN drew 84 percent of its input from a single licensed news service.
- Process automatically. A pipeline flags near duplicate articles, adds metadata, classifies each article against a hazard taxonomy, machine translates foreign language text and scores relevance. GPHIN's automated stage has these five phases. WHO's EIOS 2.0 adds AI tools for automated analysis and signal detection and processes radio broadcasts with automatic transcription and translation.
- Analysts triage. Analysts read the filtered articles, check the source and decide whether an item is a signal under the International Health Regulations criteria. They send alerts and a daily report to subscribers.
- Verify and assess. Public health authorities verify the signal through official channels and assess its risk. A signal can reach senior decision makers within two days: the pneumonia signal GPHIN detected on 30 December 2019 was raised to senior tables across government by 1 January.
- Maintain. Analysts add new vocabulary to the taxonomy, for example the word "COVID" after it appeared, and review false positives and missed events.
- Audience
- Employee facing
- Autonomy
- Assist
- Adoption
- Early adopters
- Channels
- Internal tools, Email
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Interactions handled | Not pooled | at least 7000 | 1 | 1 independent |
Value drivers: Risk and loss reduction, Speed and cycle time, Employee productivity.
Indicative value
A national public health agency whose analysts review open source articles every day
EUR 109,500 to EUR 1.5 million
Analyst reading time avoided by filtering per year
How this is calculated
Formula: articlesPerDay * filteredShare * minutesPerArticle / 60 * 365 * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Articles collected per day articlesPerDay, articles per day | 3,000 | 7,000 | The review panel reports that GPHIN pulled over 7,000 articles a day into its platform. The low end is an editorial assumption for a smaller service. Source |
| Share of articles filtered out before analysts read the rest filteredShare, fraction of articles | 0.3 | 0.5 | The panel reports that about half of the articles were filtered out before analysts manually reviewed the rest, without saying how much of that is automatic. Counting the filter as work the software does is an editorial assumption, as is the low end. Source |
| Analyst minutes to skim one article minutesPerArticle, minutes per article | 0.5 | 1 | Editorial assumption. Replace with a time study of your own analysts. |
| Fully loaded cost of one analyst hour hourlyCost, EUR per hour | 40 | 70 | Editorial assumption. Replace with your own cost. |
What it leaves out: This values only the reading time that filtering saves. It does not value the earlier warning, which is the reason to run the service, and it leaves out the cost of the news licenses (GPHIN's main feed cost CAD 1.8 million every two years according to the review panel), the platform and the analysts' remaining work, which together can exceed the reading time saved.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Public Health Agency of Canada (PHAC)
Canada · Government and public sector · 2021
The Global Public Health Intelligence Network (GPHIN) is a Canadian early warning system that scans global news in nine languages for chemical, biological, radiological and nuclear and other public health threats. An independent review panel described it as partially moderated: an automated stage flags duplicates, adds metadata, categorizes and translates articles and scores their relevance, using machine learning and algorithmic filtering, and analysts review the rest and write alerts and daily reports.
GPHIN's analyst on duty detected a signal about a pneumonia outbreak on 30 December 2019, which the panel found came at about the same time as other surveillance systems; BlueDot and ProMED flagged the outbreak the same day. The panel also found the platform unstable, in use in its 2018 version with later fixes not migrated, and recommended changes to mandate, governance, technology and the role of analysts.
The panel's report is dated May 2021 and was published in July 2021.
- Interactions handled: at least 7000, articles pulled into the system per day
"Analysts have created (and continue to refine) sets of complex logical queries that pull over 7,000 articles a day into the system."
Claimed by: independent
Pan American Health Organization (PAHO/WHO)
Global · Government and public sector · 2025
PAHO has trained countries in the Americas on the WHO Epidemic Intelligence from Open Sources (EIOS) system since 2019. A published analysis covering 17 countries trained between August 2019 and October 2023, after 105 PAHO training sessions, found that most of them put EIOS into routine event based surveillance within a year of training, at different intensity: 24 percent became frequent users, 18 percent moderate, 18 percent minimal and 29 percent ad hoc users.
PAHO reports that in countries using EIOS the share of health events first detected and reported by national authorities rose from 68 percent to 85 percent in the first year after training, while it fell in countries with minimal or no use. PAHO reads this as evidence that EIOS strengthens early warning; the release is PAHO's own summary of the analysis.
No outcome disclosed.
World Health Organization (WHO Hub for Pandemic and Epidemic Intelligence)
Global · Government and public sector · 2025
The Epidemic Intelligence from Open Sources (EIOS) system, hosted at the WHO Hub in Berlin, lets ministries of health and public health agencies find relevant content in websites, social media and other public sources to spot health events early, which authorities then verify and assess. WHO offers it free of charge to its Member States and eligible organizations, with training and communities of practice, to complement formal channels such as laboratories and hospitals.
On 13 October 2025 WHO launched version 2.0, which adds new data sources, AI powered tools for automated analysis and signal detection and processing of radio broadcasts with automatic transcription and translation. WHO reports that more than 110 Member States and around 30 organizations and networks use it.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Licensed news feed and RSS sources with rights to process and share article links
- A hazard taxonomy and multilingual query sets maintained by analysts
- A history of verified signals and false alarms to measure the system against
- Rules on which social media sources may be processed and how personal data is handled
Systems to integrate
- News and media monitoring feeds
- Signal and event management system for analysts
- Email or portal distribution of alerts and daily reports to national and international partners
Complexity: Medium
A working service needs licensed news and monitoring feeds, a multilingual hazard taxonomy and query sets, a processing pipeline with deduplication, classification and machine translation, analysts who read the languages of the sources, and a channel to subscribers and to national and international health authorities. For GPHIN the main licensed news service accounted for 84 percent of input at a cost of CAD 1.8 million every two years, according to the review panel.
- 1
Name the decision the service supports
Write down which events must reach which decision makers in what time. WHO presents EIOS as a complement to formal channels such as laboratories and hospitals, with authorities verifying and assessing what the system surfaces.
- 2
Secure sources before models
List the feeds, the licenses and the languages first. GPHIN's panel recommended that the team continue to explore incorporating further social media, and was told that the platform lacked the technical capacity to monitor open source social media in full.
- 3
Keep analysts in the loop and invest in language coverage
GPHIN is a partially moderated system: machine learning and algorithmic filtering do the first pass and analysts working in nine languages decide what is a signal. The panel judged that language ability is one reason GPHIN's algorithms work.
- 4
Treat the platform as a product
Those closest to GPHIN's system told the panel that it was unstable, often slow and prone to crashes, and the panel notes that fixes made by the National Research Council have not moved to production since fall 2018. The panel heard differences of opinion about the state of the system. Assign an owner, a release process and a budget.
- 5
Evaluate against outcomes and peers
The panel found that earlier attempts to evaluate the performance of PHAC's event based and indicator based surveillance systems appear to have relied primarily on self assessments. Compare detection times with other systems and record false positives and misses.
Guardrails
- A signal is a lead, not a confirmed event; nothing is published as fact before verification
- Analysts check the source and apply a written criteria set before an alert goes out
- Personal data from social media is minimized and processing is limited to what the service needs
- Machine translations are checked by an analyst before they inform an alert
- A named owner approves changes to queries, taxonomy and models
KPIs to instrument
- Articles collected, filtered out and reviewed per day
- Share of alerts confirmed as events and share judged false positives
- Time from first public report to alert, against peer systems
- Share of health events first detected through the tool
- Active users and share of countries or teams using it in routine work
Human in the loop
Analysts read the filtered articles, decide what counts as a signal and write the alerts. Epidemiologists verify and assess a signal with national and international partners before any response. People decide whether an event exists; the software raises leads.
Common failure modes
- Platform decay
- GPHIN's review panel described a system analysts called unstable, with fixes not migrated to production since 2018. A tool that is slow or crashes loses analysts' trust and time.
- Source concentration
- When one licensed service supplies 84 percent of input, a price rise or a license change changes what the system sees. Keep a plan for the second source.
- Alert fatigue
- Accepting some false positives is intended in this field, but a daily flood of weak alerts lowers the value of each one. Tune relevance scores against analysts' decisions.
- Unclear mandate and governance
- The panel found GPHIN's mandate, governance and information flows needed work, even though GPHIN's analyst on duty had detected the pneumonia signal. A technical upgrade does not fix who decides, who is told and who pays.
What are the risks and rules?
EU AI Act
Minimal risk
Processing public text and broadcasts to find health events is not a use listed in Annex III and makes no decision about a person. GDPR still applies where social media posts or other personal data are processed, and the assessment would change if the system were used to profile individuals.
Rules that apply
Guidance
- Global Public Health Intelligence Network (GPHIN) Independent Review Panel Final Report (Public Health Agency of Canada, external review panel, North America). Independent review of a long running system, with findings on mandate, governance, technology and analysts, and recommendations.
- WHO upgrades its public health intelligence system to boost global health security (World Health Organization, Global). Describes EIOS 2.0, its AI tools, who uses it and how authorities verify what it surfaces.
Controls to put in place
- Written criteria for what counts as a signal, applied by analysts
- Verification step with named epidemiologists before any response
- Source register with licenses and a second source plan
- Review of false positives and missed events each quarter
- Data protection rules for social media sources
Frequently asked questions
- What is event based surveillance?
It is the monitoring of unstructured sources such as news, social media and reports for signs of health events, as a complement to indicator based surveillance from laboratories and hospitals. GPHIN's review panel describes GPHIN as a public health event based system, and WHO describes EIOS as complementing formal channels.
- Does AI replace the analysts?
No deployment on this page does. GPHIN is partially moderated: machine learning and algorithmic filtering do the first pass and analysts work in nine languages. WHO says authorities verify and assess what EIOS identifies.
- Who uses it and how widely?
WHO reports that more than 110 Member States and around 30 organizations and networks use EIOS. PAHO reports that 88 percent of 17 trained countries in the Americas used it in routine surveillance within a year of training, though 29 percent only as ad hoc users.
- Did it detect COVID early?
GPHIN's analyst detected a signal about a pneumonia outbreak on 30 December 2019, and the review panel found the timing in line with similar systems, with BlueDot and ProMED identifying the outbreak on the same day. The panel saw no evidence that earlier detection would have been possible.
How to cite this page
Blits.ai AI Use Case Library, "AI epidemic intelligence for detecting public health threat signals in open source news and social media", last verified 14 October 2026, https://www.blits.ai/ai-use-cases/epidemic-intelligence-signal-detection. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 14 October 2026: First published