Canada's Global Public Health Intelligence Network (GPHIN), run by the Public Health Agency of Canada, is a working example of AI in public health: software collects about 7,000 news items a day in nine languages, discards roughly half, and passes the rest to human analysts who decide what counts as a threat. An independent review panel reported in 2021 that the system flagged the first signal of the COVID-19 outbreak on December 30, 2019, and that its analysts, not its algorithms, were the core of the system.
Key takeaways
- GPHIN is an event-based surveillance system. It reads news and similar open sources, not patient records, to spot possible outbreaks early.
- Machine learning and natural language processing handle de-duplication, translation, tagging, and relevancy scoring. Analysts review the middle band of scores.
- The 2021 independent review found no evidence that GPHIN could have detected the COVID-19 outbreak earlier than it did.
- The panel flagged heavy dependence on one news aggregator, bugs, and the cost of labeling training data as weaknesses.
- The World Health Organization runs a comparable tool, Epidemic Intelligence from Open Sources, used by 120 countries by the end of 2025.
What event-based surveillance means
Traditional disease surveillance counts cases reported by clinics and laboratories. That data is accurate but slow. Event-based surveillance works from the other direction: it looks for unusual reports, such as a cluster of pneumonia cases in a local newspaper, and asks analysts to check them. GPHIN, created about 25 years before the 2021 review, monitors global news sources for chemical, biological, radiological, and nuclear threats as well as infectious disease. The review panel described it as a partially moderated system, meaning automation and human judgment share the work. It monitors Arabic, Chinese (traditional and simplified), English, French, Russian, Spanish, Portuguese, and Farsi.
Where the machine learning sits
According to the panel's final report, the platform has used machine learning and natural language processing since its 2004 launch. The automated part has five phases: removing duplicates, adding metadata, categorizing articles, translating them, and scoring relevance. About 7,000 items arrive each day and about half are filtered out. High-scoring items are published automatically, low-scoring ones are discarded, and the middle band goes to an analyst. Between 1,000 and 1,500 records are published to the database daily, and 2,500 to 3,000 during the height of COVID-19.
The algorithms learn from labeled examples, so their accuracy depends on how well past articles were tagged. The panel noted that labeling is costly and that analysts were not given enough protected time for it. It recommended a standardized labeling process.
The December 30, 2019 signal
The report records that GPHIN first picked up a signal on December 30, 2019, at 10:30 p.m. Eastern, when an analyst on duty read a South China Morning Post article about a pneumonia outbreak. The item ran in the December 31 daily report, and a special report reached senior leadership on January 1. The panel found no evidence that earlier detection by GPHIN was possible, and it noted that BlueDot and ProMED identified the outbreak on the same day. A detection on the same day as other systems is the honest summary: the tool worked as designed, and it was neither first nor late.
What the panel criticized
| Issue found in 2021 | Detail from the panel's report |
|---|---|
| Single-source dependence | One news aggregator supplied 84 percent of inputs; RSS feeds supplied 11 percent and Twitter 2 percent |
| System reliability | Analysts reported slowness, crashes, an unreliable search function, and incomplete de-duplication |
| Missing support role | The technical advisor position was phased out in 2017 and not replaced |
| Isolation from risk assessment | GPHIN products were not well connected to the process that turns a signal into a decision |
| Limited social media | Verification, signal-to-noise, and deepfake concerns kept social media monitoring modest |
The panel members were Margaret Bloodworth, Mylaine Breton, and Dr. Paul Gully. The panel dated its letter to the Minister of Health May 28, 2021, and the report was published on July 12, 2021. Its recommendations included keeping GPHIN's surveillance capacity within the agency, restoring a dedicated technical advisor, and building the next upgrade as a modular system. The status of those recommendations in 2026 was not confirmed for this article.
A comparable global tool
The World Health Organization hosts Epidemic Intelligence from Open Sources (EIOS), which it describes as an initiative that analyzes open-source data to detect public health threats in near real time. WHO reports that by the end of 2025, 120 countries and 30 organizations and networks were using it. Its public page does not describe AI methods, so the comparison between the two systems is one of purpose, not of algorithms.
Questions to ask about AI claims in public health
- Does the system detect events, forecast cases, or support decisions? These are different tasks with different evidence.
- What are the data sources, and how many does a single supplier control?
- Who reviews the output, and who can overrule it?
- How was it evaluated against a known event, and was the review independent?
- How are false alarms and missed signals tracked over time?
This article gives general information about surveillance technology. It is not medical or public health advice, and it makes no statement about any individual's health risk. Questions about outbreaks in a specific area belong with the local public health authority.
FAQ
Does GPHIN use personal health data?
The panel describes it as monitoring global news and other open sources, so it works from published reports and not from patient records.
Did AI detect COVID-19 first?
No single system was first on the evidence in the report. GPHIN's signal came on December 30, 2019, and BlueDot and ProMED flagged the outbreak the same day.
Can the algorithm replace public health analysts?
The panel called analysts the heart of the system and said their language skills and judgment cannot be automated.
Is GPHIN still operating?
The 2021 report describes it as an operating system. Its current configuration and any changes since the review were not verified for this article.
Related reading
- What health education and promotion mean
- AI and cancer care
- Documented risks of artificial intelligence
- Where human judgment still leads
More in AI in health.
Last verified: October 2026
Sources
- Global Public Health Intelligence Network (GPHIN) Independent Review Panel Final Report, Public Health Agency of Canada, 2021
- Epidemic Intelligence from Open Sources (EIOS), World Health Organization, 2025