How AI Can Improve Event-Based Surveillance for Infectious Diseases

Infectious disease surveillance is a core activity of public health agencies, and AI tools are poised to dramatically improve the speed, accuracy, and completeness of pathogen-based surveillance, i.e., the systematic collection, analysis, and use of reports from laboratories of confirmed infectious disease cases. In this blog, I’ll focus on the ways in which AI is likely to enhance the other major approach to infectious disease surveillance—event-based surveillance—and discuss how AI can address its current weaknesses and improve its utility for public health practice.

What Is Event-Based Surveillance?

The World Health Organization (WHO) broadly divides infectious disease surveillance into two categories. Indicator-based surveillance involves the systematic collection of structured data, such as routine health reports or laboratory results. I covered this in the previous blog post, where I used the term “pathogen-based surveillance” to refer to how public health agencies collect reports from clinical laboratories about bacteria, viruses, fungi, and other pathogens that are legally mandated to be reported. 

The other category of infectious disease surveillance is event-based surveillance (EBS. In contrast to indicator or pathogen-based surveillance, EBS, is the organized and systematic collection, monitoring, assessment, and interpretation of information about health events that could pose risks to public health using unstructured data from a wide range of sources, such as news reports, social media posts, online forums, and community-level observations such as from a medical provider or lay person. The goal of EBS is to augment traditional indicator-based surveillance systems, because EBS is more sensitive to acute events, for events involving new diseases that do not yet have laboratory tests, and for events in which specimens have not yet been collected or tested.

Public health agencies continuously seek new ways to improve their ability to monitor infectious diseases. Starting in the 1990s, as concerns about bioterrorism rose, public health experts began looking at methods that are potentially more sensitive than reporting by physicians and labs, such as relying on reports of clinical syndromes from emergency response services and emergency departments. 

From the 1990s to 2010s, the emergence of influenza H5N1, SARS, MERS, and influenza H1N1 prompted many middle and high income countries to develop surveillance systems to monitor influenza-like illness and severe acute respiratory infection. It was not until the 2014-16 Ebola outbreak in West Africa, however, that WHO and other public health agencies (like Africa CDC) began developing formal global guidance and protocols for what was eventually given the name “event-based surveillance.”

During Ebola, it became clear that public health agencies could find, isolate, and trace Ebola cases faster if they received reports from minimally trained lay persons in the community about people who were severely ill or dead, rather than waiting until ill persons presented to hospitals or lab tests returned from tests on deceased people.

Sources of Event-Based Surveillance Data

Multiple systems and data sources exist under the umbrella of EBS. Data sources include:

Community Reports: Community leaders, and citizens often serve as the first line of detection, reporting unusual clusters of illness to local authorities by calling specialized hotlines or public health agencies.

Non-traditional Health Entities: Pharmacies (e.g., sales of cough and cold medications), hotlines (911), schools (e.g., attendance and health records) and emergency medical services can provide useful information about infectious and non-infectious health conditions that may be increasing in a community. 

Traditional Media: Newspapers, radio broadcasts, and television news reports are often the first to highlight disease outbreaks or unusual health events in remote or underserved areas.

Digital Media: Social media platforms (e.g., X, Facebook), blogs, and online forums often include individuals reporting illness about themselves or in their community, as well as healthcare providers discussing emerging events in their community. During COVID-19, Twitter (now X) proved invaluable for identifying places with outbreaks.

Official Sources: Government or non-governmental organizations may release press statements or reports that indicate a health threat. For example, a school may tell parents about an outbreak among its students without having notified a public health agency.

Emerging Technologies: Increasingly, non-traditional sources such as satellite imagery, video feeds, and audio recordings are being explored for their potential to provide actionable insights.  Implementing each of these systems may require anything from minimal investment—e.g., identifying lay personnel in communities who are willing and interested in reported—to extensive protocol development, training, ongoing monitoring and evaluation, and electronic data collection systems, such as for pharmacy and digital media monitoring. 

In the US, many of these systems have long existed, but were called “syndromic surveillance” (e.g., school absentee reports, emergency medical reports) or were simply never given a formal “event-based surveillance system” moniker. For example, the “astute clinician”—a physician who suspects a patient has a disease that is of public health significance—is often the first person to notify a public health agency of an emerging threat, as during the first cases of inhalation anthrax from bioterrorism in 2001. Now such reporting would fall under the term EBS. 

Risk Assessment for Events

Because EBS identifies many more potential events than indicator-based systems, risk assessment is an essential part of the workflow for epidemiologists running these systems. Most health agencies used a systematic approach similar to that outlined in WHO’s Strategic Toolkit for Assessing Risks, which involves.

Event Detection and Characterization: Identifying the event and gathering initial data from multiple sources, including indicator-based surveillance systems, on its nature, geographic location, affected populations, and transmission dynamics.

Hazard Assessment: Understanding the biological agent (e.g., virus, bacterium) and its properties, including transmissibility, severity, and potential for spread. This often involves contacting subject matter experts to help assess how unique this event is compared to what is already known about the pathogen.

Exposure and Vulnerability Analysis: Evaluating the population’s exposure to the hazard and identifying vulnerabilities, such as immunological status, healthcare capacity, and socio-economic factors.

Impact Analysis: Estimating the potential health, social, and economic consequences of the event.

Risk Characterization: Combining data from the above steps to prioritize the event and determine response actions.

If the event is potentially of large significance–such as an outbreak involving a special population like infants or a potential bioterrorism agent—these assessments could involve large numbers of humans working together to synthesize, debate, and reach consensus on a final recommendation.

Challenges and Limitations in Event-Based Surveillance

Some of the major challenges and limitations with EBS include: 

Data Volume: Public health agencies are inundated with vast quantities of information from disparate sources, much of it unstructured and noisy. It can require extensive time and effort to manually clean and process this into usable data.

Verification: Determining the validity and relevance of signals is resource-intensive, often requiring manual intervention. In simple terms, event-based product surveillance produces a lot of smoke, but very little of it is fire; pathogen-based surveillance only tells you about fires and, even then, only the fires that you know for sure are occurring.

Language and Accessibility: Media and community reports may be in local dialects or inaccessible to agencies lacking linguistic expertise.

Coverage: Many EBS systems are limited to digital or print media, leaving out other critical sources like radio broadcasts or video feeds.

In December 2024, major legacy media outlets began sounding the alarm about an outbreak of “Disease X” in the Democratic Republic of Congo. Large numbers of children were reported to be severely ill and dying from what was feared to be a new infectious disease. After news outlets began running stories, discussion boards lit up with people postulating the emergence of a new hemorrhagic fever or the evolution of mpox into something more virulent. 

EBS systems detected this outbreak, but lab testing eventually showed that there was no new disease. Children were not dying from “disease X”; they were dying from malaria, measles, and other routine, endemic infectious diseases, overlayed on top of severe malnutrition. 

The upside to EBS is that it raised attention to a community that desperately needed resources for nutrition, vaccination, antimicrobial treatment for malaria and bacterial pneumonia, and supportive care for measles and other severe illnesses. The downside is that it likely diverted attention from public health officials and donors to the ongoing, devastating outbreak of mpox already occurring in the country.

Another example of the strengths and weaknesses of event-based surveillance was the large amount of publicity that Google received in the early 2010s when it released “Google flu trends.” There was tremendous optimism that monitoring internet searches for flu-related terminology would provide an early warning of influenza in a community; ongoing use, however, demonstrated that while it did provide an earlier indication of influenza trends, it was also highly susceptible to people searching for other infectious conditions or to non-respiratory conditions and provided data that was not particularly actionable for hospitals, providers, or public health agencies. 

Do We Need AI to Enhance Event-Based Surveillance?

In my public health career, I have designed, developed, and used EBS systems in China (for severe acute respiratory infections and influenza like illness), in New York City (using EMS, emergency department, school absenteeism, and pharmacy records), and in Africa (developing the first EBS program for Africa CDC). My experience is that they are extremely useful as adjuncts to traditional indicator-based surveillance systems, helping, in particular, to understand the degree of under-diagnosis and under-reporting that may be occurring as well as the potential impact of a disease in causing outbreaks or social disruption. 

But they also require substantial time to build, maintain, and operate. The lack of specificity–lots of smoke, not so much fire; too much noise, not enough signal–results in large amounts of human cognitive labor. I believe that AI could have a transformative impact on EBS, because it can automate many of these tasks performed by humans right now. In fact, this is probably the one area in which machine learning and other AI algorithms have already shown the biggest impact in helping monitor epidemics as seen by the publicly-available HealthMap and the WHO-operated Epidemic Intelligence from Open Sources (EIOS).

AI Will Improve Data Ingestion and Integration

AI tools can automate the collection of data from diverse sources, including radio, video, and social media platforms. For example:

Natural Language Processing (NLP): NLP algorithms can analyze radio broadcasts, print articles, and social media posts in multiple languages, identifying relevant keywords and patterns indicative of health events.

Speech-to-Text Technology: AI-driven transcription tools can convert audio content from radio or video reports into text for further analysis.

Video Analysis: Machine learning models can process video footage, such as televised news or surveillance camera feeds, to identify scenes suggestive of health crises (e.g., overcrowded parking lots of hospitals or clinics, unusual animal die-offs).

AI Will Improve Signal Detection and Prioritization

AI excels at identifying patterns in large datasets, making it ideal for detecting health signals amidst noise.

Anomaly Detection: Machine learning algorithms can identify unusual spikes in online discussions, news reports, or community complaints, flagging them for further investigation.

Sentiment Analysis: NLP tools can assess the tone and urgency of social media posts or news articles, helping prioritize potential threats.

AI Will Improve Verification and Contextualization

One of the most resource-intensive aspects of EBS is verifying whether a reported event is legitimate. Currently, when a new event occurs, humans need to review the text from the electronic report then often make phone calls or send emails to health officials or other trusted persons in a community to understand if the report is authentic and accurate. 

AI can streamline this process:

Cross-Referencing: AI systems can compare signals from multiple sources (e.g., matching a social media post about an outbreak with a corresponding news report).

Automated verification calls or messages: AI can auto-generate emails, text messages, or even voice calls to health officials in a community to do the work currently done by human epidemiologists.

Geospatial Analysis: Geographic information systems (GIS) powered by AI can contextualize events, identifying proximity to known outbreak hotspots or vulnerable populations.

All of that said, I also think this represents one of the biggest risks of current AI systems. LLMs essentially scrape the internet for unverified data, which is why they “hallucinate” when asked to provide detail for queries that they may not have sufficient data to support – of course, just like some humans do. Careful human review–described in tech sectors as “human in the loop”–will be particularly critical for verification and contextualization of AI-assisted work in this area. 

AI Will Improve Translation and Accessibility

AI tools can overcome language barriers, enabling public health agencies to monitor data in local dialects or languages they might otherwise overlook.

Real-Time Translation: AI-powered platforms can translate audio, text, or video content into multiple languages, ensuring global reach and inclusivity.

Linguistic Models: AI can train on local slang or idiomatic expressions to better understand culturally specific signals.

AI Will Improve Risk Assessment

As discussed above, risk assessment is one of the most important and most labor-intensive parts of the event-based surveillance workflow. I think that AI will play a particularly important role in accelerating this process by ensuring that analyses are done using a systematic approach, consider the full range of scientific and public health data, and integrate data from multiple population-level sources, including softer criteria, such as political impact.

Enhanced Hazard Assessment: AI models can analyze genomic data, epidemiological studies, and historical outbreak patterns to provide a detailed understanding of the infectious agent.

• Genomic Sequencing: Machine learning algorithms can analyze pathogen genomes to identify mutations, predict transmissibility, and assess potential resistance to treatments.
• Historical Comparisons: AI systems can compare the new event to similar past outbreaks, offering insights into probable outcomes and control measures.

Population Vulnerability Analysis: AI can evaluate vulnerabilities within affected populations by integrating demographic, healthcare, and socio-economic data.

• Predictive Models: AI can forecast healthcare system strain, identifying regions where hospital capacity, medical supplies, or workforce shortages may impede the response.
• Geospatial Analysis: AI analysis of geographic information systems data can assess population density, mobility patterns, and access to care, highlighting areas of greatest risk.

Impact Forecasting: AI models can simulate various outbreak scenarios to estimate the potential impact on health systems, economies, and social structures.

• Scenario Planning: Machine learning algorithms can generate projections under different response strategies, helping policymakers evaluate trade-offs and plan resource allocation.
• Economic Impact Analysis: AI tools can predict economic disruptions, such as workforce absenteeism, supply chain interruptions, and healthcare costs.

Risk Prioritization and Communication: AI can assist in synthesizing data from multiple sources to provide a clear, actionable risk characterization.

• Dashboard integration: AI-powered dashboards can present risk assessments visually, allowing decision-makers to understand complex data at a glance.
• Automated reporting: Generative AI tools can create concise, real-time summaries for public health officials and stakeholders.

The Future of AI and Event-Based Surveillance

EBS is a critical tool for early outbreak detection, offering a complementary approach to pathogen-based systems. Public health agencies need AI developers and companies to help them make EBS systems less dependent on human epidemiologists scanning, aggregating, verifying, analyzing, and writing.

Assisted by AI, epidemiologists can make these systems more timely, accurate, and useful for public health practice. In the future, AI developers and companies will need to collaborate with public health agencies on the integration of even more diverse data sources, including wearable health devices and drone surveillance in remote areas.