Detecting outbreaks before they spread
An early warning system turns routine health data into timely signals, so that public-health teams can investigate and respond while an outbreak is still small. Here is how they work, and how EMERGENT applies AI to laboratory data.
What is an early warning system?
An epidemic early warning and response system aims to detect outbreaks at their earliest stage, when timely corrective measures can still prevent new cases and interrupt the natural course of the outbreak.
It does so by continuously screening health data, generating signals when observed values depart from what is expected, and routing those signals to people who can verify them and act.
Public-health surveillance usually combines two complementary streams:
Structured, routine data
Regular counts of notified cases, laboratory results or syndromes, analysed against thresholds and historical baselines. EMERGENT belongs to this family.
Unstructured reports
Rumours, media reports, social-media posts and community alerts about unusual health events, screened and verified by analysts.
From manual review to automated, explainable signals
Our systematic scoping review of AI-based early warning systems found that AI can identify patterns in data that signal the onset of epidemics. In practice, AI contributes at several stages:
Data integration
Automated extraction and harmonisation of records from different laboratory and hospital systems, using standards such as HL7 FHIR and LOINC.
Anomaly detection
Statistical and machine-learning models estimate the expected number of cases and flag weeks that are unusually high.
Text & social signals
Natural-language processing can screen news and social media for event-based signals that complement laboratory data.
Forecasting
Time-series models project short-term trends to support preparedness and resource planning.
Antimicrobial resistance
Monitoring resistant organisms in laboratory results helps detect the spread of extensively drug-resistant bacteria.
Decision support
Dashboards and alerts present signals to experts, whose feedback improves the system over time.
Surveillance in Lebanon
Communicable-disease surveillance in Lebanon is led by the Epidemiological Surveillance Program of the Ministry of Public Health (MoPH), established in the mid-1990s. Over the years, the Ministry added an Early Warning, Alert and Response System (EWARS) and web-based reporting, with data increasingly entered at source through the DHIS2 platform.
The system relies primarily on mandatory notification by physicians, hospitals and laboratories. Successive crises — economic collapse, the COVID-19 pandemic, a cholera outbreak, population displacement and conflict — have placed heavy pressure on these reporting channels.
The gap EMERGENT addresses
Notification-based surveillance can have low sensitivity, limited coverage, data-validity issues and reporting delays. Hospital laboratories already hold coded, objective evidence of infection — EMERGENT analyses it automatically, at the source, to complement national reporting rather than replace it.
A hospital-anchored, AI-driven early warning system
Co-designed with clinical and laboratory stakeholders and built as a web platform, EMERGENT combines automated laboratory-data extraction with a hybrid anomaly-detection approach.
Development in three stages
- Evidence review — literature reviews to select target pathogens and effective detection methods.
- Co-design — designing the system with clinical and laboratory stakeholders to fit local workflows.
- Platform build — a web application with automated extraction, detection, dashboard and email alerts.
Why an ensemble?
No single algorithm performs best for every pathogen: some series are seasonal, others are rare and noisy. Combining several detectors makes alerts more robust, and the user-adjustable threshold lets teams balance early detection against false alarms.
Read more in our chapter “Conception and Development of an AI-Driven Hospital-Based Early Warning System” (Springer, 2026).
How an alert threshold works
This interactive example uses simulated data containing hidden outbreaks. Move the sensitivity slider to see the trade-off every early warning system must manage: a lower threshold catches outbreaks sooner but raises more false alarms.
Illustrative only — simulated counts, not real Lebanese surveillance data. The two methods shown are simple textbook detectors, not the full EMERGENT ensemble.
How do we know it works?
| Metric | Question it answers |
|---|---|
| Sensitivity | What share of real outbreaks are detected? |
| Specificity | How often are normal weeks correctly left un-flagged? |
| PPV / NPV | When an alert fires (or doesn't), how likely is it to be right? |
| AUROC / AUC-PR | Overall discrimination across all thresholds, including for rare events |
| Timeliness | How many days earlier than routine reporting is a signal raised? |
See: Practical approach for evaluating ML anomaly detection algorithms for epidemic early warning systems (2025).
Principles we follow
- Human oversight — alerts support, never replace, expert judgement.
- Privacy by design — detection works on aggregated weekly counts rather than individual patient profiles.
- Transparency — detectors and thresholds are explainable and adjustable.
- Context and equity — tools are designed for fragile, resource-constrained health systems, with local stakeholders.
- Alignment with national policy — built to follow MoPH standards and priorities.
Interested in piloting EMERGENT?
We welcome laboratories, hospitals and public-health partners who want to explore AI-driven laboratory surveillance.