Epidemic intelligence & infectious-disease outbreak forecasting (incl. wastewater surveillance)

Health agencies must spot an outbreak before official case counts arrive and project how it will evolve. AI scans multilingual news and reports for early signals (event-based surveillance), and models wastewater, syndromic and laboratory data to produce short-term forecasts with uncertainty. Analysts stay in charge of verification and risk assessment; the benefit is days of lead time and a consistent, evaluated forecast instead of ad-hoc spreadsheets.

Problem & value

• Today: epidemic intelligence officers read hundreds of articles and bulletins daily, in many languages, and judge which are relevant. Indicator-based surveillance (notified cases, sentinel networks) is reliable but lags by days to weeks. Forecasts are often built per institute with different methods and no shared evaluation.
• AI value: (1) automated filtering, translation, classification, deduplication and geo-tagging of open-source texts to surface weak signals; (2) wastewater signals that lead clinical data and do not depend on testing behaviour; (3) ensemble forecasts that are scored against observed data.
• Evidence: WHO's EIOS (Epidemic Intelligence from Open Sources) platform is used by ECDC and other agencies for event-based surveillance [1]. ECDC's RespiCast hub combines weekly submissions from external modelling teams into ensemble forecasts for ILI, ARI and COVID-19 indicators; a 2025 evaluation found ensembles generally outperformed a persistence baseline and most single models [2]. Wastewater monitoring is being scaled up in the EU following Commission Recommendation (EU) 2021/472 [3].
• Caveat: published forecast skill drops sharply at turning points and for novel pathogens; do not promise fixed lead times without local back-testing.

Variants

• Event-based surveillance (EBS): LLM/NLP triage of news and reports; EIOS-style workflows.
• Wastewater-based epidemiology (WBE): SARS-CoV-2, influenza, RSV, polio, antimicrobial-resistance genes; per treatment plant, with normalisation and trend detection.
• Syndromic and short-term forecasting: ILI/ARI, COVID-19 hospitalisations, regional hubs.
• Genomic surveillance: variant detection and lineage growth estimation from sequencing data.
• Local health office (Gesundheitsamt) / hospital infection control: ward-level cluster detection (touches HLTH-012, HLTH-020).
• Animal and One Health: avian influenza, vector-borne disease with climate covariates (PRED-017).

How it works

1. open sources (news, reports, social, ProMED-type feeds)
2. crawl/ingest
3. language ID + translation
4. relevance classifier + event extraction (disease, place, date, count)
5. dedup/clustering
6. analyst triage queue wastewater lab data + clinical/syndromic + sequencing
7. harmonised time-series store
8. normalisation, outlier handling, trend and anomaly detection
9. forecasting models (mechanistic + statistical + ML)
10. ensemble with calibrated intervals
11. dashboards, alerts, situation-report drafts
12. epidemiologist review
13. risk assessment & response

• Signal detection: multilingual encoders and LLMs classify relevance and extract structured events; embeddings cluster duplicates. LLMs also draft summaries with citations to the source articles; every alert links back to the original text.
• Forecasting: renewal-equation and compartmental models (e.g. EpiEstim-style reproduction-number estimation, SEIR), statistical time-series models, gradient boosting and neural networks; individual models are combined into an ensemble, scored with proper scoring rules (weighted interval score).
• Wastewater: viral concentration normalised by flow or faecal markers (e.g. PMMoV); trends compared with clinical data to estimate incidence.
• Human-in-the-loop: analysts verify signals, decide on escalation to EU Early Warning and Response System (EWRS) and own the risk assessment; forecasts are decision support, never an automatic trigger for measures.

Data preparation

• Data and location: wastewater lab results (lab systems, spreadsheets, CSV), sentinel and syndromic reports (PDF bulletins, portals), notifiable-disease case counts (national reporting systems), sequencing data, and open-source news and feeds.
• Typical gaps: spreadsheets with changing layouts, missing normalisation values (flow, PMMoV), reporting delays and retroactive revisions, differing catchment and region codes, changing testing policies, no history of what was known on which date, and scattered sharing agreements.
• Preparation steps: inventory sources, owners and sharing rules (PREP-001, PREP-004); convert spreadsheets and PDF bulletins into tidy, machine-readable tables (PREP-011, DATA-003); harmonise region and pathogen codes in a shared glossary (PREP-012); store data as vintages so back-tests use only what was known at the time; build analyst-labelled relevance sets for news triage (PREP-018, PREP-020); aggregate or pseudonymise line lists and apply small-number thresholds (PREP-022).
• Tools: Airflow or Dagster pipelines, a time-series store, Hubverse data formats.
• Effort share: around 60 percent; harmonised, versioned time series are the main asset. See GUIDE-003 for a readiness check.

Implementation paths

Which path for whom

• Public: Path C. Health agencies need reproducible, auditable methods and data sovereignty, and HW1-HW2 plus the R/Python epi ecosystem is enough; use free EIOS (Path A) as an additional signal source.
• Enterprise: Path B. Hospital groups and commercial analysts need signals and forecasts quickly on EU-region models without building a research platform.
• Developer: Path C. Open epi tooling (EpiNow2, Hubverse) and open multilingual models are the natural base for building and back-testing.

Path A — Buy (off-the-shelf)

• Event-based surveillance: WHO EIOS 2.0 (free for public-health institutions, via WHO Hub for Pandemic and Epidemic Intelligence, Berlin) [1]; commercial intelligence services such as BlueDot (Canada) and Metabiota/Ginkgo Biosecurity (US) for organisations without an EBS team.
• Wastewater: laboratory service providers plus data platforms; national/EU programmes (e.g. EU Wastewater Observatory, Germany's AMELAG) supply data rather than software [3][4].
• Choose A when: a small health office or hospital group needs signals, not a research platform.

Path B — Build on cloud APIs

• LLM layer: EU-region frontier models (Mistral, Azure OpenAI EU Data Zone, Claude/Gemini via EU regions) with structured output for event extraction and multilingual summarisation; machine translation via DeepL (EU) or similar.
• Forecasting: managed notebooks/compute on EU cloud; time-series foundation models such as Chronos / TimesFM can serve as additional ensemble members next to epidemiological models.
• Data: public sources (ECDC Surveillance Atlas, ERVISS, Eurostat), national feeds; orchestration with Airflow/Dagster; dashboards with Grafana or Metabase.
• Note: surveillance data is mostly aggregated, but line lists and sequences can be personal data.

Path C — Self-host (open source)

• Models: open-weight multilingual LLMs (Qwen, Llama, Mistral families) for classification and extraction; XLM-R/multilingual-E5 encoders; NLLB / Opus-MT for translation.
• Epi tooling: R/Python ecosystem — EpiNow2, EpiEstim, epidemics (Epiverse-TRACE), Nextstrain (genomic), Hubverse tools for forecast submission and scoring [2][5].
• Hardware: HW1–HW2 is sufficient (small LLMs, CPU time-series fitting); a single 24–32 GB GPU covers multilingual triage of tens of thousands of articles per day.
• Choose C when: data sovereignty rules prevent cloud use or you need reproducible, auditable methods.

Path D — Custom training

Worthwhile for (a) a relevance/event classifier trained on the agency's own labelled analyst decisions (a few thousand examples), and (b) site-specific forecasting models trained on years of local wastewater and syndromic history. Back-test on past seasons; retrain each season because pathogens, testing and behaviour change.

State of the art (2026)

• Forecast hubs (Hubverse, US CDC FluSight/COVID hubs, European RespiCast) have made evaluation standard: simple ensembles of diverse models are consistently hard to beat [2][5].
• Wastewater is now routine for SARS-CoV-2, influenza and RSV in many EU states and expanding to pathogens such as polio and AMR; quantitative links to hospitalisation are catchment-specific.
• LLMs improve multilingual triage and summarisation of open-source intelligence; the open problems are false positives from rumours, regional bias in media coverage and hallucinated details.
• Time-series foundation models are competitive for short horizons but not clearly better than tuned epidemiological ensembles at turning points.
• Limits: reporting delays and revisions, changing testing policies, sparse data for novel pathogens, and non-stationarity. Always show prediction intervals and track calibration.

Shared resources & combinations

• Same stack: LANG-018 (media monitoring: crawler, translation, classification), PRED-011 (volume forecasting, time-series stack), PRED-017 (climate covariates), PRED-040 and PRED-027 (anticipatory-action forecasting practices).
• Downstream: HLTH-010 (pharmacovigilance signals), HLTH-020 (population risk stratification), HLTH-007/HLTH-012 (individual-level monitoring, kept separate from population surveillance).
• Shared infrastructure: time-series store, workflow orchestrator, EU-hosted multilingual LLM server.

Effort, cost & KPIs

• Path A: weeks; EIOS is free for institutions, commercial services priced per subscription.
• Path B: 3–6 months for ingestion, triage UI and first forecasting pipeline.
• Path C/D: 6–12 months with epidemiologists and data engineers; ongoing season-by-season recalibration.
• KPIs: signal precision and analyst time per relevant signal, detection lead time versus official notification, forecast skill (weighted interval score relative to baseline), interval coverage, wastewater-to-clinical lag, share of alerts leading to action.

People & change

• Involve early: agency or institute leadership, the epidemiology lead as process owner, epidemiologists and analysts as daily users, IT, data owners of laboratories and sentinel networks, partner agencies that supply data, and the DPO.
• Works council: depends. Where an elected staff council (Personalrat under BPersVG or state law) or works council exists, inform it when work processes are planned with AI (§90(1) no. 3 BetrVG, analogous in public law); co-determination under §87(1) no. 6 is triggered mainly if usage logs of analysts are kept, so keep them anonymous (GUIDE-004).
• Data protection: aggregated counts are mostly non-personal, but line lists, small-area counts and social-media content can be health data; involve the DPO, use the Art. 9(2)(i) basis, minimise and run a DPIA where needed (GUIDE-005).
• AI literacy: take measures to support AI literacy (Art. 4): forecast uncertainty, calibration, and LLM triage errors (GUIDE-006).
• Pilot and communication: run in shadow mode next to the manual process for one season (GUIDE-009). Typical concerns are alert fatigue, over-trust in forecasts and loss of expertise; address them with source links on every alert, shown prediction intervals, and analysts owning risk assessment and escalation (GUIDE-008).

Risks, regulation & compliance

• EU AI Act: typically minimal risk — population-level decision support, no individual-level decisions and no Annex III category; if outputs drive individual measures, access to services or are embedded in a medical device, reassess. AI literacy (Art. 4) applies. Not a prohibited practice as long as no individual scoring or social-media profiling of identifiable persons occurs.
• EU health-security law: Regulation (EU) 2022/2371 on serious cross-border threats to health governs preparedness, EWRS notification and ECDC's epidemic-intelligence role [6].
• GDPR: aggregated wastewater and case counts are generally non-personal; line lists, small-area counts, sequences and social-media content can be personal or health data (Art. 9); use Art. 9(2)(i) public-interest basis, minimisation, aggregation thresholds and DPIA. Do not profile individuals from social media.
• Wastewater: the recast Urban Wastewater Treatment Directive (EU) 2024/3019 introduces surveillance obligations for health parameters; ethical review for sub-community sampling (e.g. single facilities) [7].
• Risks: false alarms and alert fatigue, missed events from media bias, over-trust in forecasts, stigmatisation of regions or communities. Mitigations: analyst verification, source transparency, calibrated intervals, communicated uncertainty, retrospective evaluation, versioned models.

Getting started

1. Inventory available data (wastewater plants, sentinel networks, lab feeds) and agree data-sharing rules.
2. Join EIOS and the relevant forecast hub; reproduce the hub baseline before building your own models.
3. Prototype multilingual triage on three months of archived articles, labelled by your analysts.
4. Build a back-test harness (weighted interval score, coverage) before any dashboard.
5. Run in shadow mode for one season next to the manual process, then agree escalation procedures.

Sources

1. WHO — Epidemic Intelligence from Open Sources (EIOS) — https://www.who.int/initiatives/eios
2. ECDC — European Respiratory Diseases Forecasting Hub (RespiCast) — https://www.ecdc.europa.eu/en/news-events/respicast-new-european-respiratory-diseases-forecasting-hub
3. Commission Recommendation (EU) 2021/472 on wastewater surveillance of SARS-CoV-2 — https://eur-lex.europa.eu/eli/reco/2021/472/oj
4. RKI — AMELAG, wastewater surveillance in Germany — https://www.rki.de/DE/Aktuelles/Publikationen/Epidemiologisches-Bulletin/epidemiologisches-bulletin-node.html
5. Hubverse — collaborative forecasting hub tooling — https://hubverse.io
6. Regulation (EU) 2022/2371 on serious cross-border threats to health — https://eur-lex.europa.eu/eli/reg/2022/2371/oj
7. Directive (EU) 2024/3019 (recast Urban Wastewater Treatment) — https://eur-lex.europa.eu/eli/dir/2024/3019/oj
8. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj