> ## Documentation Index
> Fetch the complete documentation index at: https://www.worldmonitor.app/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Disease Outbreak Alert Level

> How World Monitor derives ALERT, WARNING, and WATCH tiers for disease outbreaks using a keyword classifier over item titles and descriptions.

The "Disease Outbreaks" panel labels each item with one of three categorical
levels: **ALERT**, **WARNING**, or **WATCH**. This page documents how that
label is derived so the value is reproducible from the inputs displayed
on the panel.

## Editorial weights

Levels are assigned by a keyword classifier in
`scripts/_disease-outbreaks-helpers.mjs::detectAlertLevel` that matches
whole-word tokens (case-insensitive) in the concatenated item title and
description.

| Level     | Keywords (any match)                            |
| --------- | ----------------------------------------------- |
| `alert`   | `outbreak`, `emergency`, `epidemic`, `pandemic` |
| `warning` | `warning`, `spread`, `cases increasing`         |
| `watch`   | (fallback when no keyword matches)              |

The keyword set is **editorial** — it is not derived from a published
classification index (ECDC, WHO EWARN, ProMED). Treat the level as an
opinionated newsroom triage signal, not a clinical assessment.

## Inputs

The classifier sees the item's `title` and `desc` fields after they are
normalized by the per-source parser:

* WHO Disease Outbreak News (`whoNormalizeItem`)
* CDC HAN and Outbreak News Today RSS (`rssNormalizeItem`)
* ThinkGlobalHealth disease tracker, backed by ProMED-sourced real-time alerts (`tghNormalizeItem`)

There are no numeric scoring weights. The label is purely categorical and
the order shown in the panel is `alert > warning > watch`.

## Known limitations

1. **Newsroom phrasing, not epidemiology.** A press release that uses the
   word "outbreak" once will be promoted to `alert` even if the situation
   is minor. Conversely, a serious situation reported without any of the
   keywords will fall through to `watch`.
2. **English-only keywords.** Non-English titles will systematically
   under-classify until the source is translated upstream.
3. **No multi-word context.** "False outbreak rumor" will still match
   `outbreak` and be promoted to `alert`.

These limitations are accepted for now because the source feeds are
predominantly English-language press releases and the panel is positioned
as a "what to look at next" surface rather than a risk score.

## Change protocol

If the keyword sets in `DISEASE_ALERT_KEYWORDS` or `DISEASE_WARNING_KEYWORDS`
change:

1. Bump `ALERT_LEVEL_METHODOLOGY_VERSION` in the same file. The seeder
   stamps this onto the published payload as `alertLevelMethodologyVersion`,
   so a bump observably propagates to clients and the canonical key.
2. Update the table above to match.
3. Add a regression test in `tests/disease-outbreaks-seed.test.mjs` covering
   the new keyword behavior + at least one expected non-match (substring
   guard).
4. Note the change in the next release's changelog so cached / pre-rendered
   payloads can be invalidated if necessary.

## See also

* Source: `scripts/_disease-outbreaks-helpers.mjs` → `detectAlertLevel`
* Consumer: `src/components/DiseaseOutbreaksPanel.ts`
* Tests: `tests/disease-outbreaks-seed.test.mjs`
* Tracking: [#3791](https://github.com/koala73/worldmonitor/issues/3791)
* Pattern reference: [CII risk scoring methodology](/methodology/cii-risk-scores) (same disclosure pattern, larger scope)