Skip to main content

Signals — anomaly detection

Automated screening of 1,200+ notified infectious disease time series (diagnosis × region, ÚZIS ISIN) for exceedances of the expected level using the Farrington/Noufaily method.

Our data Zdroj: ÚZIS CZ — ISIN Open Data (CC BY 4.0) Aktualizace: With every data pipeline run anomaly detectionsurveillanceearly warningÚZISstatistics

Infectious disease notifications form more than 1,200 time series (114 diagnoses × 14 regions plus national totals) — far too many for anyone to watch by eye. This page therefore screens them automatically: for every series it computes the expected endemic level from seasonality and history, and shows the months where the notified count exceeded the threshold.

Series above the expected level

načítám data…

Note: diagnosis and region names in the table come from the Czech source data.

How to read the table

ColumnMeaning
ReportedHow many cases were actually notified in the given month.
UsuallyThe model’s endemic level: how many cases this disease, in this region, at this time of year, would have in an ordinary year. Computed from the 2018–present history with past epidemics down-weighted.
Still normal up toThe upper limit of what ordinary fluctuation can still explain (99th percentile of the prediction interval). A count between Usually and this limit is business as usual; a signal starts above it.
Exceeded byHow many times further past the limit of normal than that limit is from the usual level. = exactly at the limit, = twice as far beyond it. The higher, the less likely it is chance.

An example from the table: hepatitis A in the South Moravian region — usually 1 case, still normal up to 5, notified 80. The observation is ~19× further beyond the threshold than ordinary fluctuation reaches (strength 19×) — chance practically cannot explain that.

Two badges replace strength where a statistical model makes no sense:

  • rare disease — a disease with at most ~5 cases in the entire history (diphtheria, yellow fever…). The threshold here is “more than an isolated case”: every cluster of 2+ cases is flagged.
  • outside prior occurrence — cases in a period where the disease previously did not occur at all (no history for this part of the year). It may be a genuine novelty or a change in reporting.

What to look at first: high strength together with high case counts (an epidemic under way). A row with small counts and strength just above 1× may be chance — with 1,200 series scored, we expect a few such rows every month.


What a signal means — and what it does not

A signal says one thing only: the number of notified cases is statistically well above what would be usual for this disease, in this region, at this time of year. It is not a confirmed outbreak. An exceedance may reflect a genuine cluster, but also a change in reporting practice, delayed notifications catching up, or plain chance. A signal is an invitation to look closer, not a conclusion.

The reverse also holds: silence is not evidence of calm. Data for the most recent months are always incomplete (notifications arrive with a delay), so a fresh rise may only become visible retrospectively.

How the expected level is computed

We use the Farrington/Noufaily method — the same algorithm the UK Health Security Agency runs weekly across thousands of laboratory reporting series:

  • the expectation comes from a quasi-Poisson regression model over the full available history (2018–present) with seasonality and trend,
  • past epidemics in the history are down-weighted, so that last year’s wave does not raise this year’s “normal”,
  • the threshold is the upper bound of the prediction interval (99th percentile); the strength of a signal states how many times the observation exceeded the threshold distance from the expectation,
  • diseases too rare for a statistical model (e.g. diphtheria) have their own rule: every cluster of cases is flagged.

How well the method works

We measured the method’s operating characteristics with a simulation study with known ground truth, and by backtesting against real epidemics:

  • an epidemic adding ten times the usual monthly variation is caught with a probability of ~96%, usually in its first or second month; five times the usual variation is caught in ~3 cases out of 5,
  • the 2024 pertussis epidemic (37,918 cases): first signal in November 2023, three months before the epidemic became a public topic — with no false alarm in the calm period 2021–2023,
  • real major epidemics are far stronger than these detection limits; November 2023 for pertussis was roughly fifteen to twenty times the usual variation.

Method: Farrington et al. (1996), Noufaily et al. (2012) · implementation, backtests and the simulation study live in the pathogensportal-db repository · Data source: ÚZIS CZ — ISIN, monthly aggregation