Public health surveillance outbreak detection rests on a handful of numbers, and most of the published ones come from simulations, single sites or older studies. That matters before you read any of them. The figures describe specific settings, and none of them is a current national audit of a US system.
How an alert gets triggered
Most systems compare routine counts, such as reported cases, laboratory results, clinic visits or sequence fingerprints, with an expected baseline. They raise a flag when counts exceed a threshold. Germany's SurvNet, described in an undated arXiv preprint, analyses thousands of weekly case-count time series for aberrations from an endemic baseline, mostly using regression or statistical process control. That description has not been checked against the institute's own documentation.
A measles modelling study defines an outbreak alert as a seven-day rolling average of test-positive cases above a threshold. In that model, the optimal threshold ranged from about 1.72 to 18.73 test-positive cases per day, depending on testing rate and background noise, and it generally rose as testing rose. Those values depend on the model's assumptions. They are not the setting of any real health department.
Error rates: sensitivity, specificity and PPV
Sensitivity is the share of real outbreaks an alert catches. Specificity is the share of non-outbreak periods left unflagged. Positive predictive value (PPV) is the share of flagged cases that turn out to be real. Raising a threshold usually trades one measure against another, so no single figure settles whether a system is good.
A simulation of 21 outbreak-detection methods, published in PLOS ONE, reported false-positive rates from 0.7% to 59.9%, detection probability from 43.3% to 88.7%, PPVs from 6.5% to 68.4%, and negative predictive values above 94%. Methods with high specificity often had low sensitivity. These are ranges across methods in a simulation, not measurements of a deployed system.
Single-setting studies show a similar spread. In a Malaysian dengue evaluation covering 2017 to 2023, a proposed endemic-channel method had sensitivity 0.60, specificity 0.82, PPV 0.35 and negative predictive value 0.93. The conventional method it was compared with had sensitivity 0.56 and PPV 0.23. The research retrieved only the abstract, and the journal name could not be confirmed.
What a PPV of 0.35 means, and whose reading it is
Read directly, a PPV of 0.35 means about 65% of flagged cases in that evaluation were not cases. That subtraction is our arithmetic on the reported figure, not a finding the research attributes to the authors. For a health department, a low PPV points to many follow-up investigations that turn up nothing.
A single-site evaluation in Peru, which the research describes as a U.S. Navy study, reported sensitivity of 100%, specificity of 97% to 99%, and PPV of 27% to 46% for three syndromic algorithms on diarrhoeal disease. The 100% sensitivity rests on a small number of events and should not be generalised.
How quickly signals arrive, and how much they capture
Timeliness and completeness are separate problems. A signal can be complete and late, or prompt and partial.
For dengue in Rio de Janeiro, a Statistics in Medicine analysis found that cases should be reported within seven days, yet under 50% were reported within one week, under 75% within four weeks, and no more than 90% within nine weeks.
A US analysis of notifiable diseases, retrieved only as a ResearchGate copy of a paper roughly twenty years old, reported that for several enteric and bacterial diseases under 40% of cases were reported within one incubation period. Median reporting delays ran from 10 to 19 days depending on the disease. Treat these as low-confidence figures until they are checked against the journal version.
On completeness, a state health department's foodborne-outbreak training guide gives an indicative range: depending on the disease, only 5% to 80% of cases are reported. The guide gives no primary citation, so the range shows how wide the uncertainty can be, not a national rate.
Routes that do not start with case counts
PulseNet, the CDC network for foodborne illness, starts from DNA rather than from counts. Laboratories produce whole-genome sequence fingerprints, CDC staff compare them to find clusters of matching patterns, and state and local health departments are then notified of a possible outbreak. An NCBI Bookshelf chapter describes PulseNet as a national network that made rapid identification of multistate clusters possible, but that chapter is historical and is not used here for current figures.
Event-based systems, which draw on media and online reports, vary just as widely. In one time-series comparison, HealthMap's PPV ranged from 9% to 100% and its timeliness from 13% to 94%. EIOS ranged from 0% to 100% for PPV and from 0% to 92% for timeliness. The research saw only the abstract.
Why figures get revised
CDC says FluView data each week are preliminary and may change as more data arrive. The system shows whether influenza activity is rising or falling, but it does not directly give the number of illnesses. A figure can be right about direction and still be revised later in count.
Wastewater is often presented as an earlier signal than clinical case counts. A secondary explainer says so, but the research did not confirm that timing or the claim against CDC or peer-reviewed sources, so treat it as unverified.
Where the evidence runs out
Most performance figures here come from simulations, single sites, other countries or older sources. No current national PPV or sensitivity audit for an operational US system was retrieved. Failures of digital and search-based systems are not documented in the research, and the track record of Google Flu Trends could not be confirmed, so this article makes no claim about either.
Before repeating any alert figure, check which measure it is, which setting produced it, and whether it comes from a simulation, a single site or an audit. A sensitivity figure answers a different question from a PPV, and a number from Malaysia or Peru does not describe your own region.