Epidemiological surveillance

It adds a statistical reference for what is "expected" at each point in time, so that a deviation can be detected before it becomes obvious to the eye. It applies when health indicators are monitored routinely and you want to react earlier, not once the rise is undeniable. Eyeballing an indicator or waiting for the monthly report picks up the anomaly late, because the eye only picks out what already stands out clearly. A surveillance system defines the expected behaviour of each indicator from its history, with its trend and seasonality, and flags when an observation moves outside that range, separating a statistically relevant signal from the fluctuation to be expected in the series. The difference is moving from "this month looks high" to "this value is outside the expected range for this time of year, based on the historical data", which is a basis for deciding where to look.

The expected value is modelled from each indicator's history, incorporating its trend and seasonality, so that the reference takes into account that what is normal at one time of year is not normal at another. This applies to all surveillance based on health time series. Without a model of what is normal, you cannot judge whether a value is anomalous: a value that looks high may be entirely expected for its season, and one that looks normal may be a deviation if it should have been low. Alert thresholds are then set from that expected behaviour, either with historical percentiles or with models that predict the expected value for each point in time. The alert is thus assessed against a constructed statistical reference, not against a hunch. The quality of that reference depends on having a sufficient and consistent history: with short or highly irregular series, the expected value is estimated with more uncertainty.

We use aberration detection algorithms that are established and widely used in surveillance, such as EARS, the Farrington algorithm or CUSUM, rather than improvised rules like "raise an alert if it goes up by 20%". This applies to the statistical detection of deviations in surveillance series. A fixed rule ignores the seasonality and inherent variability of each series, and produces either too many false alarms or too many late warnings. Depending on the algorithm chosen, these methods compare observations with expected references or thresholds built from historical data, and they are designed for different profiles: detecting gradual changes, sudden jumps or departures from a seasonal pattern. The algorithm is chosen and calibrated to suit the series and its reporting frequency, setting the balance between sensitivity and false alarms according to what each one costs the system. There is no single optimal algorithm: there is a suitable one for each type of indicator and surveillance question.

By adjusting the balance between sensitivity and false alarms, which is the central decision in any alert system: increasing sensitivity to possible deviations means tolerating more false alerts, and reducing false alarms means risking later detection. This applies to any surveillance system with alert thresholds. There is no universal sweet spot: the right balance depends on what each type of error costs the particular system. An alert that fires too often overwhelms whoever receives it and ends up being ignored; one that fires too rarely comes too late to act on. Calibration sets the thresholds and the algorithm's parameters with two things in mind: that asymmetric cost, meaning whether a false alarm or a delay is worse in each case, and the real response capacity of whoever receives the alert. The aim is not to eliminate both errors, which is impossible, but to place them where they are manageable for the use the system will be put to.

Yes. In fact, robust surveillance rarely rests on a single source: case notifications, laboratories, emergency departments and primary care can provide complementary signals about the same phenomenon or related ones, and combining them lets you corroborate a signal instead of depending on a single route. This applies when several sources record related indicators with different reporting frequencies and definitions. The work consists of harmonising those definitions and frequencies to build a coherent surveillance series or set of indicators, so that the sources are comparable and can be cross-checked. This matters because a signal that appears in only one source may be an artefact of that source, such as a change in how it is recorded, whereas checking that it appears consistently across several sources makes it more credible. The client provides the sources and knows how they are generated; the service integrates and harmonises them and builds the series or set of indicators that surveillance runs on.

This service detects statistical signals of indicators deviating from their expected behaviour and issues alerts on them; it provides the surveillance and alert system, not the epidemiological field response. This is an important clarification of scope. What the service delivers is the statistical layer: it models the expected value, detects when an indicator moves outside its range, integrates the sources and leaves the system running with its alerts. What it does not do is the epidemiological investigation that follows a signal, such as contact tracing, investigating the source or mode of transmission, or intervening in the field. That work belongs to the client's public health team, who know the context and have the competence and the mandate to act. The distinction protects the client: a statistical alert shows where and when to look first, but the epidemiological decision and the response are theirs. The service flags a statistical deviation worth reviewing; the public health interpretation and the action are the client's.

They share time series methods, but they differ in the data they monitor. Epidemiological surveillance works on health indicators recorded by the health system, such as incidence, emergency department, laboratory or primary care data, in other words notified cases and health events. Search-based infodemiology works on web search volume, an indirect signal of the interest or concern of people who search online, not a record of cases. They do not answer the same question: surveillance detects deviations in recorded health indicators; search-based infodemiology captures the perception dimension from what people search for. The two can complement each other, since the search signal can run ahead of the healthcare signal in some settings, but one monitors notified health data and the other an indirect syndromic signal. If the question is about how recorded cases or health events evolve, it is epidemiological surveillance; if it is about the search interest of people searching online, it is search-based infodemiology.

We deliver a statistical surveillance system configured for the agreed sources and reporting frequency, designed to run continuously rather than as a one-off analysis that is out of date on delivery. It includes a model of each indicator's expected behaviour, calibrated aberration detection algorithms, integrated time series and a dashboard with alerts when observations cross the defined thresholds. The system can update only as quickly as its sources: near-real-time data allow near-real-time monitoring, while weekly data produce a weekly view. We document the assumptions, thresholds and sensitivity calibration so that they can be reviewed and recalibrated as the series change. We build and configure the statistical layer; epidemiological review of each alert and the response remain with the client, because an alert is a signal of deviation worth reviewing, not confirmation of an outbreak.