Social media and digital infodemiology

Because reading what circulates by hand gives a sample biased towards the most visible and the most recent, whereas a systematic analysis quantifies the captured corpus of conversation on a topic. This applies when you want to know what is really being said about a vaccine, a disease or a public health measure in the digital space, with data instead of impressions from a handful of posts. Reviewing a few messages leads you to mistake the most striking for the most widespread, and to miss the narratives that do not catch the eye. The analysis collects the available corpus through defined APIs and search criteria, identifies the narratives present and quantifies the weight of each, giving a systematic reading of what the data cover, not a selection of examples. It is the difference between an impression formed from whatever you happened to see scroll past and systematic surveillance of the conversation about a health topic.

It can detect a narrative accelerating before it dominates the captured space, which is the window in which a communication response has the most room to act. This applies when the aim is not only to describe what circulates but to react in time to a misinformation narrative that could erode trust in an intervention. Once a hoax is trending, containing it usually comes too late; the value of surveillance lies in distinguishing a real rise from a normal fluctuation before that point. To do this, the propagation dynamics are modelled and the evolution of volume and sentiment is monitored, so that an alert is triggered by a robust signal and not by noise. The alert should be read as a way to prioritise review and response, not as a sure prediction of virality: it estimates where and when to look first, but it cannot ensure that all misinformation is detected or predict with certainty what will go viral. It is an early signal that improves reaction time, not a crystal ball.

Narratives are approximated by combining topic modelling with expert interpretation, grouping the messages by the latent topics they share instead of reading them one by one. This applies to analysing large volumes of digital conversation about a health topic. Methods such as LDA or BERTopic propose thematic groupings that emerge from the corpus, showing what is being said and around which axes, and sentiment analysis adds the tone in which a policy, a vaccine or a measure is discussed. This turns thousands of scattered messages into a map of the dominant narratives and their relative weight. It is worth being clear about the scope: topic modelling proposes statistical groupings, but the narrative is constructed and interpreted by the analyst, and the stability of the topics and their semantic coherence should be checked before they become interpretive categories. It is not an automatic, infallible reading of the discourse but a structuring of the content that is then validated and interpreted, and whose quality depends on how the query is defined and on the captured corpus.

Network analysis can identify the nodes that amplify a narrative most and estimate what share of the activity appears automated or coordinated, as opposed to organic conversation. This applies when what matters is not only what is said but its diffusion structure: the content of a hoax matters less than how, and through whom, it spreads. The analysis reconstructs the propagation cascades, locates the amplifiers and characterises patterns consistent with automated or coordinated activity, using behavioural indicators and the structure of the network. It is important to be precise about what this allows you to claim: it estimates the probability that an account or a pattern is automated or coordinated based on its behaviour; it does not label a specific account as a bot with absolute certainty. The result directs a communication strategy towards the points where the diffusion network concentrates most amplifying power, instead of responding diffusely to all the noise.

No. Classifying a piece of content as false is anchored in health evidence and in recognised institutional or fact-checking sources, not in the service's own judgement. This is an important clarification, because "misinformation detection" can be read as the analyst ruling on the truth, and that is not the case. What the service does is identify and quantify circulating claims, check them against existing fact-checks and established health evidence, and analyse how they spread. The boundary is clear: the service detects, measures and characterises how claims circulate, relying on external fact-checking references; it does not set itself up as the arbiter of scientific truth, nor does it replace the judgement of health authorities or fact-checkers. This keeps the analysis on its own ground, the surveillance of the conversation and how it spreads, and leaves the substantive verdict where it belongs: with the evidence and recognised sources.

Working with data from digital platforms has ethical and legal limits that determine what can be collected and how results are reported, and respecting them is part of the design of the analysis. This applies to any collection of social media data. The extraction is designed to respect each platform's current access terms and the applicable data protection law, working with public conversation and avoiding the exposure of personal information. We also avoid reporting results that would identify or needlessly single out specific individuals when the aim is population-level or communication-related. In addition, the scope and bias of the sample are documented (which platforms, period and language it covers, and which are left out), because the conclusions are only valid within what the data actually represent. Bear in mind that no capture covers the whole conversation: each platform has its own access limits, and what is analysed is a sample with known biases, not the full universe of what is said. That transparency about scope is part of what makes the analysis sound.

They differ in what their data capture. Social media and digital infodemiology analyses the conversation on platforms (what is said, with which narratives and what sentiment, how it spreads and who amplifies it) using text and network analysis methods. Search-based infodemiology analyses aggregated, anonymous search volume (what people search for about health, when and where) as a syndromic epidemiological signal, using time series methods. They do not answer the same question or work with the same type of data: one studies the content and circulation of messages; the other, how search interest evolves through trends, seasonality and geographical comparison. If the question is what is being said and how a narrative or a hoax circulates, it is social media infodemiology; if it is how a population's interest in or concern about a topic is evolving, it is search-based infodemiology. They are sister services covering different sides of the same information phenomenon.

You receive the surveillance analysis of the captured digital conversation on the defined topic: the map of identified narratives with their relative weight, the sentiment analysis, the reconstruction of the diffusion networks with the amplifiers and the likely automated or coordinated activity, and the modelling of the propagation dynamics, with the rises that may warrant an alert or priority review. This is for those who need to shape a health communication strategy knowing which narratives dominate, who amplifies them and how they spread, or to detect a misinformation narrative in time. The deliverable documents the scope of the capture (the platforms, the period and the language, and what is left out) so that the reading stays within what the data cover, and anchors the classification of claims in external fact-checking sources. The service collects the public digital data, and analyses and interprets them; the health authority or the communications team decides on and implements the communication response.