Millions of patients may be reporting adverse events. Is pharma listening?

Somewhere online, right now, a patient is describing what happened to them after they started a new medication. They are not filing a report. They are not calling a helpline. They are simply telling their story, in their own words, to whoever might be reading. That story contains information which, in the right hands, could matter enormously for drug safety. In most cases, nobody in a position to act on it will ever see it.

This is not a technology problem. It is a prioritisation problem. The technology, as we now know, is ready.

What we have been missing

Pharmacovigilance is defined as ‘the science and activities relating to the detection, assessment, understanding and prevention of adverse effects or any other medicine related problem’. As such it has historically depended on formal reporting mechanisms. These systems are valuable and necessary. They are also, by their nature and design, rigid and often incomplete. Patients do not always report adverse experiences through official channels. What they do, with remarkable consistency, is talk, often online, and increasingly at scale.

Social media platforms host millions of patient-generated conversations about medicines, vaccines, and treatment experiences every single day. Spontaneous, unsolicited, and unfiltered, they represent something structured reporting systems can rarely capture: the lived experience of patients in real time, expressed in real language.

The challenge has traditionally been that this data is extraordinarily difficult to work with. Informal spelling, fragmented sentences, colloquial descriptions of symptoms, and a complete absence of clinical structure have historically defeated systematic analysis. For this reason, social media data has largely been set aside, treated as too messy to be reliable.

That assumption deserves to be challenged directly: messy data is not the same as low-value data.

Proving the signal is real

At CREATION.co, we put this to a rigorous test. The problem was immediately apparent. Someone writing “my chest felt tight after the shot” or “had a pounding headache all night” is reporting a potential adverse event, but extracting that information reliably, understanding it in context, and distinguishing it from unrelated health commentary requires something more than pattern recognition.

Our approach layered different analytical capabilities on top of one another. A biomedical language model identified relevant clinical entities within the text. A second model interpreted complete phrases rather than individual words, better reflecting how patients actually construct their descriptions. A large language model then evaluated each candidate signal against its surrounding narrative, asking in effect: is this person actually describing something that happened to them as a result of their treatment?

That final reasoning layer proved decisive. It filtered out sarcasm, hypothetical mentions, and unrelated commentary. It interpreted informal patient language and aligned it with clinical terminology. It transformed a noisy stream of text into a structured, reviewable set of candidate safety signals, at a level of performance that closed the gap between controlled test conditions and the unpredictable reality of live social media data.

Critically, the system was designed so that human safety experts remain integral throughout. Machine learning surfaces the signals. Clinicians evaluate them. The judgement, ultimately, remains human. This is worth calling out, in an age where AI is increasingly the buzz term and uncertainties around AI’s impact on jobs are, in some cases, real. However, in this case, the clinician in the loop is a vital part of the process, and so it should be.

The question has changed

For years, some had the working assumption that social media data is too unreliable to be useful for pharmacovigilance. But CREATION.co’s work in pharmacovigilance on social media spans more than fifteen years – we first proved that it was possible to identify and classify adverse events reported on social media in 2010. Our work has consistently demonstrated that there was no lack of data available on drug safety incidents on social media. Our latest research, however, leverages cutting-edge tools to classify pharmacovigilance signals of interest consistently at scale.

In drug safety, time matters. An adverse event pattern identified earlier means earlier investigation, earlier intervention, and potentially, earlier protection for patients who might otherwise be harmed. Social media, approached with the right architecture, can function as a genuine early-warning layer in the safety ecosystem, surfacing emerging concerns that might otherwise remain invisible for months or even longer.

The question has shifted, therefore, and it is no longer a case of whether credible safety signals can be detected in social media data. Indeed it is a question of whether the industry, and those who regulate it, are moving quickly enough to act on that fact. Regulatory frameworks in this area have not kept pace with what the science can now deliver. That gap will close. The organisations that engage now will be better placed to shape how it does.

The same opportunity exists beyond pharmacovigilance

This capability has implications well beyond drug safety monitoring, because the underlying challenge is the same wherever valuable intelligence is embedded in large volumes of informal, unstructured data.

Consider what happens when significant new clinical trial data enters the public domain. The conversations that follow among healthcare professionals, on open social professional networks and in digital communities, represent real-world expert commentary on emerging science, forming in real time and, in the main, are largely unobserved. For any organisation seeking to understand how new evidence is being received by the clinical community, or how medical opinion is forming around a new therapy, the cost of not listening is not abstract. Decisions about medical education, clinical engagement, and commercial strategy are often being made against an incomplete picture, because the most candid and immediate professional reaction to new science is happening in spaces that are not being monitored.

The analytical capability required to surface intelligence from this kind of data is, in essence, the same capability I already mentioned in the pharmacovigilance context earlier. The technological architecture exists; the question is whether organisations are aware, and whether they are prepared to use it.

Conclusion: the conversations are happening. Is your organisation listening?

The signals in the social media data are there. Patients are describing their experiences with medicines in ways that contain genuine safety-relevant information. Healthcare professionals are discussing clinical evidence in ways that contain genuine strategic intelligence. This information is not hidden. It tends to merely be unstructured, messy, informal, and idiosyncratic in nature, so therefore, for most organisations, it tends to be invisible.

At CREATION.co, we have changed this. We specialise in the analysis of digital and social data to generate intelligence that matters for healthcare and life sciences organisations. Our insight platforms and bespoke consulting enable teams in many of the world’s leading organisations to discover what’s changing, and why, and to take action faster.

What could it mean to you, if you were able to learn from healthcare professional and patient experiences shared online? Which signals may your organisation not be aware of currently, and what decisions are being made in the absence of this potentially vital information?

We would very much welcome that conversation.