A common misconception worth correcting 💊
It is tempting to believe that an authorised medicine is one whose effects are all known. That is false, and health authorities know it perfectly well. Authorisation marks the start of monitoring, not the end of assessment. This system is what we call pharmacovigilance.
It is tempting to believe that an authorised medicine is one whose effects are all known. That is false, and health authorities know it perfectly well. Authorisation marks the start of monitoring, not the end of assessment. This system is what we call pharmacovigilance.
This morning we reported on an analysis of 400,000 online messages about side effects. Here is the system within which this kind of study makes sense.
Why authorisation is not enough
Before being authorised, a medicine goes through clinical trials. They are rigorous, and they have three structural limitations.
Numbers. A few thousand participants, sometimes a few tens of thousands. An effect that affects one person in fifty thousand stands a good chance of never appearing.
Duration. A few months to a few years. An effect that only shows up after long-term exposure cannot be observed.
Selection. Trials often exclude pregnant women, frail elderly people, and those taking other treatments. The real-world population is far more diverse.
When millions of people take the treatment in real life, effects that were previously invisible appear. Pharmacovigilance exists to catch them.
How it works
The principle rests on reporting. Doctors, pharmacists and, in many countries including France, patients themselves can report a suspected adverse effect.
These reports are centralised and analysed. A single report proves nothing. But when several reports describe the same effect with the same treatment, more often than chance would explain, a signal emerges. It then triggers a more rigorous investigation, which may lead to a change in the patient information leaflet, a restriction of use, or, exceptionally, a withdrawal.
The well-known weak point 📉
The system relies on reporting, and people report little. Generally accepted estimates suggest that only a small fraction of adverse effects are actually reported. Patients do not think to do it, doctors lack the time, and many mild effects are never linked to the treatment. This is precisely the gap that analyses of online messages attempt to fill: going out to find what people say without going through the official channel.
The system relies on reporting, and people report little. Generally accepted estimates suggest that only a small fraction of adverse effects are actually reported. Patients do not think to do it, doctors lack the time, and many mild effects are never linked to the treatment. This is precisely the gap that analyses of online messages attempt to fill: going out to find what people say without going through the official channel.
What AI brings
Three genuine contributions, and none of them changes the rules.
Reading what nobody reads. Hundreds of thousands of patient messages, consultation notes, the medical literature. AI can spot regularities in volumes that no human team would process.
Understanding ordinary language. A patient writes "I'm knackered all the time" where a form expects "asthenia". Matching these formulations is exactly what language models do well.
Detecting earlier. Spotting a signal a few months sooner can prevent thousands of exposures.
What AI does not change is the next step: a signal must always be confirmed by rigorous methods before it becomes knowledge. That is the division of labour we described in relation to medical AI: spotting, yes; concluding, not without evidence.
What you can do
The most useful step is also the simplest: if you think a medicine is causing you an unexpected effect, report it. In France, patients can do this themselves on the official portal for reporting adverse health events, and discuss it with their doctor or pharmacist.
Every report counts, precisely because too few are made. This is how the effects that trials could not see come to light.