Since I now have a piqued interest in the application of artificial intelligence in pharmacovigilance, I was interested to read the findings from a study using machine learning to identify ICSRs in social media. Until now, my impression of ADRs reported in the social media was that identifying valid ICSRs was more than challenging due to their diversity, complexity, colloquialisms and anonymity. Also, it seems the regulatory authorities acknowledge these challenges as it is not yet a regulatory requirement to monitor social media as a source for ICSRs. But is this looking set to change if machine learning and automation become the norm?
This proof-of-concept study by Shaun Comfort and colleagues, recently published in Drug Safety, showed that machine learning can be successfully used as a classifier to identify valid ICSRs, at an accuracy level similar to humans. But here’s the thing, the rate at which the machine was able to identify valid ICSRs was strikingly faster than humans – taking just 48 hours versus 44,000 hours.
We already know that there is potential added value in reports in social media, where patients can freely express their drug experiences, including effects on quality of life which may not be reported through the traditional channels. And although there are worthy initiatives such as WEB-RADR which are exploring the use of mobile apps and social media in this setting, this still relies on proactive reporting by the patient. Even if use of these tools becomes widespread, there will likely still remain a significant volume of informal reports of ADRs discussed in various social media channels. Especially as social media usage trends are not showing any signs of slowing down, with the number of active users worldwide expected to reach over 3 billion by 2021. So finding smart, efficient and scalable machine-based approaches to identify ICSRs looks to be an effective solution to overcoming the accepted challenges.
In summary, this study adds to the growing evidence that social media is a valuable source of safety information, and that modern-day machines are best placed to handle this modern-day reporting style, be it for the purposes of signal detection, ICSR reporting or risk-benefit analysis.
Image credit: NordWood themes