“Far more people die from adverse reactions to prescriptions and over-the-counter medications than illegal drugs.”

Stephan Fried, Bitter Pills

Adverse drug events (ADEs), one of the most prevalent health care related harmful effects that affects around 2 million Americans annually (vs 10,000 who die from illicit drug use), can be preventable (medication errors) or non-preventable (adverse drug reactions). More than half of these ADEs are preventable. As ADEs are documented via diagnostic codes in health systems, they are very likely to be vastly underestimated.

This Lancet Digital Health collaborative work from academic and industry is a very good scoping review of the use of AI for reducing the frequency of adverse drug events. The AI techniques searched included: neural networks, tree-based algorithms, support vector machines, and natural language processing. There is a differentiation between prediction models (86%) and early detection models (14%) in the total number of studies (n=78) that were international in scope.  Four medications (analgesics, antineoplastics, antibiotics, and anticoagulants) make up about half the studies (not surprising).

The majority of studies were published within the past 5 years, therefore indicating that this is a focused area of investigation with AI. The majority of studies, however, assessed the technical algorithm performance but not the utility of the tool in the clinical setting. In addition, most studies did not utilize the precision-recall curve with these unbalanced data sets. Lastly, this lack of accountability of AI tools beyond the publication remains a major issue with AI studies in clinical medicine in general. While there is not an intentional attempt to forego the clinical utility of these tools, it is nevertheless a real issue with the investigators.

The authors do correctly point out that the addition of genetic information (presumably pharmacogenomic information in particular) and unstructured notes (even more relevant with innovative transformer NLP tools) will further advance this burgeoning field of ADEs and AI tools. In addition to the aforementioned authors’ comments, another important future source of information is real world data and the resultant evidence generation; this source of data will be perhaps the most important in the future for ADEs. Lastly, the highest value of these AI tools will be from a real-time capability of these tools to stop a preventable ADE before it happens.

Read the full paper here: