Trying out Google NotebookLM as a pharmacist
Pharmaceutical-driven analysis of Google's new service in relation to tasks analogous to medication management scheme
I recently came across Google’s new service, data analysis, that can be organized into workbooks. Workbooks were already available under the name Gemini, but NotebookLM represents a different level.
I immediately felt the professional need to analyze drug information using this new tool. Specifically, I was interested in whether this tool (if all the information is entered) would be able to answer a few questions when an average patient comes to me with questions about their medications.
I did not specifically simulate a medication management task, but I searched for answers to partially similar questions using Google’s language model.
It’s true that there have been websites dealing with drug interactions before, but I was thinking of more complex analyzes than that.
To do this, I prepared a new NotebookLM workbook and loaded some randomly selected summaries of product characteristics (SmPC) for healthcare pros, which I downloaded from the NNGYK drug database or found through a direct web search. This created a data lake consisting of some documents (in PDF and DOCX format). The disadvantage of the free NotebookLM account is the limited number of data sources, but it was still sufficient for an experiment.
The data sources are therefore text files in PDF or DOCX format containing official medical/pharmaceutical information relating partly to over-the-counter (OTC) (commonly known) medicines and partly to prescription (Rx) medicines:
eucalyptus oil, peppermint oil, and magnesium salicylate-containing drops to support liver and gallbladder health (OTC)
an over-the-counter (OTC) hemorrhoid ointment containing ephedrine/procaine/menthol/camphor and yarrow oil
a nasal spray containing xylometazoline (OTC)
a drotaverine-containing antispasmodic tablet (OTC)
one drop of vitamin D containing cholecalciferol (OTC)
a tablet containing diosmin and hesperidin used to treat varicose veins (OTC)
a paracetamol-based painkiller tablet (OTC)
two garglin insulin-containing injection preparations (Rx) of different brands
film-coated tablet containing levetiracetam (Rx)
three different brands of antihistamine preparations containing desloratadine (Rx)
a somatropin-containing injection preparation (Rx)
This resulted in a rather strange combination, but life can bring even more bizarre situations.
I tried to take the questions from everyday practice. In its answers, NotebookLM analyzed the documents throughout. Some of my test questions:
Summarize what products are mentioned in the sources!
Which source mentions that the medicine contains lactose?
Is there any indication in the sources that any two medicines mentioned in the sources cannot be taken together?
Which preparations should not be used during pregnancy?
Which preparations can cause skin rashes?
Which preparations should not be used together with iron preparations?
Are there any preparations among the sources that have the same active ingredient, meaning that taking them together could lead to an overdose?
Overall, NotebookLM, as a closed world and the processing of resources within it, enabled a sufficiently quick overview and summary, as well as a pharmacist-oriented evaluation in a matter of minutes.
It is true that accurate questions and a reliable, comprehensive set of source material are necessary for the language model to work. The answers provided are easy to check, as all it takes is a click on the given material, and we can check the answer ourselves.
This analysis is intended for a professional audience (healthcare professionals and decision-makers, health statisticians, data scientists, manufacturing and wholesale professionals, etc.) and is for informational purposes only. It should not be used as a basis for any layperson’s decisions regarding individual therapy! These analyses reflect private opinion!
Artificial intelligence solutions were not used in the preparation of this report. The report was written in Hungarian and is the result of human work. The translation was assisted by the DeepL tool, with additional linguistic verification by LT Assistant.