I have read with some interest recent publications about the application of AI in pharmacy practice. Ben Michaels' Update describes a number of issues that need to be addressed before the plethora of EMR data can be used to try to drive pharmacy API models that might be useful, chief among which is the poor quality of such data and the wide variety of (or complete lack of ) standards by which concepts in that data are expressed.
I also noted Scott Nelson's report on which also calls out the need for both data stewardship and participation of pharmacists in the development of such models. You will need a valid login to AJHP's website to get a copy. It is the latter matter (pharmacist involvement in model development) that I think needs some discussion beyond the assertion that we must be involved. It has taken me some time to formulate my concerns, and I write this with some sense of trepidation.
I am one of a very few (Ben seems to be another) pharmacists who has actually participated in model development. It is not for the faint of heart. It can involve mind-numbing attention to detail to uncover what may be issues in the model's "thought process".
The easier (but still not easy) part of this governance involves identification of cases in which the model reports conclusions that do not align with current clinical or professional thought and research. The much harder part involves determining whether that misalignment is the result of an error in the model, or the discovery of associations (or even causations) that have yet to be described. If the goal of such governance is to make models replicate what we currently believe to be true, then we miss the opportunity to have AI reveal new knowledge to us that may improve clinical outcomes or enable automation of current rote professional tasks.
If a model indicates that certain diseases are more prevalent within certain populations, do we reject the model as biased, or do we attempt to learn whether such information might be true?
So who are the pharmacists we entrust to perform this governance? What training (technical, pharmaceutical, societal) should these governors have? How do they determine that sources of data are sufficiently large, sufficiently clean, and sufficiently diverse to yield reliably usable results? Are there cases where some model results are useful and others are not? Do they restrict the use of some models to certain fields or populations?
I don't know the answers to these questions. I know only that they need to be asked.
What do you think?
As always, the thoughts in this blog are my own.
Dennis A. Tribble, PharmD, FASHP
Ormond Beach, FL
tribbledennis@gmail.com