An artificial intelligence (AI)-powered chatbot shows potential as an educational tool for clinicians and patients who have questions about pharmacogenomics. But the technology isn't yet ready for deployment, says Philip Empey, associate director of pharmacogenomics at the Pitt/University of Pittsburgh Medical Center Institute for Precision Medicine in Pennsylvania.
In a report published in the June 2024 issue of the Journal of the American Medical Informatics Association, Empey and colleagues examined how well an experimental chatbot, or AI assistant, interpreted the results of pharmacogenomics tests related to the use of statins.
“There’s a lot of enthusiasm in all fields about the opportunity for chatbots to support activities that are happening in all business areas — including healthcare,” Empey said.
In the field of pharmacogenomics, he said, test results are often actionable, but clinicians and patients may need guidance on how test results influence treatment decisions.
“So, AI is being looked at as a potential tool that could help us with that,” Empey said.
The AI assistant was developed using GPT-4, the latest iteration of OpenAI’s generative AI system, and it was intended as a proof-of concept experiment for pharmacogenomic counseling.
One of the benefits of GPT-4 is its support of retrieval-augmented generation (RAG) involving user-specified data sources. The RAG portion of this study included a dataset derived from Clinical Pharmacogenomics Implementation Consortium (CPIC) prescribing guidelines and relevant literature.
The study evaluated the AI assistant’s responses to specialized pharmacogenomics questions, such as the appropriateness of atorvastatin in patients with diminished organic anion transporting polypeptide 1B1 function.
A baseline assessment identified some deficiencies in the AI assistant’s responses. For example, the tool gave potentially problematic dosing guidance to patients, and it did not identify itself to patients or clinicians as an AI tool whose responses should not be considered medical advice.
Those problems were solved by constructing system and user prompts that defined behaviors and boundaries for the AI assistant, emphasized the reliance on evidence- and context-based responses, and reduced the likelihood of hallucinatory output.
Empey said he was surprised by how genuine some of the tool’s responses were when it was guided by the prompts.
“It’s not really empathy,” Empey said. “But the ability to provide context-aware and appropriate responses was more empathetic, I think, than I would have expected.”
He said additional research is needed to determine how this type of AI assistant might perform in clinical practice, where accuracy and appropriate context are vital.
“This was a very early and preliminary evaluation, just to kind of see where things were at,” Empey emphasized.
The study focused on statins because of the wealth of pharmacogenomic data available for this drug class. Empey said medications used in cardiology, mental health, and primary care could be potential candidates for evaluating the AI assistant.
“There are now 25 CPIC guidelines that ASHP has endorsed,” Empey noted. He added that it’s not known how tools like the AI assistant might respond to queries about multiple drug–gene pairs and clinically complex situations.
“That’s where I think future evaluation is needed to understand the [clinical] capability of these tools,” he said.
For more general uses, Empey said, AI chatbots could serve as educational tools for patients and clinicians and, potentially, to help broaden the implementation of genomics-based medicine.
ASHP’s position statement on the use of AI calls for implementing the technology where it is known to work and where it helps pharmacists make better decisions and appropriately focus their problem-solving expertise. The position statement, which is undergoing review and revision, describes potential roles for AI in informatics, clinical decision support, pharmacy operations, and education.
Empey urged pharmacists to take advantage of opportunities to learn about AI and its potential uses in pharmacy practice.
“I think the more pharmacists are involved in this burgeoning field, the more we could use it to enhance what we do,” he said. “It could be in medication reconciliation; it could be in pharmacogenomics. But it could also be any areas where we would play a key role in the medication use process."
