AI’s Role in Advancing Critical Care Pharmacotherapy: A Researcher's Insights

The integration of Artificial Intelligence in health care has gained traction in recent years, particularly with large language models that excel in answering medical questions. Yet, one area that remains underexplored is their application to critical care pharmacotherapy — an essential domain where quick, accurate medication decisions can be a matter of life or death.
In a new study published Jan. 9, Sheng Li, a Quantitative Foundation Associate Professor of Data Science, was part of a research team that tested whether large language models (LLMs) could match the decision-making ability of Doctor of Pharmacy students in the high-stakes environment of intensive care units.
Exploring the Potential of LLMs in Critical Care
Joining Li in this study were researchers from the University of Virginia Department of Computer Science, the University of Georgia College of Pharmacy, and the University of Colorado Skaggs School of Pharmacy and Pharmaceutical Sciences. The team sought to evaluate the effectiveness of various LLMs — GPT-3.5, GPT-4, Claude 2, Llama2-7b, and 2-13b — on a set of 219 multiple-choice pharmacotherapy questions designed to assess knowledge in critical care. The results revealed that ChatGPT-4 led the way in terms of accuracy, particularly in knowledge recall questions. However, the Pharm.D. students still outperformed the models in applying their knowledge to more complex, real-world clinical scenarios.
And in critical-care venues like an ICU or emergency room, precise and rapid decision-making is absolutely crucial to saving lives. Having an LLM quickly and efficiently pull knowledge from vast datasets is great. But without the nuanced understanding and reasoning to apply that knowledge effectively, AI’s usefulness in this kind of complex health care environment will remain limited.
Bridging the Gap: Challenges and Opportunities
One of the major challenges in this field is the customization of AI models for specific health care specialties. Li emphasized the importance of models with domain-specific materials and training to improve performance and outcomes. A customized ChatGPT model named PharmacyGPT was built on a dataset of relevant pharmacy school course materials as a proof of concept and tested in the study. While this model showed promise, more sophisticated training and access to specialized data are needed to achieve substantial improvements.
According to Li, a critical barrier remains the lack of access to real patient and pharmaceutical data, which is essential for further training and refining these models.
"If we have access to this kind of data, I believe the model performance will be improved significantly," Li said.
However, privacy concerns and the ethical handling of patient data continue to complicate this path forward.
The Public Interest: AI in Medication Safety
As AI continues to evolve, its potential to reduce medication errors and improve patient safety has generated public interest. Li suggests that LLMs can serve as invaluable tools to assist health care providers, particularly in reducing the cognitive load and the risk of human error.
“I would view the current AI models as tools of assistance to doctors,” Li said. “Even humans make mistakes, right? I think the better way to frame this is: Can we use these AI tools to assist the doctors to ensure that there will be a reduced number of mistakes in the future?”
Li believes the AI tools can be used to provide alternative suggestions or project potential risks. This would allow doctors to make more informed decisions based on their own experience along with these new kinds of inputs to make better plans for patients.
However, he was quick to add that the final responsibility must always lie with the healthcare professional.
"Doctors should be the final gatekeepers to make sure everything is safe and correct before implementing plans to their patients,” he said.
The Future of AI in Health Care Education
This study also highlights the role AI can play in medical and pharmacy education. Another line of research Li is pursuing is how AI tools can help generate personalized training materials and questions for students. This could enhance learning outcomes by providing tailored feedback and challenging scenarios based on individual progress, ultimately helping future health care professionals build stronger decision-making skills.
Additionally, as AI models become more transparent and interpretable, they could potentially serve as teaching tools — not only answering questions but also revealing the reasoning behind their answers.
“In the medical domain, we believe transparency and interpretability play a major role,” Li said. “We want to showcase the intermediate reasoning process to doctors and patients so that they can believe the AI decision-making process is actually meaningful to them.”