Identifying the best uses for these tools, whether in patient education, data collection, improving efficiencies, or other uses, could significantly improve the future of pharmacy.
Although there are still some limitations, the advent of artificial intelligence (AI) and machine learning (ML) could significantly accelerate and support pharmaceutical development and research, according to Shawn Xiong, PhD, an assistant professor in the James L Winkle College of Pharmacy at the University of Cincinnati. On Friday, Xiong presented the American Pharmacists Association (APhA)-Academy of Pharmaceutical Research and Science (APRS) Keynote at the 2024 APhA Meeting and Exposition.
Image credit: Tierney | stock.adobe.com
AI is a broad field that encompasses multiple tools and systems, Xiong said. However, the overarching theme is that AI utilizes computers to simulate human cognitive processes and intelligent actions including learning, logical reasoning, cognitive thinking, and strategic planning.1
“Basically, the computers can learn and improve their performance over time,” Xiong said.1
In the pharmaceutical industry, the top 3 use cases are predictive maintenance, quality inspection and assurance, and manufacturing process optimization. In clinical practice, AI tools can be used for early detection and personalized treatment, automated image diagnosis, prescription error analysis, and much more. Xiong said early detection and personalized treatment is a particularly exciting area for the use of AI in clinical practice.1
“This is not only about identifying the disease itself,” Xiong said. “It’s also about providing insights into the unique situations of the patient so that we can make personalized treatments for the patient, especially considering the concept of patient-centered care.”1
ML is a subset of AI in which algorithms enable computers to learn patterns from empirical data. In this case, Xiong described computers as a chef. Humans provide the ingredients (i.e., the data), but the computer creates the recipe and the final product.1
The theoretical basis for AI and ML began in the 1950s and 1960s, although practical applications did not emerge until the 1990s. with the big data revolution of the 2000s, development of these tools accelerated exponentially, and has only continued to grow since then. With the AI boom since 2010, Ziong said researchers have now developed deep learning and neural networks. The use of AI and ML in health care research has also grown, increasing dramatically since 2015.1
Advanced machine learning (AML) models are more sophisticated and have algorithms designed to handle significantly more complex data, Xiong said. Whereas other ML models use structured data that is neatly organized (i.e., tables in a database, insurance claims, pharmacy records), AML models can make sense of unstructured data (i.e., text information, audio, video, images). Deep learning is an example of AML models that utilizes deep neural networks that are more suitable for unstructured data. These neural networks mimic the human brain to learn from data and make predictions.1
“Just like the neurons connect our brain, these neural networks use multiple layers to connect and network and then, based on this structure, the neural networks can learn from data just like how we learn from experiences,” Xiong said.1
Xiong also reviewed some case studies highlighting potential uses of these tools. In one, by Hsu et al., investigators used temporal modeling to predict medication adherence in cardiovascular disease management. The models were found to outperform non-temporal models and long short-term memory (LSTM) models produced the best predictive performance. The authors concluded that deep temporal models can be used to integrate patient history and predict adherence and said the use of LSTM models was particularly advantageous.2
In the second case study, published in JAMA Network Open in 2022, investigators Sumner et al. used multiple proxy data sources to estimate weekly national opioid overdose deaths in near-real time. Signals from 5 overdose-related, proxy data sources including health, law enforcement, and online data from 2014 to 2019 were combined using a statistical model. The model was able to estimate national opioid overdose death rates with an approximate 1% error, highlighting the ability of these tools to provide a more timely understanding of overdose mortality trends.3
Finally, in the third case study, Alexander et al. used unsupervised ML to identify and analyze clinical subtypes of Alzheimer disease in care electronic health records (EHRs). Notably, each clustering approach used in the study produced different clusters, but 1 cluster appeared consistently, suggesting the presence of a distinct disease subtype. The authors concluded that it is essential to systematically evaluate different ML approaches to identify disease subtypes using EHRs.4
Despite these encouraging findings, Xiong did note some limitations, particularly with natural language processing models or large language models (i.e., ChatGPT, Llama, Gemini). Ambiguity or nuances in medical language or terminology could be a challenge, necessitating careful oversight and human interpretation. Privacy and security concerns with sensitive patient data is also a concern, particularly given recent data breaches. Finally, regulatory compliance will need to be figured out as these tools continue to advance.1
Regardless, AI and ML tools are sure to continue evolving and becoming more common in the pharmacy space. Identifying the best uses for these tools, whether in patient education, data collection, improving efficiencies, or other uses, could significantly improve the future of pharmacy.1
“AI and machine learning can help automate the medication filling process, improving accuracy and also saving time,” Xiong said. “The AI can help check and double-check errors in the medication orders, and they can help us identify errors and reduce errors so that patients are getting what they need and how they need it.”1
