Pharmacy Today logo

Researchers wield AI to address some of pharmacy’s most serious problems
Roger Selvage 1750

Researchers wield AI to address some of pharmacy’s most serious problems

Previous Article Previous Article Pharmacist-provided care may significantly help Hispanic patients lower A1C
Next Article Lipoprotein(a): A crucial influence on atherosclerotic risk Lipoprotein(a): A crucial influence on atherosclerotic risk

On The Cover

Sonya Collins

Illustration of pharmacist and patient immersed in a digital pharmaceutical environment.

Artificial intelligence (AI) may have the capacity to extend its reach into virtually every industry and profession, and pharmacy is no exception. While professionals in many industries may envision ways in which the technology could make their jobs obsolete, there might be more reason for excitement than dread about the potential role for AI in pharmacy practice.

“As AI grows, pharmacists have a lot of expertise to help develop these models to make them more effective, especially given our expertise in medications,” said Adrian Wong, PharmD, MPH, FCCP, a clinical pharmacy specialist in the medical intensive care unit at Beth Israel Deaconess Medical Center in Boston, MA. “Involving pharmacists from the beginning to help develop these models will be very helpful. AI won’t necessarily take away our jobs. It will just improve patient care.”

While most uses of AI in pharmacy practice remain theoretical, research underway examines AI as a tool to address some of the most serious issues affecting pharmacy practice and the patients in pharmacists’ care.

AI to quantify the value of pharmacists’ care

At a time when many hospitals are reporting shortages of clinical pharmacists in both the acute and ambulatory settings, the OPTIM study is engaging AI to quantify the effects of such a shortage.

The study, in which 20,000 patients are currently enrolled and which has a goal of 30,000 enrollees, will use traditional methods and unsupervised machine learning to analyze the relationship between patient mortality in the ICU, pharmacists’ patient load, and many other components of ICU care. Unsupervised machine learning identifies distinctive clusters within a dataset, and researchers may then determine whether the clusters are meaningful.

“My hypothesis is that we’re going to find a cluster of ICU patients who had the most ICU resources—the most physicians, pharmacists, nurse practitioners—are going to have the best outcomes,” said Andrea Sikora, PharmD, MSCR, clinical associate professor of clinical and administrative pharmacy at University of Georgia College of Pharmacy in Augusta, GA, and the study’s principal investigator. “The ability to figure out which resources, or which patterns, are associated with the best outcome would be an impossible study to do otherwise.”

Sikora, along with her coinvestigators, expects OPTIM to be a landmark study for the pharmacy profession. “It’s going to show for the first time that pharmacists’ workload affects patient outcomes. If you overload your clinical pharmacists, they don’t do as good of a job taking care of patients, and patients have worse outcomes.”

Better allocation of pharmacists’ time and expertise

In the face of an aging population and shortages across most all health professions, pharmacists practicing at the top of their licenses could fulfill many unmet health care needs. The COVID-19 pandemic demonstrated this. In fact, some state scope-of-practice laws would allow pharmacists to diagnose, prescribe, and treat many common health conditions. But basic pharmacy operations—many of which do not require a pharmacist’s high level of training for execution—often keep pharmacists from providing more direct patient care. AI may offer solutions here as well.

AI platforms are already in use in some pharmacies to help predict times of day that get the most patient traffic or see the longest patient wait times so that pharmacies can plan, schedule, and allocate staff accordingly. Researchers in Taiwan taught machines to identify blister-packaged medications with greater than 90% accuracy, which freed up pharmacists for more clinical tasks.

AI’s predictive abilities may help with drug shortages by heading them off at the pass, as well. Researchers at the University of North Carolina reported in the American Journal of Health-System Pharmacy that their machine learning model could predict drug shortages based on drug characteristics and manufacturer-related variables.

Planning ahead for drug shortages can save time and resources later.

AI and patient safety

Medical errors, including medication-related ones, are a leading cause of death. Since the advent of EHRs, medical errors have plummeted, but the digital medical records haven’t eliminated the problem. Typically, pharmacists have had to check prescriptions against patient records for possible errors, and the most common triggers of these checks have been data about the patient or about the drug. AI may provide a means of reducing a different type of medication error substantially.

Researchers at New York University found that certain prescriber behaviors may predict risk for medication errors, too. When they used both machine learning and traditional methods to analyze the relationship between prescriber behavior and risk of error, machine learning outperformed other approaches. It found that multiple provider-specific factors, including prescriber experience and number of patient interactions in the hour before placing the medication order, predicted risk for errors.

In another study aimed at preventing medication errors, machines learned to link prescriptions for certain medications with diagnoses of corresponding GI disorders in the patients’ medical records. Prescriptions that were ordered outside of diagnoses of these conditions were flagged for pharmacist intervention.

AI has also been used to mine social media for reports of adverse drug events, though this area needs more work, as humans must screen the data first for social
media–specific phenomena, such as emojis, that machines cannot interpret.

Findings like these could power more effective clinical decision support tools that might flag potential errors for prescribers at the point of order entry and for pharmacists when they fill the prescription. The tool could streamline pharmacy operations by more accurately identifying prescriptions that require pharmacist intervention and save pharmacists from reviewing orders that are less likely to be erroneous.

AI approaches to more precise precision medicine

Increasing recognition that medications are not one-size-fits-all drives clinicians and scientists to search for tailored approaches that will have the best outcome for each individual patient based on their genes, comorbidities, lifestyle, environment, and disease state. AI may help hone these approaches.

For example, while canagliflozin has improved outcomes for many patients with T2D, a small subset of patients who take this drug face increased risk of lower-extremity amputation. Researchers at the University of Florida College of Pharmacy have used machine learning to determine who these patients might be. Their study found that those with a previous lower-extremity amputation and those on loop diuretics were at greater risk for this outcome.

“We can use the drug to protect the heart and kidney in most of the diabetes population but avoid using it in those for whom it could cause worse or life-threatening outcomes,” said Serena Jingchuan Guo, MD, PhD, an assistant professor in the Department of Pharmaceutical Outcomes and Policy at the University of Florida College of Pharmacy in Gainesville.

In another study, Guo and colleagues used AI to home in on the reported protective effects of metformin against Alzheimer disease. The machine learning platform found that patients who didn’t have neuropsychiatric disorders and those who didn’t have long-term use of NSAIDs seemed most likely to receive the protective benefit of metformin.

“AI/machine learning helped us not only identify these subgroups, but also gave us more consistent results,” Guo said. “Previous studies about this benefit of metformin had given weak signals or been inconsistent.”

Similarly, researchers are using AI to improve warfarin dosing, which, Wong said, “is quite variable depending on the patient, and available guidance only explains approximately 60% of the dose variation by patient.”

Wong’s recent narrative review in the November 2023 issue of the Journal of the American College of Clinical Pharmacy, which described the role of AI in multiple facets of pharmacy practice, highlighted two papers that used AI to tease out the remaining variables in warfarin dosing. Both successfully employed various machine learning techniques to better predict warfarin dose in Black patients.

“This is an important area for AI research,” Wong said. “These populations are not well represented in these models, so we don’t have models that we can generalize to them.”

An AI response to a public health crisis

Provisional data from CDC show that deaths due to opioid overdose topped 100,000 again in 2022, a projected slight increase over 2021. Researchers have begun to use AI to tackle this problem from multiple angles.

A study published in the International Journal of Medical Informatics used AI to screen EHRs for ICD-9 codes as well as natural language that might indicate risk of problem opioid use. Ten percent of the patients identified by the machine learning platform had been previously missed by more traditional means of flagging risk. In another study, machine learning identified some 50 factors, including concurrent medication use and laboratory data, that helped identify overdose risk with greater than 80% accuracy.

Researchers at the University of Florida College of Pharmacy have analyzed AI’s potential to predict overdose risk among those receiving an opioid prescription. While the model proved highly accurate at quantifying overdose risk in many cases, the false negatives illustrated a common problem with AI.

“Sometimes AI/machine learning creates bias and can exacerbate existing health disparities,” said Guo.

In this case, the false negatives—that is, those who were at risk of overdose but not captured by the AI model—were more likely to be Black patients than white patients. In fact, Black patients were twice as likely as their white counterparts to be classed as having low risk. One factor that raised a patient’s risk for overdose was previous interaction with the health care system for overdose.

“We used claims data and most of the people who had sought medical attention for opioid overdose were white,” Guo explained. “For many underlying reasons, Black people were less likely to pursue medical attention for opioid overdose.”

Remember who’s in control

The bias that AI introduced in the overdose risk study serves as an important reminder for pharmacists and any clinician who might lean on AI for support in clinical or other decisions.

“Before you do anything with the data, you need to understand the model that created it,” said Almut Winterstein, RPh, PhD, distinguished professor and director for the Center for Drug Evaluation and Safety at the University of Florida College of Pharmacy.

She cites numerous examples of variables that AI might identify a causal relationship between but that human researchers would not. AI could suggest, for example, that people with diabetes have more depression than others, but that could be because they are more frequently seen by doctors and therefore more likely to be evaluated for depression than others. AI might suggest that if a patient stops statins, they are very likely to die in the near term, but a human clinical researcher would recognize that chronic therapies are typically discontinued close to death.

“These are the sorts of things we must be aware of,” Winterstein said. “This is where the design portion and the human brain piece comes in because the machine just sees the association and runs with it. AI is a tool and depends on how we use it.”

For these and other reasons, many potential clinical uses of AI are not yet ready for widspread adoption. AI models contain bias. Many populations are underrepresented in AI models. And there’s a lack of thorough validation of AI technology. But as it gradually makes its way into the clinic, it’s important for pharmacists to educate themselves on how the data were generated and the limitations of therein because it could eventually play a significant role in patient care and outcomes.

AI massively improving all of clinical pharmacy is still quite a ways away, Sikora said. “But I imagine it’s eventually going to be providing clinical decision support that’s just much smarter than what we have now.” ■



Documents to download