Skip to main content

Advertisement

Advertisement

Advertisement

ADVERTISEMENT

Conference Coverage

Using Artificial Intelligence to Understand and Improve Medication Nonadherence

Hannah Musick

In 2018, an information technology (IT) employee presented an idea to Hy-Vee’s Health and Wellness Division: What if artificial intelligence could be leveraged to improve their patients’ pharmaceutical outcomes? 

This idea sparked a new chapter in Hy-Vee’s health care services, according to speakers at Asembia’s AX23 Summit. 

“Underneath Hy-Vee’s Health and Wellness Division, we have a vast portfolio of offerings that are interwoven to help our patients live their best life,” said Kristin Williams, PharmD, executive vice president, chief health officer, Hy-Vee.

Dr Williams shared an overview of Hy-Vee’s many health care-related services for patients and employees across the United States and Puerto Rico, including Amber Specialty Pharmacy. Amber Specialty Pharmacy was the first division in Hy-Vee to test the use of artificial intelligence (AI) in what Dr Williams called the “machine learning project.” 

“Artificial intelligence is the theory and development of computer systems to perform tasks that normally require human intelligence,” said Kelli Wyant, PharmD, senior vice president of operations, Amber Specialty Pharmacy. “Machine learning uses data to analyze data. Siri, Alexa, and Google Assistant are some of the most obvious examples of artificial intelligence in our everyday lives.”

In health care, AI is used for robotic surgeries, virtual nursing assistants, precision medicine, and imaging analysis, Dr Wyant said. For Hy-Vee, the decision to integrate AI was driven by machine learning’s ability to rapidly analyze large volumes of data, efficiently act based on the data, work as a scalable model, and accurately interpret trends in patient behavior. 

Integrating machine learning also presented an opportunity to achieve closer alignment with payers and manufacturers, identify hidden challenges driving lapses in therapy, and understand previously unknown barriers to therapy, speakers said. 

“As the programs are exposed to new data over time, they're continually independently adapting to that data,” Dr Wyant said regarding how machine learning can predict patient adherence. “It’s an opportunity for pharmacists to identify possible interventions and resulting outcomes in improving patient care.” 

The speakers said a discussion about AI would be incomplete without acknowledging ChatGPT, an internationally recognized example of AI technology. 

“Chatbots are software that makes humanlike conversations with users via chat. While chatbots also use machine learning to digest data, the difference between bots and our programming is in the datasets,” Dr Wyant said. 

Hy-Vee’s machine learning program differed from ChatGPT because it is trained with patient data and does not have interface or interaction capabilities. The random-forest model used by Hy-Vee is trained using data on past patient adherence, drug costs, prescription details, location, patient proximity to their provider, age, sex, and more. 

To reduce variance, the random-forest model averages multiple deep decision trees trained on different datasets. Thresholds are used to label patients with nonadherence and can be adjusted over time based on changes in therapy decision-making and doctor feedback. The model was set to run weekly, and the results were stored in a data warehouse for future analysis.  

The pilot phase for the machine learning program, which began in September 2018, focused on gathering data from patients with cancer via phone calls. In comparing a control group to an experimental group of patients at risk for nonadherence, the model was successful in identifying high-risk patients. 

During the pilot phase, the Hy-Vee team also found that patients in smaller population cities were 5.12% more adherent than in larger cities, and patients located 100 miles or less from their provider office were 5.26% more adherent than patients living farther out. 

“To make the project more scalable and have success long term, we needed better integration between teams, and to create a process that wasn't as manual as phase 1. We also identified the need for more robust dose reminders via our apps,” Dr Wyant said. 

Today, the Hy-Vee team has built upon the foundation from 5 years ago to process data from patients with cancer, tardive dyskinesia, and transplants, speakers said. The goals of phase 2 include developing a more efficient process that fits seamlessly into clinical workflows and expanding patient populations to capture and report intervention data. 

Patients with cancer were identified as being uniquely motivated to be adherent and therefore fitting candidates for the program, as the side effects of many oncology medications may lead to dose interruptions or therapy discontinuation if not addressed. Patients with tardive dyskinesia caused by long-term use of antipsychotic medications may have adherence issues because of underlying conditions and treatment side effects. 

Because of their often-lifelong medical journeys and complex medication regimens, patients with transplants may not immediately see the effect of nonadherence. Dr Wyant also said transplant facilities have a vested interest in the success of patients from a business perspective. 

Connie Kerkhove, PharmD, clinical pharmacist, Amber Specialty Pharmacy, shared that the first challenge of phase 2 was to build a platform in-house which was customized for the specific needs of their patients and business. 

“Cascade, which is what we're calling our clinical platform, has given us the opportunity to document, track, and report intervention assessments without the pharmacist having to manually create the assessment and toggle back and forth,” Dr Kerkhove said. “Cascade has finally allowed us to all speak the same language, to speak across departments and build procedures together.” 

The Cascade workflow still relies on a human component to check for data accuracy. Once per week, the machine learning model outputs high-risk patients. The IT team generates an assessment in Cascade for the clinical team to review. Clinical staff, usually a nurse or a pharmacy intern, use these assessments to check if patients qualify for additional assistance because the model is accurate roughly 80% of the time. 

“Follow-up assessment cadence will depend on the conversation that's had between the clinician and the patient. If no short-term follow-up is needed, patients are recycled through the system and then can be assessed by the machine no sooner than every 6 months, so they're not falling back onto the next weekly report,” Dr Kerkhove said.

Since March 2023, the phase 2 system has led to 118 interventions completed with 250 patients. Thus far, the process seems to be more helpful for engaging internal teams while increasing the number of recorded pharmacist interventions, speakers said. 

“Cascade has finally allowed us to all speak the same language, to speak across departments and build procedures together,” Dr Kerkhove said. “The number one barrier for us is still reaching the patient when they don't want to be reached.” 

In the future, the speakers shared that they hope to expand to other therapies or condition categories, partner with manufacturers and payers, find potential external point integrations, and investigate further ways to integrate their app with the portal. 

Advertisement

Advertisement

Advertisement