A Better Way to Motivate Our Patients? Generative AI and Adherence, Part I.
Prompting Our Way to Better Health
“Drugs don't work if people don't take them,” C. Everett Koop once said. The former surgeon general was being witty, but his point is serious: Patient adherence is a profound challenge for the US healthcare system. Inadequate adherence amounts to $500 billion in avoidable healthcare costs and causes 125,000 avoidable deaths each year in the US. Currently, clinicians are putting health apps, telemedicine and wearable technologies to work on the issue. But with the startling arrival of generative AI (GenAI) technology and large language models (LLMs), such as ChatGPT and Google Med-PaLM 2 (not public yet), new possibilities for patient-friendly interactions, with their context-relevant, human-like responses, are making the rounds.
We believe that by responsibly integrating LLMs, healthcare can empower patients in unprecedented ways, potentially leading to improved care outcomes and significant cost reduction. But with technology as new and ever-evolving as GenAI, leaders must consider what constitutes responsible AI while identifying its realistic capabilities and applications. This post aims to address these considerations.
Guardrails: Watchouts and What to Do
EPAM experts are looking at GenAI with a holistic view, considering reliability and safety, costs versus benefits and privacy laws. They say it’s essential to keep these ideas in mind when thinking about how GenAI in the context of patient adherence.
Andrea Sorkin, our Director of Data Analytics Consulting, cautions against relying on LLMs for medical diagnosis or advice until they achieve a high degree of confidence because GenAI's diagnostic current capabilities are not as reliable as those of human doctors. Liz Winter, Director of Innovation Consulting, notes the absence of data for certain patient groups, such as racial minorities and women, historically underrepresented in healthcare data.
Healthcare providers must first recognize that GenAI is a potential means to improve care outcomes. Jonathan Rioux, Director of Data Analytics Consulting, recommends asking “How can GenAI play a role in reaching the outcomes we want?” rather than “How can I integrate GenAI into my product?” Analyzing the benefits and costs of introducing a new technology is always prudent, but given the allure of GenAI, it’s particularly important. Not every challenge can be solved by GenAI.
Ira Livshits, Senior Data Scientist, stresses the importance of using caution while staying up to date on GenAI's limitations. As the field advances, “its boundaries are being pushed, and its capabilities are expanding, but at the moment, one has to be very, very careful.” Healthcare solution providers must prioritize compliance with privacy laws and patient safety when exploring GenAI integration.
It's important to note that LLMs currently have no access to Electronic Health Record (EHR) data due to regulatory constraints. Sorkin stresses the importance of protecting health and privacy data even as we promote “data interoperability.” A “data bifurcation” approach should share administrative data, such as provider availability, while safeguarding personal health information. The illustrations of these journeys are forward-looking, reflecting future value, and are unconstrained by current limitations.
The GenAI Health Coach Is Always on Call
Patient adherence requires continuous behavior modification, commitment and lifestyle adaptations. Livshits anticipates that the most significant contribution from GenAI will be its constant availability and responsiveness, provided it is “trained not to overstep or hallucinate.”
Sorkin suggests that, within certain boundaries, GenAI could serve as a helpful adherence coach. Regarding the significant challenge of an LLM generating incorrect information without a cited source, EPAM experts recommend using Retrieval-Augmented Generation (RAG), a method that limits an LLM to reference only reliable, authoritative sources of information. Documents from clinical providers or pharmaceutical companies could be fed to generate responses to patient queries.
Eugene, 34, recently diagnosed with ADHD, seeks to best understand the jargon he heard from his psychiatrist, such as “executive functioning” and “co-morbidity.” Using the GenAI tool recommended by his doctor, he gains a deeper understanding of how these terms relate to his situation and is enabled to ask follow-up questions.
After mistakenly forgetting to take his medication for two days, Eugene consults his tool for guidance rather than leave a question on his provider’s voicemail. Glad that he wouldn't be judged for his slip-up, he is pleased to receive a provider-verified answer in response to his question, as his case was in the vectorized database. Following his coach’s instructions, Eugene resumes taking his medication and the coach schedules a follow-up call with his doctor by accessing the doctor’s and his calendars.
Health Information Aggregator and Advocate
In an ideal world, healthcare-based GenAI would pull information from a multitude of sources. Naomi Korn Gold, Senior Director of Innovation Consulting, believes this capability could one day foster patient adherence by being “smarter at not just tracking but actively encouraging adherence.”
GenAI may be able to personalize and optimize the healthcare experience by acting as both an aggregator of relevant health data and an advocate for the patient's health management.
Riley, 42, manages a chronic condition through regular health measurements, exercise and medication per her provider’s recommendations.
A GenAI tool has learned Riley’s routine. She takes her meds and checks her blood pressure before leaving for work at 8 a.m. If she forgets by 7:45 a.m., the tool prompts her. The system adapts to her daily activities and respects her autonomy without imposing a rigid routine.
Riley also aims for 40 minutes of post-work walking and eight hours of standing time daily. On days when she has back-to-back meetings, the system reminds her to take laps around her neighborhood once she returns home (based on GPS data). It also analyzes weather forecasts and suggests she add extra walking minutes when it detects unpleasant weather in the next few days and recommends indoor exercise programs on rainy days.
Korn Gold believes that such a model could significantly reduce notification overload. Most importantly, this will give patients a sense of control and free up their cognitive bandwidth to live their lives without worrying about following health routines. By learning the patient's context, the system could operate as an unobtrusive yet effective advisor, providing flexible and collaborative support beyond preset alarms.
A GenAI for Koop?
Would C. Everett Koop have approved of GenAI? So much depends on how humanely we design it. GenAI might truly improve patient adherence, but its integration into patient care must be approached with caution, responsibility and a focus on user needs. Successful implementation must start with a diverse, multidisciplinary team capable of navigating the complexities inherent in such an operation. You’ll learn all about this in our next post. Stay tuned!
