As healthcare costs continue to rise, many Americans are looking to artificial intelligence to provide cost-reducing solutions. At the 13th annual UNC Business of Healthcare Conference, a panel of experts separated the AI hype from reality in a discussion of the limitations, risks and ethical questions surrounding AI solutions in healthcare.
AI describes when machines are able to learn from experience and simulate human intelligence processes. Often referring to computer systems, AI also encompasses mathematical models, which have been in use for more than 70 years, as well as natural language processing algorithms that have enabled tools such as ChatGPT. A common public perception posits that AI means computers eliminating jobs or robots taking over, yet in reality its usefulness lies in assisting humans with various tasks.
Amanda Barefoot, director of healthcare and life sciences strategic advisors at SAS, emphasized that AI is often overcomplicated, especially in the healthcare sector. Some professionals believe they need to adopt AI solutions immediately to address their challenges. Yet, she pointed out that many are already employing AI, even in limited scopes, possibly without full awareness. Marcella Dalrymple, associate director of strategic development and business partnerships at Duke Health, elaborated on this topic. She highlighted a common misconception regarding AI’s use in healthcare: the lack of understanding that AI is already integrated into routine medical practices, such as in the calculation of risk scores for chronic diseases like stroke or Alzheimer’s. This example indicates the broad, yet subtle, penetration of AI technologies in healthcare, which is more widespread than is commonly perceived.
Barefoot said that while clinical diagnostics is often the primary healthcare application that people attribute to AI, this area certainly does not represent the low-hanging fruit. Where AI can really shine is in taking on the repetitive, burdensome parts of a job, allowing more time for other important tasks that require deeper thinking and problem-solving skills. In healthcare, panelists suggested that using AI for operational and financial applications would have the greatest potential to yield cost savings.
Large language models, which are extensive deep learning models trained on vast data sets, excel in tasks such as creating summaries and identifying missing information. These LLMs, including the technology behind ChatGPT, have the potential to automate numerous tasks. For instance, they could significantly streamline the exhaustive paperwork process that physicians must complete for prior authorizations. Seth Lester, an actuary and business development consultant at Milliman Advanced Risk Adjusters, elaborated on this possibility. He suggested that an LLM could compile all the clinical information a healthcare provider has, organizing it in a way that automates the prior authorization process. “This would significantly reduce the administrative burden on healthcare providers,” he explained, “allowing them to focus more on patient care.”
Clinical trials are another promising area because AI could be used to quickly identify patients that meet criteria for certain clinical studies, a process that has traditionally required contacting individual physicians to see if they have patients who could be considered. AI can also be used to identify doctors who are qualified to act as principal investigators for clinical trials. Expediting this upfront work means clinical trials could launch earlier, allowing research results to benefit patients faster.
Dalrymple pointed out that accessibility is a critical limitation of AI in the healthcare arena. “A rural hospital or community center does not have the resources to develop these tools yet and to evaluate them effectively and to implement them into their computer systems, because they don’t have a thousand engineers,” she said. “Accessibility and equity, in that standpoint, is the first thing we have to address.”
The expense of developing AI tools limits accessibility and also means that careful thought must be given as to whether an AI solution is necessary and worth the investment. It can cost upwards of $1 million to develop a single AI model, and, because they do not always work correctly, not every model will make it to deployment. Although AI seems to be the popular choice, Lester said that sometimes a simple automation program will work while also coming with less risk.
Another consideration is deployment. Without a deployment strategy that integrates the AI tool into other systems and processes while also allowing results to be used in a way that guides actions, an AI solution is not likely to contribute significant benefits.
Panelists said that while bias is always a risk with AI, it can be avoided through careful testing and proper guardrails. “When AI is touching a patient and helping make a clinical decision, we want to make sure that it’s not going to exacerbate any type of healthcare disparities because the AI that’s already being implemented is usually trained on a very biased dataset,” Dalrymple said.
As AI becomes more common, standards and guidelines are increasingly important. At Duke Health, Dalrymple said that algorithms are not integrated until they pass an evaluation and quality framework in which ethicists, physicians, nurses, clinicians, Ph.D. researchers, and engineers conduct an exhaustive evaluation of the process by which the algorithms were derived. There are also industry-wide efforts to develop standards for AI technology use in healthcare. The Coalition for Health AI is a community of academic health systems, organizations, and expert practitioners of AI and data science working to provide guidelines and standards tied to health AI.
The panelists acknowledged that while there is a lot of excitement surrounding AI in healthcare, its most practical and impactful application would be in streamlining everyday areas like operations. They suggested that implementing AI in these essential but often overlooked segments could lead to significant efficiency gains and cost savings.