Recognizing the potential and peril of generative AI

Much has been written about generative AI’s promise for healthcare for everything from back-office task automation to clinical documentation. At the same time, only 6% of health systems have a generative AI strategy in place, and executives remain unsure where, how, and when the technology will be as transformative as promised.  

Momentum is certainly building to put generative AI to use. In recent weeks, Microsoft and Epic Systems have announced a collaboration, as have Google and MEDITECH, with both emphasizing improving the productivity of clinicians using the electronic health record. Health systems around the country are also taking note: Cedars-Sinai is developing generative AI tools internally, Northwell Health is working with a startup incubator, and many are partnering with technology vendors. 

That said, two recent reports suggest further uncertainty about the technology’s impact. McKinsey described generative AI as “a meaningful new tool that can help unlock a piece of the unrealized $1 trillion of improvement potential.” On the other hand, Health Affairs warned of “a bleak outlook for interoperability and fairness” and an abundance of low-quality products. 

McKinsey pointed to a range of potential use cases. Payers, for example, could help members search for in-network physicians or compare plan and product features. Providers, meanwhile, could expand clinical decision support offerings, personalize patient education and clinical training, and automate coding. Both entities could synthesize notes, generate summaries, automate marketing, and augment self-service offerings. 

The challenges are significant, though. Healthcare must contend with data privacy, which is something open-source generative AI tools have struggled to protect. In addition, significant investment must be made in technology infrastructure as well as staff, especially as “[a]ny content synthesized or summarized by generative AI must have human-in-the-loop involvement and undergo rigorous risk and compliance review.” 

Health Affairs cited additional cautions. One centers on interoperability and product quality. The large language models on which generative AI is trained require large medical data sets. Healthcare providers are unlikely to share data with companies that intend to develop competitive products – and, in turn, will develop their own products. As a result, a “fragmented landscape” will emerge, with an abundance of products trained on proprietary datasets “and thus of lesser quality and lesser diagnostic acuity than would have been possible had all the data been available.” 

The other concern is the potential to exacerbate inequity. It would be all too easy for the industry to funnel patients with limited means to AI-based chatbots, Health Affairs argued, with in-person visits reserved for patients with extensive insurance coverage or the means to pay out of pocket. 

With these concerns in mind, organizations such as the World Economic Forum (WEF) and the American Health Information Management Association maintain cautious optimism about the potential for generative AI.  

AHIMA said generative AI is positioned to discover “unique opportunities” to improve population health, disease surveillance, and diversity in clinical trials. Predictive models are bound to get better with time – provided they have access to reliable, unbiased data that they can analyze in a secure manner.   

WEF, for its part, said “[m]oments like this don’t come around often,” citing benefits for global health that include pandemic preparedness, personalized care plans, and faster drug discovery. “[T]hose daring to experiment and lead in this space will help create opportunities for patients, providers and healthcare institutions alike.” 

 Brian Eastwood is a Boston-based writer with more than 10 years of experience covering healthcare IT and healthcare delivery. He also writes about enterprise IT, consumer technology, and corporate leadership.