How 3 health systems decide when to buy or build AI

As health systems invest in artificial intelligence, executives are deciding when they should buy a vendor's Al product and when they should build their own models.

Health systems plan to increase spending on Al.They're figuring out when it makes sense to bring in engineers to develop in-house models and when they should adopt vendor solutions. While many systems will both buy and build Al, one factor driving health systems towards internal development is bypassing what they see as a crowded vendor marketplace.

‍

"There is an overwhelming amount of vendors," said Dr. Eve Cunningham, chief of virtual care and digital health at Renton, Washington-based Providence. "The problem is we see a lot of ideas but not a lot of proof that they've been able to get people to use [the technology)."

Developing in-house is not without its challenges. Finding and keeping the right Al talent and infrastructure to develop models internally is hard and costly, said Erik Pupo, commercial health information technology advisory director at consulting firm Guidehouse.

"Al requires more of a data science experience, which is very expensive in the market," Pupo said. "It also requires a lot of actual data and many hospitals do not have that or are able to afford access to large amounts of data." Better than you'd think Here is how three health systems are weighing their options.

‍

Providence: Too big to buy many AI solutions

Providence owns 51 hospitals across seven states. Its sheer size means a lot of small, narrowly-focused Al vendors aren't the right fit for the health system, Cunningham said. The company is looking to adopt Al solutions that can address large-scale problems across its various regions while also getting widespread buy-in from clinicians.

The nonprofit Catholic health system's Al governance structure evaluates if the technology addresses a strategic priority, can integrate into its electronic health record system and is built on ethical, unbiased datasets. The process is lengthy and not easy to navigate for a smaller company, Cunningham said.

"Any startup or company that works with us will need to have the right resourcing to be able to handle the size, scope and magnitude, and complexity of what we have inside of our system," Cunningham said. "But if you solve it in Providence, you can probably solve it for healthcare."

The system has developed its own Al-enabled tools including MedPearl, an EHR-integrated digital assistant and clinical knowledge platform for primary and urgent care clinicians. MedPearl, which provides subspecialty guidance on more than 700 conditions, is used by nearly 7,000 Providence clinicians, according to research the health system published last Wednesday in the journal NEJM Catalyst.

Providence has developed a home-grown chatbot that answers patient questions, as well, Cunningham said. The system also works with large vendors including Microsoft's Nuance for its Al scribe solution, which takes recordings of patient-doctor conversations and adds them to the EHR. If Providence adopts Al for imaging analysis, it's also likely to use a vendor solution since that's a more mature market, she said.

‍

Ochsner Health searches for AI talent

Ochsner Health employed a data science team long before other health systems jumped on the trend, said chief digital officer Dr. Denise Basow. The New Orleans-based health system has a long history of either building its own models or customizing EHR vendor Epic's models for Ochsner's patient population, she said.

For example, Ochsner created an Al model that predicts who's going to deteriorate in the hospital and may need more intensive care services, Basow said.

Finding talent to develop and customize these models is challenging. The nonprofit health system is thinking differently about pay structures and growth opportunities for employees who work on Al models compared to other roles at the organizations.

‍