Mayo Clinic’s Healthy Model for AI Success

The nonprofit health system’s technology and services are helping people throughout the organization build AI applications efficiently and safely.

Reading Time: 7 min 


AI in Action

This column series looks at the biggest data and analytics challenges facing modern companies and dives deep into successful use cases that can help other organizations accelerate their AI progress.
More in this series

Alice Mollon / Ikon Images

In 2022, we argued in the MIT Sloan Management Review AI in Action series that Mayo Clinic was probably the most aggressive adopter of AI among U.S. health care providers. Today, it’s time to review some of the factors that have made this health system successful with AI and its underlying components. One key: Mayo Clinic staff members see the data and AI team as enablers, not gatekeepers.

We were surprised in 2022 by the extent of AI activity underway at Mayo Clinic, but that’s explainable in part by the organization’s size: It’s the largest nonprofit integrated health system in the world. The system employs 76,000 staff and 7,300 physicians at three campuses in Minnesota, Arizona, and Florida.

Moreover, with the organization’s long tradition of medical research, it makes sense that many of the organization’s clinicians and administrators have attempted to find ways to use AI to improve care. And as one of the most highly rated medical institutions in the world, Mayo Clinic would naturally want to improve its patient care and administrative processes with a revolutionary technology like AI.

Since we last wrote about Mayo Clinic, it has developed an infrastructure for enabling and facilitating AI development that has led to considerably increased activity (generative AI hasn’t hurt in that regard either). And since we are both interested in enablement (making it easier for users to do the right thing in building use cases with data, analytics, and AI) as opposed to the much more heavy-handed governance (“we professionals will tell you what you can do and what you can’t”), we were interested in revisiting an organization that has taken a strong enablement focus.

Current AI Use Cases

In speaking with Ajai Sehgal, Mayo Clinic’s chief data and analytics officer (CDAO), we learned about some new use cases — applications for specific tasks — that we weren’t aware of. On the clinical side, researchers have created an algorithm to identify certain heart pump problems (low ejection fraction, among others) from 12-lead echocardiogram (ECG) readings that were previously only detectable through stress tests. The AI algorithm can also be used to detect some heart diseases, including hypertrophic cardiomyopathy and cardiac amyloidosis. The algorithm was cleared by the Food and Drug Administration to be marketed as a medical device (by a Mayo Clinic spinoff called Anumana) and has already been modified to take Apple Watch single-lead ECG signals.

Mayo Clinic researchers have also created a new class of AI called hypothesis-driven AI that may help to improve interpretability of AI algorithms for health care treatments, particularly for cancer.

Sehgal says Mayo Clinic continues to work on administrative AI use cases. During the COVID pandemic, when capacity management was a major focus, a machine learning model was created to forecast the availability of beds in intensive care units. The same approach was later used to address capacity to treat RSV in the Children’s Center.

The Power of AI Enablement

We’ve been advocating for companies to shift their focus from data and AI governance to enablement, by emphasizing the technology and services that make building AI applications easier and safer. Mayo Clinic is ahead of that game, and Sehgal extensively employs the enablement concept. It’s the primary approach to helping clinicians and administrators develop their own AI capabilities — a large-scale “citizen development” effort.

AI is a “tool that needs to be put in the hands of the people with the deep knowledge in the practice,” Sehgal says. “If you want to leverage AI, the people with the domain knowledge need to be able to leverage the tools.” Centralizing the development of AI models is not scalable, he adds. Instead, Mayo Clinic leans in to its clinicians, who are plenty smart and quantitatively oriented, to make use of machine learning toolkits.

Sehgal’s group is primarily focused on AI, technology, and data enablement, and he has 60 or so people who help him fulfill it. His group, in partnership with Mayo Clinic IT, has built two versions of an AI Factory, a platform for building applications. Mayo Clinic has a partnership with Google, and the company’s Vertex AI suite is a primary component of the platform. The platform also includes a tool set for gathering the needed regulatory information if a proposed use case has to be approved by the FDA or other regulatory bodies.

Mayo Clinic helps clinicians develop their own AI capabilities — a large-scale “citizen development” effort.

On the services side, the AI enablement function provides consulting services, education, and even a medical AI degree program for Mayo Clinic staff. There is also a lawyer in the group to consult on regulatory issues. A key part of enablement consulting is to encourage people with a use case idea to come to Sehgal’s team and discuss strategy and regulatory considerations before developing it. The team will then provide guidance on tools, data, regulatory issues, and other related use cases within Mayo Clinic that may have already been developed.

Preparing Data to Be Used Most Effectively

Sehgal sometimes uses the term governance when discussing data, but he primarily employs a more enablement-oriented term — stewardship. Data stewardship is a concept that has been difficult to sell in many organizations, but Mayo Clinic appears to be making it work. Stewardship in this context means user groups are responsible for getting the data ready for analysis — integrating it, ensuring quality and currency, and putting it in the organization’s data library. Business and clinical stakeholders are the stewards of the data; they know it best and can best leverage it with AI.

To illustrate the value of the stewardship approach, Sehgal mentioned an email he’d recently received from the head of the surgery group at Mayo Clinic. It asked Sehgal to endorse a proposal for a self-funded data and analytics team for the surgery practice. Rather than naively starting with the expectation of creating an AI team, the surgeons are starting with getting their data ready. The head of the practice recognizes the role that data plays and is setting aside funding to steward his data.

If you approach implementing stewardship as a transformational change — and take the requisite care to educate (with an extensive data literacy program starting at the senior executive level), outline the value, and so forth — the business and clinical stakeholders get it, Sehgal says. The alternative to stewardship, he believes, is not the IT function undertaking those tasks: It’s no one performing the stewardship activities, leading to a set of data problems.

The CDAO’s organization does maintain some centralized capabilities, including a data library or catalog. It is highly enhanced with links to associated clinical rules and metadata. Sehgal describes it as a one-stop shop for everything known about a data element. When, for example, the CEO asked for a new dashboard on hospital operations, he also wanted to know where every data element came from. It was easy to address the questions because Sehgal’s team could just point him to the data library. Beyond the library, however, most of the data responsibilities are distributed throughout the organization.

Enabling Generative AI

Even though Mayo Clinic has been exploring generative AI technology since Sehgal arrived in 2020, the technology is still sufficiently new and experimental that clinicians and administrators who want to develop generative AI tools need a lot of help. This is one of the primary challenges faced by the CDAO organization: The 60-person group of enablers isn’t large enough to meet the heavy demand for help with generative AI. The primary focus for GenAI use cases thus far is on automating administrative tasks — filling out forms, creating clinical notes, and the like.

As in many other organizations these days, the excitement about generative AI makes it much easier to sell the importance of data and data stewardship. In addition, Mayo Clinic is heavily focused on the outputs of GenAI models. As Sehgal puts it, “You need high-quality outputs in health care or you are in big trouble. Even using GenAI to summarize emails can get you into trouble.”

Since it has so many smart, autonomous, and innovative clinicians, Mayo Clinic is perhaps the perfect organization for an enablement-oriented approach to data, analytics, and AI. A more restrictive and punitive governance-based approach would not likely succeed given the user base at this organization. With all this enablement help and the success of the stewardship program, Sehgal and other senior executives at Mayo Clinic can be confident that data usage and analytical and AI models are both effective and safe.


AI in Action

This column series looks at the biggest data and analytics challenges facing modern companies and dives deep into successful use cases that can help other organizations accelerate their AI progress.
More in this series

More Like This

Add a comment

You must to post a comment.

First time here? Sign up for a free account: Comment on articles and get access to many more articles.