Healthcare organizations are seeking ways to prepare for a data-driven future utilizing artificial intelligence (AI) and machine learning (ML) that improves the health of individual patients and entire communities. Lessons can be learned from the work being done at Mayo Clinic, a health system that serves patients who often require complex care. Well before the COVID-19 pandemic hit, the organization was investing in better ways to collect, manage and analyze complex healthcare data.
The health system defined a simple but daunting goal: figure out how to use data in more powerful and meaningful ways. But as they began to build their internal infrastructure to house data in more sophisticated ways for both research and clinical applications, they soon discovered serious limitations. This journey was impossible to tackle alone.
To start from the ground up, they needed a partner in order to fully utilize AI and ML. That need was the driving force that launched their partnership with Google Cloud. What followed has been nearly a year of challenges faced, lessons learned and hopes ignited by the race to integrate smart technologies into research processes and clinical workflows.
For Mayo Clinic, choosing the right partner was the first critical decision. They wanted an established cloud partner to house data in a secure and private way and one that could collaborate on research and innovative ideas.
Cris Ross, chief information officer at Mayo Clinic and a HIMSS board member, said, “We manage complex cases at Mayo Clinic, with many patients who have vast medical records. So a lot of people have said to me over the years, ‘Wouldn't it be fantastic if you could use the power of the search and other tools in order to get a deeper and more meaningful insight into patient care?’ So, that was the root of what we were trying to do when we came together with Google. They have invested from the ground up to create a [cloud platform] and we’re using that environment to support patient care, research and investigation of new cures.”
In launching their partnership, Ross and team worked closely with Aashima Gupta, director of global health solutions for Google Cloud and a HIMSS North America board member, and her team. Together, the two teams gave behind-the-scenes insights on the keys to success in priming a healthcare organization to utilize AI and ML on the journey to advance healthcare.
Gupta used the metaphor of a three-layer cake for the partnership. The bottom layer is made up of the foundational elements needed to accelerate joint solutions. It represents the data and infrastructure fabric.
“The primary goal for this layer is to make data more meaningful, accessible and open,” said Gupta. “With healthcare data, we are dealing with a high degree of complexity. The data sets are often too large, they're too complex, and different data ontologies are so dissimilar that they cannot be combined easily. So another goal is to make it easy for machine learning developers to take their projects from ideation to deployment quickly. And finally, I am passionate about data interoperability to enable digital experiences.” Gupta sees all these as essential to create efficiency in both operational and clinical workflows.
The top layer—or frosting on the cake—came when Mayo Clinic and Google collaborated to jointly develop solutions to transform healthcare. But what’s in the middle matters, too. And that where the detail lives.
According to Gupta, the center of the cake is critical to the recipe for success. It consists of foundational work that must be done by a healthcare organization to move forward with utilizing AI and ML. This works consists of:
As they were preparing, the organizations needed to break down the journey into three major stages, shared Ilia Tulchinsky, engineering director for Google Cloud Healthcare & Life Sciences.
The first step to prime the organization for AI and ML was to integrate their clinical, operational and financial data together from multiple sources for a holistic view of the patient journey. “It's important to be able to connect diverse systems of records that leverage different protocols, APIs, or data formats and to be able to ingest that into the cloud,” said Tulchinsky.
Once in the cloud, the data was available and could be applied to many problem types. “This allows us to be agile and have this ‘build once, use many’ approach, as this foundational piece lends itself to many, many different problems and analysis,” Tulchinsky added.
Other key components to this early phase were implementing strong cybersecurity measures for security and data privacy, and access controls to meet privacy and regulatory requirements. “We really were committed to getting this foundation right because it serves such an important component of building upon,” shared James Buntrock, the vice chair of IT at Mayo Clinic.
Once the data was in the cloud, because it came from many sources and in many formats, the next phase was to do a quality check on the data and map it to common industry standard schemas, like FHIR.
“We also need to remember where the data came from,” added Tulchinsky. “So, provenance and image are very important because when it comes time for analysis or ML building, it really matters if the data came from a human or devices—what was the journey of the data up to [this] point.” He added that “unstructured documents, whether clinical records or reports or other types, are treasure troves of information, and so extracting structured knowledge from them at this stage is very important for subsequent analytics and AI.”
Tulchinsky also stressed the importance of developing common ontology and terminology, as well as units of measure during this phase.
With integrated, clean data the final stage was to analyze and model their data for actual implementation of AI and ML.
Mayo Clinic has several collaborative clinical examples active today, including breast cancer risk prediction based on the Tyrer-Cuzick algorithm risk assessment. The algorithm draws on a number of data elements that already exist in the EHR.
Another great example Buntrock shared is in their radiology department around polycystic kidney disease. One way to understand the disease change is to look at changes in the total kidney volume over time. While humans can make these calculations, it’s strenuous. Instead, the radiology group provided a set of annotations and markups, and trained models to better calculate the change in volume. They then use this as an overlay annotation on imaging that can be used specifically for this disease, and can be applied to a number of other clinical situations as well.
“This is actually work that's in production. It's running today,” shared Buntrock. “A number of other things are in the pipeline that bring new insight into disease progression, new insight into prognostics, as well as insight into perhaps new types of therapy applications.”
Integrating and then utilizing AI and ML in the clinical workflow is the critical endgame for any healthcare organization. Buntrock said, “We can have the greatest model and insight, but our ability to take that into the ‘last mile’ problem of applying it to the case of a patient setting the clinical workflow, of asking when does that model need to execute, and when does that decision support have to be rendered, are all important elements. A number of our clinical areas have stepped up to recognize this.”
In their partnership to better utilize AI and ML, Mayo Clinic and Google agree on essential lessons learned thus far in a shared venture. First, that technology is here to stay and has a role in keeping people healthy. Second, that for any transformation to occur, there must be agility, trust and alignment with the culture of a particular healthcare organization. And finally, that implementation is an ongoing journey that requires investment, even when new urgencies like COVID-19 arise.
Gupta summarized by saying, “In the end, it is about creating better AI-enabled tools to move clinical research, to enhance clinical and operational processes and to take the patients along with us on that journey.”