I have written frequently about the concept of a learning health system. The University of Michigan (UM) is one of the epicenters of the movement, because Chuck Friedman, formerly of the Office of the National Coordinator, has established a Department of Learning Health Sciences on its campus.
When I saw a UM presentation description about a new concept called the “Knowledge Grid,” I contacted Dr. Friedman, who put me in touch with research analyst Allen Flynn, PharmD, who is leading the project as part of his PhD dissertation.
Flynn told me that there is currently a lot of excitement about analytic processes, machine learning, and methods for building predictive models. The analytic output from these activities is going to be quite helpful to guide clinical decisions. “The question is how is the world going to organize knowledge that is in that form,” he said. “We think of those analytic results as knowledge. What kind of library and librarianship are going to be needed, and what skills and platforms are required to manage all of those outputs well?”
The Knowledge Grid platform at UM is still in its infancy. The goal is to make knowledge about health more Findable, Accessible, Interoperable, and Reusable (“FAIR”) than it is today.
One aspect of the concept is the creation of a digital repository and associated metadata. “We are talking about new kinds of knowledge infrastructure that would include a repository, Flynn told me. Here is how he explained the interest in machine-interpretable, computable knowledge about health. “Today, we see huge growth in biomedical research output in written form —through journals indexed in PubMed, for example.” He describes the Knowledge Grid not as a replacement for that but as a complement to it.
“Think of a repository of computable health knowledge instead of PDFs and articles,” he said. This is not a new concept, he admitted. AHRQ has a project to create a clinical decision support repository. Other groups are doing knowledge management focused on computable knowledge. “Our focus is on creating the infrastructure that would allow many others to stand up repositories like this,” Flynn explained. “We are creating the platform that may help others manage this knowledge in their own context, and at the same time reserve some space in the computable biomedical knowledge world for open repositories that are akin to something like PubMed Central, where when research dollars come from the taxpayers, after a year of embargo of being in a journal, then you get that article available to the public.”
The Knowledge Grid project sits at the nexus between the medical school and the school of information at UM, he said. “We are drawing on decades of work done on digital repositories in the library space. There is a profound need for that kind of capability for computable health knowledge.”
Predictive analytics is one use case, but there are others, he said, including computable guidelines. There are hundreds of guideline-generating organizations, in which experts look over evidence and create practice guidelines. Clinical decision support has been predicated on those guidelines for a long time. “There are sophisticated groups thinking about the next generation of clinical guidelines and how to improve them,” Flynn said. “We are complementary to their work, because they are also thinking about fully computable forms. We are thinking about managing fully computable forms of knowledge that scale.”
Here are some concrete examples Flynn gave about how the Knowledge Grid could be used: One pilot involves the oncology care model and patient-reported outcomes. Providers are trying to collect information on symptoms and patient experience between visits and put in place early warning interventions when a pain profile or mental health profile changes.
“One of our projects is to engage experts doing this work. Once they collect patient-reported outcomes, how do they want the data to be analyzed and visualized and delivered into work flows within care.” The Knowledge Grid can put the analytic model and visual model into externalized, modularized objects in a library that can be deployed in one or more applications. “Those could be much more easily shared than they can be today,” he said.
Flynn said the concept includes support for clinical decision support (CDS) modules. “CDS is one end point of how you can apply computable knowledge, but not the only one,” he said. “You can have behavioral apps for patients or analytics that run in the background to do population health scans and screens. As important as CDS is, we have other end points in mind as well.”
Besides the repository, the second part of the Knowledge Grid is called the “Activator.” It allows you to take a knowledge object out of a digital library platform and bring it into your own IT environment behind your own firewall. You then have a copy of that object you can run locally or in the cloud and the Activator wraps it in an application programming interface (API) for you. “It gives you an API mechanism so your EHR or any other system can call and draw on that. Epic, Cerner and other EHR vendors are making it possible to draw on external APIs, he said.
Here is another example Flynn gave of the potential. In 2013 there was a journal paper published that described a risk prediction model for lung cancer. It included smoking history and family history of cancer. Using the model and the appropriate population, it is possible to predict who is at higher risk of lung cancer over a six-year period. “How do you get something from a journal paper into practice?” Flynn asked. “As long as there is a table with logistic regression models and variable coefficients sitting in a paper, that is not how it is going to be used in practice. People are not going to open that up that and get out a pencil and calculate it. It doesn’t happen. There is no time for that to happen. For actionable things like that, we can create a modular object that we can share.”
Yet another example UM is working on involves pharmacogenomics. Although more genetic tests are available, it is not happening on a widespread basis yet. “Our small part would be take guidelines that exist from a pharmacogenomics consortium and start to make knowledge objects available so they can be widely deployed and so that not every site has to re-engineer this knowledge in their EHR,” Flynn said. “The dream we have had of shareable clinical decision support never quite got there. Our hope is that as web services and RESTful APIs and the semantic web have developed around us, we now have the minimum set of standards and capabilities to actually move the needle.”
I asked Flynn how this all tied into the learning health system concept. “We think a learning health system has profound analytical capability and profound capability to bring knowledge to practice,” he said. We are working on the infrastructure for knowledge to practice, recognizing that the infrastructure for data to knowledge is already very well on its way. We believe this Knowledge Grid is needed to bring about a learning health system at scale.”