Providence St. Joseph finds success leveraging AI and machine learning projects
Providence St. Joseph Health is already proving that artificial intelligence and machine learning technologies can have a meaningful impact on the delivery of healthcare today.
The Renton, Washington-based health system, which operates 51 hospitals across Alaska, Washington, Oregon, California, New Mexico, Montana and Texas, has created a variety of systems built on AI and machine learning.
For instance, its No Show technology is already generating a monthly return on investment through increased patient arrivals. Its Medicaid risk stratification model is being used by its care management teams to find cases. And its natural language processing systems around spinal fusion and brain tumor surgeries have enabled new use cases for and enterprise-wide tool built to identify practice styles that optimize outcomes and cost.
When it comes to using AI and machine learning technologies and reaping rewards, health systems have an advantage over other industry players with privileged access to troves of patient information, said Vijay Venkatesan, chief data officer at Providence St. Joseph Health.
"The most significant benefit that health systems have with access to their patients' data is the ability to tell the patient story," he explained. "It provides an integrated longitudinal view of the patient across touchpoints of the care delivery continuum. Additionally, it allows for population-based analytics, which can be used for evaluating and optimizing care and the methods by which it is delivered."
Health systems have an even more valuable perspective on the data: It can also be viewed as the emergent behavior of a complex system, revealing how the organization truly functions, he added.
"In this data source, the variation we naturally see in practice, both from site to site, and provider to provider, offers hints as to which arrangements, sequences and combinations are associated with lower costs and better outcomes," said Venkatesan.
"Further, having a working model of the delivery system can enable educated guesses about how big, new changes might affect staffing levels, capacity, etc. This data is system-specific and most valuable when coupled with local leader insight, so it's much more difficult for a vendor to box up and export to a new system."
Health systems also can leverage data from an engaged digital relationship with patients. Providence St. Joseph is delivering new technologies to digitally engage consumers between episodes of care and make receiving that care more convenient.
The health system has launched 33 Express Care clinics, in which patients can schedule same-day appointments with one-click scheduling, initiate a video visit, or summon a care provider to their home. These digital interactions can provide valuable information about how, when and why patients seek care, and how the health system can digitally navigate them to the lowest-cost care setting for their condition.
The health system also has launched Circle, a women's health personalization platform that serves up content, products and services to engage consumers about their health and their family's health between episodes of care.
And the health system incubated and spun out Xealth, a platform that allows a physician to prescribe from their EHR any digital content, app, product or service just as they would a pharmaceutical. The data about the content a patient reviews, their use of prescribed content, engagement with apps and so forth can help the health system better tailor care over time and can be helpful in providing insights into population health management.
"Even more exciting is the potential for harnessing genomic, proteomic and biometric data," said Aaron Martin, chief digital officer. "The Institute for Systems Biology, a Providence St. Joseph Health affiliate led by Lee Hood, MD, spun out Arivale, which partners with health systems in 'scientific wellness,' to help patients improve their health using data."
One challenge health systems and other healthcare organizations face when it comes to AI is extracting useful information from unstructured text in EHRs. Health systems have all this data, which is an advantage. So what is the challenge?
"Text data represents a rich additional source of patient information, including quite a few things that might never appear in structured data," said Tristan Markwell, principal strategic scientist. "For example, social determinants information, OTC medication, or side effects/adverse reactions. There are several related problems to extracting this information. The central issue is entity recognition, or knowing what concepts are being discussed in the note."
This is made more complicated by language ambiguity and further compounded by abbreviations or misspellings by the author, he further explained.
"Once this challenge is overcome, the correct relationship of the named entity to the patient or encounter needs to be determined: negation might indicate that the entity is being ruled out or denied; the mention may be to a personal history, family history or suspicion; or the entity may be proposed or contingent on some future event," Markwell said. "Further, the information needs to be synthesized with the structural data, which may be contradictory or difficult to reconcile."
The most straightforward options at the moment for dealing with this text data are: Developing one's own natural language processing capability, usually through hiring data scientists with a specific computational linguistics background; working with domain-specific vendors to bring NLP capabilities to one project at a time; or working with a general NLP vendor to access this capability in a way that spans any individual subdomain, he explained.
"The first approach gives the most control and ownership, but is the riskiest and probably most expensive; the second unlocks the benefits on specific projects, but constrains your ROI to the identified use-cases; and the third gives a general capability but will likely require serious work to integrate into an existing environment in a helpful way," he said.
"It seems like the NLP-as-a-service model is on the cusp of being the most reasonable solution – the basic problem of extracting entities with metadata from notes is pretty uniform nationwide and thus subject to economies of scale, vendors seem interested in offering lightweight distribution models, and it meshes well with the movement of this data to a cloud."
Most health systems are hiring a core of data scientists and partnering aggressively with large cloud AI/machine learning vendors (Microsoft, Google, Amazon, etc.) to leverage their scale of expertise and tools, he added.
Providence St. Joseph Health is leveraging its vast data repositories, yet it has just scratched the surface of the data available to it.
"When predicting events for patients, we look not only at the diagnoses and medications, but also prior behavior and temporal patterns," Markwell explained. "For text data, specifically, we've analyzed the sequence of words to form a mathematical semantic model, and then used that to improve note searching; we've also taken surgical summaries and extracted key items, such as the approach, that are unavailable in the structured data."
The keys to getting at this information have been a willingness to invest in hiring people with the right skills, starting with simple tools and investing time in complexity only when needed, and protecting time for research and development and basic pipeline development, he added.
A good example of how artificial intelligence and machine learning are running through Providence St. Joseph Health today is its No Show mobile application. It's meant as a comprehensive fix for a common problem: No-shows occur every month due to patients skipping appointments or canceling too late for the health system to refill the slot. That presents a significant system-wide challenge.
"Our first step in addressing this problem was to build a comprehensive model that we were confident could be used to intervene," Venkatesan explained. "Our second step was to get these numbers in front of clinics in an experimental setting and demonstrate that using the model to intervene – generally a direct patient reminder call – could meaningfully reduce no-shows, making the clinic more efficient."
The data teams then integrated the algorithm into a web application for clinics that offers a number of useful features and also collects the call information so the teams can monitor utilization and ROI. The teams are now working to create a feedback loop, using the intervention data, to refine the model and drive interventions to where they will be most effective.
"We're also launching marketing pilots which use machine learning to build propensity models for patients," Martin added. "Machine learning helps us identify segments of population that already have a propensity toward the service we can offer them to improve their health and those who have the ability to act on it. These become the input into a marketing automation program which then helps the patients through their journey with us."
Martin further noted that the health system is experimenting with AI and machine learning in its main digital interfaces to help navigate patients to the right site of care, and that it also is looking at voicebot-driven AI and machine learning tools to simplify charting and navigation within a clinic environment.
Big Data & Healthcare Analytics Forum
The San Francisco forum to focus on utilizing data to make a real impact on costs and care June 13-14.