By Deborah Borfitz
October 9, 2018 | If platforms built on artificial intelligence (AI) are to be useful and scalable for healthcare settings that are typically chaotic and strapped for IT talent, there are a few prerequisites. Solve a problem. Be easy to implement and minimally disruptive. And offer a tangible return on investment—ideally, help organizations meet their “quadruple aim” of enhancing the patient experience, improving population health, reducing costs and improving the work life of clinicians and staff.
That is precisely the approach taken by Digital Reasoning to make its foray into healthcare with one of the nation’s leading providers—HCA Healthcare, which three years ago invested $5 million in the machine learning pioneer. The cognitive computing firm knew the right questions to ask of the data because it asked the experts—HCA clinicians and not just its data scientists—says Paul Alexander Clark, director of healthcare research at Digital Reasoning.
The overarching problem that AI can help fix is the administrative burdens on frontline staff. “Any time you have someone reading or skimming large volumes of information and making logical, rules-based decisions, you are going to have an AI opportunity,” says Clark, notably to “augment and amplify the efforts of the care team.”
Big Win In Oncology
In late 2016, Digital Reasoning’s then-Chief Technology Officer Matthew Russell and some of its data scientists began experiments in “deep collaboration” with HCA’s Vice President and Chief Data Scientist Edmund Jackson and Chief Medical Officer Jonathan Perlin, M.D. The experiments ran the gamut from early prediction of sepsis to interpreting doctors’ and nurses’ notes to what was ultimately selected to operationalize and scale: an AI-powered cancer app that reads pathology and radiology reports in real time, and then helps ensure care is rendered quickly and in the best prioritized order, says Clark.
The reality is that referring physicians—the ones responsible for informing biopsied patients about their diagnosis and handoff to the appropriate care team—are very busy and don’t always respond to findings at optimal speed, says Clark. Response times can be particularly sluggish among patients whose cancer was “incidentally” detected by a routine screen or imaging test looking for something other than cancer.
Studies suggest that only about 30% of incidental radiologic findings of cancer are followed up on in a timely manner, says Clark, leaving some patients undiagnosed for months. Once cancer is diagnosed, there is typically no outreach from the referring physician or it’s “not as guided or comprehensive as it could be” to meet patients’ emotional needs or alleviate their treatment concerns.
Over the 12-month HCA pilot in three markets, the cancer app successfully sped up the diagnosis and treatment of cancer by an average of five days per patient, says Clark, which is “real important especially for certain types of cancer.” It also took over tasks previously the responsibility of care teams—searching through pathology reports for positive results, classifying the different cancer types, assigning the most severe and complex patients to care navigators, and extracting quality reporting measures for submission to various registries and governmental bodies. Further, the software triaged patients by severity, queuing up the workflow for local care teams.
As a result, clinicians started spending more time on care coordination and delivery and less on the documentation and computer work associated with it. The overall number of cancer patients identified, navigated and treated more than doubled, Clark says, enabling higher revenues while maintaining the same labor cost structure.
Scaling A Use Case
Digital Reasoning had a “very busy” 2017 scaling its AI solution across all 165 HCA hospitals, says Clark, where outreach and care navigation begin within 72 hours of diagnosis. “The fun part is that we’ve embedded ourselves invisibly behind the scenes so that the nurse navigators, care coordinators, registrars, and doctors especially, aren’t doing anything different than they normally do. They don’t have to learn a new system or log into something else.”
In true AI fashion, the software also continuously learns, Clark notes. “Nothing is 100% accurate, so if a particular pathology report was flagged as positive but was really negative, then that gets noted by the care coordinators and put in a discard queue. We track every time they interact with the system and feed that back into future versions of the model.”
Doctors think of this as “a concierge service for their patients, ensuring they receive high-end oncology navigation,” says Clark. The benefits of patient navigation, particularly in oncology, have been well documented over the past two decades. The practice has been linked to improved survivorship, patient satisfaction, medication adherence and other quality-of-care and quality-of-life indicators.
“The limitation has always been that employing cancer nurse navigators is very labor-intensive and highly expensive,” continues Clark. By lowering the cost per patient to receive care navigation, the
soon-to-market software should enable more health systems to either expand or start an oncology navigation program. Budgetary constraints have forced many organizations to limit their program’s scope to a small number of site-specific cancers. “HCA told us that using our software last year was equivalent to hiring 186 nurse navigators across the system.”
Jonathan Perlin, M.D., president of clinical services and chief medical officer for HCA Healthcare, says the collaboration with Digital Reasoning has provided HCA with a “meaningful, technologically advanced tool” to help access a large volume of unstructured clinical data. “While our initial use of this technology is in support of improving the timeliness and effectiveness of care in oncology, we continue to work with Digital Reasoning on other natural language processing applications that will enhance workflow for physicians and caregivers and improve outcomes for the patients we are privileged to serve.”
Three other health systems will now join HCA as first adopters committed to co-innovating with Digital Reasoning on new use cases—strictly following the scientific method and sharing all results, good and bad, during the journey, says Clark.
Clark envisions the software extending vertically into additional high-labor, low-value tasks within cancer, as well as horizontally into other clinical areas and specialties. Within oncology, that would include identifying and extracting quality measures for various reports and, outside of oncology, care management functions for patients with chronic diseases like COPD. It may also be possible for the app to expedite the clinical trial matching process.
Stepping Stones In Behavioral Health
At the recent Health:Further festival in Nashville, Clark was joined on the behavioral health stage by Jim Stefansic, president and CEO of predictive analytics startup Raiven Healthcare, to discuss the fundamentals of successfully applying AI technology in patient care environments.
The quality of data, and not the quantity of it, is key, Stefansic says. The initial opportunities with risk-bearing organizations are to find the 10-12 variables that most accurately predict and prescribe behaviors. “The good news is that more and more of this “rich” data is becoming available through case management software and treatment pathways”—e.g., if a phone call was or wasn’t effective in helping a cancer patient manage depression. The constraint is that the data tends to be sensitive and therefore can be difficult to access.
One good starting point for innovators is to “at least minimally validate” their hypothesis on real or synthetic data that’s publicly available, says Clark, citing open-source deep learning tools ranging from Intel’s BigDL to Google’s TensorFlow. With growing adoption of the FHIR standard for exchanging healthcare information electronically, IT staff will be spending less time maintaining interoperable connections, he adds, leaving them with more time for AI-related projects.
“We’ve had some health systems tell us they’ve been maintaining over 500 different HL7 interfaces,” says Clark. “FHIR API is going to significantly open the floodgates by reducing the amount of physical maintenance that must be done to standardize the communication and transfer of data.”
In the behavioral health field, the challenge is not so much identifying people with mental health needs—30% to 40% of us have a diagnosable condition we could use help with—but the limited number of treatment providers, says Clark. AI’s potential is in improving the “awareness and ability of alternative treatments,” including robots. “A couple of recent studies have shown that humans are more likely to open up about their mental health needs to a robot than to a human.” That means 20 years from now cars may not only be self-driving but also providing mental and behavioral health management and coaching.