Seattle healthcare IT startup KenSci has a tagline that helps simplify the company’s audacious aims: Death versus data science.
Unpack that a little bit and you’ll find a company—banging its drum Wednesday with an $8.5 million Series A funding round led by Ignition Partners—that’s combing of-the-moment machine learning technology and mountains of data to attack some of the wickedest problems in the healthcare industry: problems like high hospital readmission rates, hospital-acquired infections, and over-use of costly emergency room services for conditions that might have been addressed through preventive care before they became acute.
“These are the problems that are very amenable to machine learning and artificial intelligence,” says KenSci CEO Samir Manjure.
KenSci, co-founded in spring 2015 by Manjure and Ankur Teredesai, a University of Washington computer science professor, has already won 11 enterprise customers, including Fullerton Health, St. Luke’s Health Partners, and other large public and private health systems. The funding round, joined by Osage University Partners and Mindset Ventures, will be used for hiring across the company, sales and marketing, and continued research and development.
The company sets machine learning algorithms to work on digitized healthcare data: electronic medical records, claims data, demographics, psycho-social data, and patient-generated data from sensors and personal devices. The company trains scores of separate machine learning models to predict risks for individual patients and entire insured populations across various diseases.
“The idea is to predict who’s going to get sick, and how sick are they going to get,” Manjure says. “If we can understand disease progression and the complex interplay of those variables, then we can help coordinate, proactively, their care in anticipation of those adverse events…This is a radical approach in terms of how we go about saving costs for healthcare, while shooting for better outcomes.”
The startup has big competitors taking similar approaches, such as IBM Watson Health and Health Catalyst. Manjure says one way KenSci can differentiate itself as a startup is through rapid deployment and faster return on investment.
The company—funded until now with customer revenue, Manjure says—is a textbook example of what the Seattle area has to offer when it comes to healthcare IT innovation.
Teredesai, the company’s CTO, is executive director of the Center for Data Science at University of Washington, Tacoma, where the company was incubated. His research over the last half-decade gave KenSci its foundation, and the company has licensed intellectual property from the UW and continues to collaborate with researchers there. Manjure left Microsoft after 17 years to found KenSci, but his startup has benefitted from close ties to the company. KenSci was part of Microsoft’s Seattle startup accelerator class last fall and has been highlighted for its use of Microsoft’s Azure cloud.
Manjure and Teredesai go way back—they grew up in the same neighborhood, and their mothers went to the same high school. They had been talking about launching a project together for some time, Manjure says, and they both wanted it to be more than just a money-making endeavor.
Reducing the death rate from preventable killers like sepsis fits that bill.
Where a doctor might typically diagnose sepsis by looking at a handful of symptoms—temperature, heartrate, respiratory rate—KenSci’s machine learning model crunches hundreds of variables for each patient and provides a predictive score of that individual’s risk of infection.
That score and the other outputs of KenSci’s growing bank of models deliver results to various individuals in the healthcare system—from front-line care providers to insurance CFOs—through data visualizations, reports, and application programming interfaces. Manjure showed off a 23-year-old piece of software used by hospital discharge planners to track who is getting better. Through an API, KenSci’s scores are integrated into that tool.
“The discharge planners and the doctors don’t have to change their workflow, but yet are able to get this very advanced insight from machine learning algorithms in a way that is natural to them,” Manjure says.
But are doctors ready to defer to the judgement of a complex algorithm, whose inner workings they may or may not understand?
Greg McKelvey, a doctor and KenSci’s head of clinical insights, says the company emphasizes transparency in all of its design decisions.
“We cannot be operating in black boxes,” he says. “We have to show why we’re making predictions [and] how well they’re performing on the populations that are actually being cared for. For a clinician—to really trust this—they need to see this for the patients that are in front of them, is it actually performing as described?”
When a new customer begins using KenSci, the company’s machine learning models are refined using two years of data from the customer’s specific population. The models then make predictions for a third year, and those predictions are validated against actual outcomes for that year.
The performance of KenSci’s predictions are shown against baseline clinical prediction models that may already be in use in another bid to show physicians how this new tool compares to their status quo practices.
For now, KenSci’s predictive models are meant to be “natural extensions of the tools that physicians are already using,” McKelvey says. “They’re merely an adjunct to their clinical judgement.”
But it’s easy to envision a system like this going from prediction to prescription, which will open “a much thornier discourse” about things like the autonomy of physicians, reimbursement, regulation, and liability, he says.
The insights KenSci promises to physicians are multiplied for large healthcare organizations—such as accountable care organizations and health insurance providers—which are struggling to better manage their covered populations and control costs. Healthcare payers are increasingly tying reimbursements to patient and population outcomes, rather than simply paying for procedures performed.
“Imagine the world where you can predict [the] 1,000 people that are at the highest risk for you next year, and these are the reasons why they are at the risk,” Manjure says. “Now you can enroll them into the right kind of programs.”
That might mean identifying someone at imminent risk of cardiac arrest before they arrive at an emergency room with chest pain, and before they incur a $25,000 medical bill. “If you can predict about this patient up front, then you can actually treat them, even if it means sending a nurse or a doctor to their home, which may cost less than $1,000,” Manjure says. “So right there, you’re having not only a better medical outcome for a patient, but also reducing significant costs in healthcare.”