Premier Inc., a health IT firm that employs data and machine learning to solve a variety of health care industry problems, is using it to modernize syndromic surveillance. They call on states to work with their local providers in using the system to monitor COVID-19 symptoms in real time. Based on machine learning, the system should work much faster and produce more accurate predictions.
Because the Premier system is new, we don’t have on-the-ground results to confirm these hopes. But I am publishing this article to prompt states and providers to take a close look at the solution.
A Serendipitous Initiative
I talked this past Monday with Mike Alkire, President of Premier, and Ryan Nellis, the Chief Commercial Operations Officer of Stanson Health, a Premier Company. They told me that they got the COVID-19 system working by repurposing a system already embedded for a very different use in several big electronic health record systems.
As background, Premier offers predictive analytics and clinical decision support quickly, sometimes in real time, to hospitals and ambulatory physicians. This can serve various needs ranging from clinical interventions to improving supply chains. To derive the insights, Premier integrates its analytics into all the most popular U.S. patient EHRs, including Epic, Cerner, and Athenahealth. Over 4,000 hospitals and 175,000 other provider organizations tap into Premier’s analytics today, a network that can be useful to scale up COVID-19 surveillance fast.
The system they developed was designed to help hospitals comply with imaging regulations designed by the Centers for Medicare & Medicaid Services, as part of implementing the Protecting Access to Medicare Act of 2014 (PAMA). In brief, radiologists and imaging centers are penalized for prescribing unnecessary CT scans or other expensive imaging tests without first consulting a qualified clinical decision support technology. It’s extremely valuable to a health care provider to gather all the data relevant to this decision in real time for each patient before prescribing the test. Premier’s Stanson Health designed a system to get the data on which they could provide clinical decision support and analytics to guide decisions.
After the onslaught of COVID-19, Premier realized they could repurpose this system to track symptoms in an ambulatory population in real time. They wanted not only to identify who had COVID-19 symptoms, but to determine which patients are likely to require an upcoming hospitalization before the actual admission occurs. This called for intensive analysis of patient data: age, vitals, gender, comorbidities, and more.
They based their initial model on data published by a hospital in the COVID-19 hot spot New York City, describing all the data the hospital collected on its patients and what the data told them about the likely trajectory of the patient.
Premier could get all the data they needed from the health records, both in-person visits and telehealth. But interpreting it programmatically wasn’t always easy. Sometimes, key diagnostic criteria are buried in unstructured text notes, and doctors may use a lot of different words or abbreviations for essential symptoms such as “shortness of breath”. Premier used natural language processing (NLP) to extract the necessary data from the text.