Leaders at Mount Sinai Medical Center in New York City have been stepping up their efforts to identify and address social determinants of health. One impetus is the hospital’s increased engagement in value-based payment models, says Varun Gupta (pictured at right), IT director of analytics and data management.
But uncovering social determinants can be difficult, given that most of the “clues” remain hidden in unstructured data within the clinical notes, progress notes and discharge summaries in the electronic health record (EHR). So Gupta and his team developed an algorithm that uses an artificial-intelligence approach called natural language processing (NLP) to detect patients’ social needs from the EHR’s vast amounts of unstructured data.
Gupta’s team built the algorithm to look for four categories of social determinants of health:
- Economic stability
- Physical environment
To test the algorithm, they analyzed the clinical notes from more than 7 million encounters with 220,000 Medicaid patients during an 18-month period. Nearly one-third of the patients had at least one social factor that had not been uncovered through other means, they found.
The algorithm was most accurate (97%) when determining economic stability factors, such as income, and least accurate (86%) when determining physical environment factors, such as whether the patients’ living quarters placed them at risk for falls.
Putting the information to use
Leaders at Mount Sinai are determining how best to equip care managers and physician practices with the social determinants data, which is now located in a single repository.
For care managers, they plan to create target lists and dashboards that list high-priority patients. For physician practices, they are exploring the possibilities of integrating the data in the EHR so that patients with greater social needs are “flagged” for clinicians at the point of care. Ultimately, they plan to apply the same approaches to patients with commercial insurance.
For these advanced algorithms to succeed, Gupta says, clinical and technical subject-matter experts should create a comprehensive set of “rules” that define how the models will work. For example, an effective NLP algorithm would not tag a patient as currently homeless if the EHR reads that “the patient has a hx [history] of homelessness.”
Despite the promise of these approaches, finance leaders should not expect immediate results, Gupta says. “These projects are in uncharted territory, and you can’t always predict a particular outcome or ROI,” he says. “These projects are experimental and need to be done to drive an organization toward better outcomes, so there is always some risk on the financial side.
“But senior leaders and CIOs should be encouraging these initiatives, as they have potential to generate better patient outcomes and provide insights when they are successful.”
Interviewed for this article:
Varun Gupta, IT director, analytics and data management, Mount Sinai Medical Center, New York City.