Searching for social prescribing: A.I. and digital tools for social determinants of health

The patient-provider relationship has evolved.

Healthcare is important, but the increasing focus on whole-person care is shifting how doctors and other providers assess patient well-being beyond traditional clinical considerations.

A recent study points to an imbalance between spending on the types of care that can improve longevity and reduce risk of premature death for patients: Behavioral and social factors account for between 16% and 65% of premature deaths, compared to 5% to 15% tied to healthcare.1

While there are “complex interrelationships” between healthcare and the various social determinants of health (SDoH), as writes Steven H. Woolf, MD, MPH, Virginia Commonwealth University, Richmond, Va., “the average lifespan of Americans will probably continue to shorten unless society quickly shifts its focus … to root causes” that exist beyond healthcare.2

While some provider organizations have built specific methods to survey patients for SDoH factors (see page 98), the broader goals of population health management require the collected data to be shared broadly. This hurdle remains particularly crucial, as health information exchange (HIE) leaders across the country point to SDoH and behavioral health data as the most difficult types of information to share, per an eHealth Initiative survey of HIE technological priorities.3

Even the phrase “social determinants of health” may not impart its importance, according to Patricia Birch, MBA, senior vice president and global practice leader, Cognizant Healthcare: “It’s sort of an awkward name for things that are at the bottom of Maslow’s hierarchy of needs,” Birch said earlier this year at the HIMSS 2019 Global Conference & Exhibition. Nevertheless, Birch contends that focusing on SDoH is a key mission to help move the needle on healthcare costs and quality: “We can’t have our social services as a totally separate ecosystem from our healthcare services.

“There’s a lot of other things that impact healthcare, other than your biology, genetics or the type of medical care you get,” Birch added. “The bulk of what impacts your health is your physical environment, your social circumstances. And those, in turn, drive your individual behavior.”

This is where Mount Sinai Medical Center in New York City enters the picture. With 7,800 physicians and 6,340 faculty members across eight hospitals, a medical school, nine ambulatory surgical centers and 190 remote clinical and administrative sites, Mount Sinai has more than 3 million outpatient visits annually, along with more than 500,000 ER visits and more than 130,000 inpatient visits. That scope alone positions Mount Sinai as a place where getting accurate data could drive a lot of improvement of clinical models and understanding of disease progression through SDoH work.

“When I worked in a health system, one of the big issues we were always confronting was we have this community mission,” Birch said, which meant hosting health fairs and other community events, “but that didn’t really move the needle. … What health systems are coming to grips with is that if they’re really, truly going to fulfill that part of the mission, [they’re] going to have to address … more business-oriented things like cost, financial risk … they have to really think about treating the whole person, which means taking a look at these social determinants of health.”

The challenge, however, is in finding the mechanism to capture that information: Most patient screenings or surveys are manual, aren’t integrated into an EHR or easily accessible in a place where the providers and care managers who need it can access it. “This creates a very, very inefficient type of system, and there hasn’t been a real emphasis on trying to close this loop,” she said, beyond organizations that have social service departments or discharge planning.

But with advancements in artificial intelligence (A.I.) and machine learning, “we really are starting to have the types of tools that can help us address this problem,” Birch said, such as taking existing EHR data from clinical notes and building models with natural language processing (NLP) technology to identify SDoH.

Varun Gupta, IT director of analytics and data management, Mount Sinai Medical Center, noted that NLP has been around for some time but has “come a long way” to serve new purposes in healthcare specific to patient interaction, engagement and care coordination.

Specific clinical data, despite living in the EHR, was “never analyzed at an aggregate level,” Gupta said, until the organization brought NLP technology to the forefront to help clinicians and care managers get more meaning out of the data.

That meant using on-premise infrastructure — in this case, an Oracle platform — to take EHR data sources, clean them up and prepare them for a combination of cloud platforms — including an FHIR-based API platform, a data lake and hybrid cloud-based architecture — to create a true analytics ecosystem that could pull insights from clinical notes.

SDoH categories

Mount Sinai Medical Center came up with 11 categories into which to divide individual determinants of health to phase the project:

Phase 1

• Economic stability: housing, income, food, nutrition
• Education
• Healthcare system: medication, prevention care, insurance, language barriers
• Physical environment: transportation, overall safety

Phase 2

• Behavioral health: depression/anxiety, substance abuse, sleep
• Care-giving responsibility
• Legal: divorce, child custody, alimony
• Support system: home health, friends, neighbors, relatives
• Physical activity

Phase 3

• Fall safety
• Special health needs: vision, hearing, urinary incontinence

 
Gupta said the technological teams worked with clinical subject matter experts (SMEs) to customize the NLP workflow to define how to structure the format of information extracted from clinical notes with terminology from various specialties, such as terms specific to clinical oncology. “The main effort here was preparing the ‘bag of words,’” Gupta said.

To test the model, all Medicaid clinical notes added to the database from December 2017 to July 2018 and then sample outputs for four different SDoH areas — economic stability, education, healthcare system and physical environment — were validated by the team’s clinical SMEs to set the model’s parameters. Accuracy of that testing ranged from 86% to 97% for the four areas. This included negation work to not count instances of a patient saying something like, “I don’t have a housing problem,” or a patient referring to someone else having a problem that otherwise would be identified by the NLP technology.

What they found: Out of 7.24 million encounters with 226,368 Medicaid patients, 70,541 patients had some type of SDoH found in the clinical notes.

But not everything was automated, Gupta noted: “We actually looked at 5,000 to 6,000 clinical notes manually to improve the accuracy to a level where we wanted to,” he said. “It was not the innovation, not the technical part that was responsible for the success — it was the clinical involvement which actually made the project successful.”

With a single source of SDoH data, Mount Sinai can work to flag patient needs and build dashboards for care managers to handle high-need patients. In terms of moving the needle for cost, Gupta said, it’s shifting patients out of expensive, acute-care settings and into primary care and community services.

“There’s also been some studies that have said what we need to also introduce into the whole clinical workflow is something called social prescribing … to address the social determinants,” Birch said. “You have to have enough information to do that.”

Notes:

1. Kaplan R, Milstein A. “Contributions of health care to longevity: A review of 4 estimation methods.” Ann Fam Med. May/June 2019, vol. 17, no. 3, 267-272. doi: 10.1370/afm.2362.
2. Woolf SH. “Necessary but not sufficient: Why health care alone cannot improve population health and reduce health inequities.” Ann Fam Med. May/June 2019, vol. 17, no. 3, 196-199. doi: 10.1370/afm.2395.
3. “2019 Survey on HIE Technology Priorities.” eHealth Initiatives. May 15, 2019. Available from: bit.ly/2JvAKAZ.