Healthcare Analytics, Population Health Management, Healthcare Big Data

Tools & Strategies News

Using Machine Learning to Target Behavioral Health Interventions

Machine learning tools can help providers target behavioral health interventions to patients who need help managing mental and physical chronic conditions.

Behavioral healthcare and machine learning

Source: Thinkstock

By Jennifer Bresnick

- Behavioral healthcare has been getting a great deal of attention lately from health IT experts, and not just because of a growing recognition that good mental health is key to improving overall patients outcomes. 

Behavioral health is one of the most complex, highly individualized, and notoriously underfunded components of the care continuum, which makes it a perfect test bed for deploying advanced machine learning tools that aim to tighten up care processes, improve access to resources, and uncover new insights into hidden patterns of treatment and disease.

At Beacon Health Options, a behavioral health management service provider partnering with health plans and employers, identifying high-risk patients before they enter a crisis is a financial and clinical imperative – one that has broad implications for an incredibly vulnerable and difficult-to-treat population.   

Emma Stanton, Associate Chief Medical Officer at Beacon Health Options
Emma Stanton, BM MRCPsych MBA, Associate Chief Medical Officer for Beacon Health Options Source: Emma Stanton

“Mental healthcare is a perfect example of chronic disease with wide-ranging impacts and a massive number of variables,” explained Dr. Emma Stanton, Associate Chief Medical Officer for Beacon Health Options.

“We don’t always have a clear start time for these illnesses, there isn’t a final resolution date, and quite often it isn’t the acuity of the disease that predicts the outcome.  There are so many complex factors, like economic status, social support, and access to care, that will determine whether or not someone ends up in the hospital.  It’s not just the extent to which an individual is feeling suicidal or hearing voices.”

READ MORE: How Healthcare Can Prep for Artificial Intelligence, Machine Learning

Patients with serious mental illnesses, including depression, schizophrenia, or bipolar mood disorders, die more than twenty years earlier, on average, than patients without these conditions, says the World Health Organization.  They also tend to receive less comprehensive and effective preventive care and have a higher prevalence of chronic disease.

“They’re not dying from the mental illness itself – they’re dying from physical health comorbidities which are often untreated because of an inability to access the necessary pieces of the healthcare system,” Stanton told HealthITAnalytics.com

“They’re dying from complications from diabetes, obesity, or cardiovascular disease, because they often have problems staying on track with managing chronic diseases due to the range of daily challenges they face.”

Beacon helps approximately 50 million patients across the country cope with these problems by connecting them with the clinical and community resources they need to navigate their illnesses.

But guiding vulnerable patients through the numerous facets of their illnesses is even more difficult when providers, care coordinators, and behavioral health specialists don’t have access to the big data they need to make informed, tailored decisions, Stanton said.

READ MORE: The Difference Between Big Data and Smart Data in Healthcare

Mental healthcare and substance abuse care are incredibly siloed, often stigmatized, and chronically under-resourced.  In many of the states we serve, there is a shortage of providers who take insurance for behavioral health issues, which complicates the problem.”

“Not to mention the fact that we have rising rates of opioid use disorders, and it is a challenge to drive access to medication-assisted treatments, like naloxone and methadone, which are real lifesavers for patients at the moment,” she added.

To address these issues, Beacon is turning to one of the most promising health IT strategies available: machine learning that can extract actionable insights from structured and unstructured data alike.

“We sit on an awful lot of data, which is organized in a very traditional claims-based model,” said Stanton.  “We can see whether someone has had an outpatient appointment or an inpatient admission, but the data doesn’t tell us a great deal about whether or not the patient has actually gotten better as a result of accessing that care.”

“So while we are a data-driven company and rely on that information for everything we do, we are keenly aware that there are limitations in our insights due to the way that data is organized and analyzed.  There is a tremendously exciting opportunity to use machine learning to improve those processes and dig deeper into that data and all the other variables that impact an individual’s life.”

READ MORE: Using Risk Scores, Stratification for Population Health Management

Beacon has entered into a partnership with a machine learning company called Cyft to integrate more individualized risk stratification technology into its patient management processes.  

“The way the healthcare system uses data right now is to create one-size-fits-all models and then bring them from healthcare organization to healthcare organization under the assumption that all populations are generally the same,” said Leonard D’Avolio, PhD, Founder and CEO of Cyft and an Assistant Professor in the Brigham and Women's Division of General Internal Medicine and Primary Care.  

Leonard D’Avolio, PhD, Founder and CEO of Cyft
Leonard D’Avolio, PhD, Founder and CEO of Cyft, Assistant Professor at Brigham and Women's Hospital Source: Leonard D'Avolio

“But those models don’t always account for the outliers who are in the most need of more intensive care, like those with the lowest socioeconomic status and the ones with the greatest number of challenges.”

Unstructured data can help to close those gaps, said D’Avolio, by supplementing data traditionally sourced from the electronic health record with a slew of supporting information about what happens to patients outside of the clinic.          

“Behavioral healthcare gives us an opportunity to take data that exists in non-traditional formats – whether it’s patient satisfaction surveys or call center transcripts or free-text data in the EHR – to create far more precise, tailored, and actionable recommendations to help members,” he said.

“Traditional risk modeling often just considers claims data and uses between five and seven variables to tell the user who needs attention.  In another one of our projects which is around preventing hospital admissions for diabetics, our highest-performing models are considering 30 different data sources, structured and unstructured, that add up to 438 different variables.  That gives us some impression about the potential of machine learning and why folks are so excited about it.”

While there is enormous potential for unstructured data to contribute to providers’ understanding of the socioeconomic, environmental, and other non-clinical factors impacting a patient’s care, these data sources can also be frustratingly messy, incomplete, and fragmented.

Providers, vendors, and researchers employing machine learning techniques including natural language processing, pattern recognition, optical character recognition, and voice recognition have already had success turning audio recordings, scanned PDFs, imaging studies, and other non-traditional data sources into machine-readable input. 

Computer scientists have been exploring the potential of machine learning in earnest for at least thirty years, D’Avolio said, but despite the long history of interest in the subject, few healthcare organizations have yet been able to integrate these analytics into the daily clinical workflow in a meaningful, impactful manner. 

“While we are well aware that machine learning works, we are also aware that it can take nine months and a thousand man hours for a team of highly specialized data scientists to create a single model for a specific problem.  The trick is to reduce the complexity of this process to the fewest number of steps so that we don’t have to spend months and months on issues like data cleaning and interoperability.”

“The missing piece right now is how to translate that information into better care, but it’s just a matter of time before providers fine tune those strategies and start to see some amazing results.”

Applying these enriched, streamlined tools to the behavioral healthcare space will give case managers the ability to identify patients at high risk for falling through gaps in the care continuum, said Stanton. 

One of Beacon’s first use cases will be enhancing care coordination for patients with serious mental illnesses and accompanying physical illnesses.

“Our goal is to move from being a reactive model that solely looks at what has happened historically to being a much more predictive, proactive, and targeted service provider,” she said.

“It’s an opportunity to bridge the siloes that exist in the healthcare delivery system, and it’s an example of where machine learning can help to bulldoze through those traditional barriers to make progress for an incredibly vulnerable segment of the patient population.”

More tailored risk models and patient identification tools will help Beacon ensure that patients have the support they need to stay engaged with their providers, maintain safe and stable living conditions, and remain adherent to their care plans. 

“The data will go directly to our program director and the care managers and enrollment coordinators to target the people who are most at risk so that we can get out there and find them,” Stanton said. 

“Machine learning is exciting because it is helping us tease out the complexity around behavioral healthcare, and it is helping us to look at these deep-seated issues in a brand new light.”

Continue to site...