HIT Consultant | June 5, 2020
By: Jeremiah Stone, Ontrak, Inc. Chief Technology Officer
Prior to the COVID-19 pandemic, a study found that almost half of Americans will experience an episode of mental illness in their lives, but may never get the diagnosis and treatment they need. In a more recent finding, over 25% of American adults now meet the criteria for a diagnosis of severe mental distress.
The looming impact on our collective mental health is concerning enough, but there are also serious implications for physical health. Left unaddressed, behavioral health issues exacerbate chronic conditions such as diabetes, hypertension, obesity, and other conditions—further continuing a tragic cycle of chronic disease in this country that has put so many at severe risk if they contract COVID-19.
Moreover, not all who struggle under the burden of chronic disease and untreated mental health challenges will engage in care. The reasons vary—and can include stigma, worry over costs, and limited access.
However, it is now possible to identify those with unmanaged or undermanaged health conditions due to underlying behavioral and social barriers, and proactively reach out to them with an individualized approach. Advances in machine learning have opened up the ability to analyze medical claims and other data to predict the probability of an underlying behavioral health condition – like anxiety, depression, or substance abuse – and to tailor outreach appropriately to encourage engagement.
Most health plan analysis focuses on population health based on standard risk groupers. Even social determinants of health tend to be analyzed by zip code rather than on a case-by-case basis. With machine learning, we can take a more individualized approach, using algorithms to link certain behavioral attributes in individual claims with a likely underlying behavioral health condition that is driving overutilization.
In a specific example, consider someone who suffers from Type II diabetes. In the past year, claims showed the person experienced several emergency department and inpatient visits due to insulin shock. After each admission, the patient received support from the health plan’s care team, yet the pattern persisted. Unknown to the hospital and health plan, the patient also suffered from acute pancreatitis— which, to a physician who specializes in this area, is a key indicator for alcohol substance use disorders.
We can train a machine-learning algorithm to “score” this individual based on their behavior as having a high probability of a substance use disorder and flag them as a candidate for targeted outreach.
Of course, a reliable algorithm relies on massive sets of clean data—as in data that has been properly normalized and standardized. And the average set of claims data needed to make accurate predictions is comprised of dozens of arbitrary formats. But even there we can apply machine learning to map all of the data into a uniform format.
Although the above work is highly automated, it is a precursor to sustained engagement between experienced care coaches and health plan members. The truth is, many digital apps and other programs are unable to successfully engage people
to manage their anxiety and depression which often accompanies chronic conditions. However, when care coaches use non-judgmental listening to create a foundation of trust, members find it easier to be forthcoming about their situation and identify the root causes of their distress.
In turn, this motivates patients to collaborate with coaches on finding coping mechanisms, tools, and resources to effectively self-manage their situation in ways that improve their quality of life, as well as clinical outcomes. These insights are currently informing the design of next-generation digital coaching tools, so in a not too distant future, we can expect machine learning to play an essential role in engagement, too.