Understanding whole person indicators of wellness using machine learning methodologies

In the world of machine learning and healthcare, one of the newest and emerging areas of exploration is understanding how to account for social and behavioral determinants of health (SBDoH) in predictive modeling. SBDoH factors decisively influence the interaction of four other types of determining factors: genetics, environmental factors, health-related personal behaviors, and the degree of access to health care (such as the level of insurance coverage in the population).

There is overwhelming evidence to suggest that SDBoH impacts both onset and progression of health conditions, clinical care processes to address conditions, and ultimately health outcomes. Access to healthcare is a critical determinant of health status, but personal behaviors are actually a much greater influence on health over a lifetime. Environmental factors and genetics are also important factors, while socioeconomic status (especially educational attainment) influences every determinant of our health.

However, our national expenditure on health care delivery is far greater than our expenditure on other critical determinants of health. According to the CDC, nearly 90 percent of personal health care expenditures in the U.S. are spent on direct care; much less is spent on changing risky behaviors or reducing environmental risks which means we continue to spend the majority of healthcare dollars on access to medical care even though it drives approximately 10 percent of health outcomes.

Integrating SBDoH within clinical contexts has been growing greatly in the US, given the focus on the Triple Aim of healthcare: improved population health, better patient experience of care, and reduced per capita cost of health care. In 2015, the National Academy of Medicine (NAM) recommended social and behavioral domains to include in electronic health records which included five main domains of variables:

• Sociodemographic (race/ethnicity, education, employment);
• Psychological (health literacy, stress, depression, anxiety);
• Behavioral (diet, activity, tobacco or alcohol use);
• Individual-level social relationships and living conditions domains (social connections, work conditions, exposure to violence);
• Neighborhoods and communities (neighborhood and community compositional characteristics).

This nationwide group recommended standard domains to be collected in healthcare settings and also integrated into accessible electronic data sources. Interest in the integration of SDBoH within clinical environments will continue to grow considering the Center for Medicare and Medicaid Innovation (CMMI) began a 5-year, $157 million grant initiative to implement social needs screening and community resource linkage programs for Medicare and Medicaid beneficiaries.

With this increase in resources to better understand the SDBoH space, however, we find that evaluations of SDBoH interventions are still limited by poor study quality for many reasons. Current research has reported that higher-quality studies are needed.

In addition, a recent analysis in the New England Journal of Medicine reported that more investment in health information technology and methodologies are required to better integrate scalable population-based SBDoH interventions within clinical environments. Considering the complexity of the pathways between social and behavioral risk factors and health/utilization outcomes, researchers are exploring machine learning methodologies to better predict patient outcomes and healthcare cost and utilization.

Researchers have discovered that underlying social factors are important predictors of biological indicators of chronic disease, and commentary about the opportunity that emerging datasources present for better understanding how to improve health outcomes. However, how to use and interpret information from consumer or social media data from models predicting health outcomes is still not clear.

There are efforts underway to apply gold standard definitions and metrics to envision how to account for SBDoH factors in case-mix adjustment and risk stratification. It’s important to note how can we use patterns of behavior data, along with sociodemographic and neighborhood-level data, to more accurately measure an individual’s wellness. There is a need to better recognize the ‘whole person’ associated with health and wellness, beyond quantitative measures of blood pressure, cholesterol, or blood glucose which are often measured in direct clinical contact settings.

Expanding the definition of wellness to include the five domains presented by the NAM recommendations as well as integrating additional (and emerging) data types describing indications of psychological or behavioral well-being can provide a more complete glimpse into an individual’s well-being. This will improve the precision of models that can be built to predict an individual’s ability to respond to treatment, trust a provider, be adherent to medication to treat chronic or acute disease, regain post-operative functionality, or deliver a healthy baby.

Population-based models that are adjusted for SBDoH may also provide ways to understand a patient’s readiness for change which is crucial as the population of people with multiple, chronic conditions increase. Precise measures of wellness will also contribute to more complete information to inform clinical decision making, personalized care and risk stratification, optimizing care patterns, and real-time course-correcting management of care journeys. By linking, analyzing, and interpreting datasets that may not yet be fully explored in the healthcare space, we can better understand how a more complete historical view of patient care is formed, thereby exposing indications of wellness.

Photo: metamorworks, Getty Images