Identifying & Taking Action on Social Determinants of Health
Blending insights gleaned from traditional claims and clinical data with insights from consumer data and survey data provides us with a more focused view of an individual’s predicted future behavior, and their barriers and motivations to improved health.
The key to all this, however, is to make the data actionable and straightforward to use. With so much data available from disparate data sources, how can we distill what’s important and incorporate it in our day-to-day conversations with patients and members. More importantly, how can this data be used in a way to improve the health of populations?
A Twist on Using Traditional data
Traditional administrative and clinical data is “bread and butter” data that healthcare organizations have been using for years to make determinations on historical and predicted behavior. This data delivers tremendous value in determining each member’s disease history, utilization patterns, age, gender and so forth. That said, this data can deliver additional value beyond the straightforward assessment of disease and utilization. By manipulating the data, we can create additional insights and create more value in this data, and use these inferences to test hypotheses.
A simple example is determining each member’s distance to their primary care provider. If we know the member’s address and the primary care provider’s address, we can calculate distance. If we then overlay the availability of public transportation in the member’s neighborhood (available with census and other third-party data), the member’s historical and predicted engagement with primary care, and even the member and provider’s primary language, we can identify potential access issues. If our goal is primary care engagement, can we boost primary care engagement by helping the member find a doctor that speaks their language and practices closer to their home?
A more complex example involves observing multiple patterns, and taking action on these patterns. For example, a member may present with intermittent clusters of emergency room and out-of-network usage, with little historical or predicted primary care engagement. This alone is a concern. However, if the member also presents with historical and predicted future hospitalizations coupled with no follow-up primary care engagement, the data strongly signals potential health literacy, health system literacy and access issues. A care management approach to educate the member on their condition, on using the healthcare system, understanding and addressing their barriers to care, and developing a relationship with their primary care provider will go a long way to lower avoidable utilization.
Non-Traditional Data – Consumer Behavior
When integrated with traditional data, consumer data can deliver significant value in providing insights into member behavior. There is tremendous breadth and depth in consumer data, and to use this data effectively, healthcare organizations need to understand which specific data element can deliver value in solving population health.
Machine learning can help with this. The question that machine learning can help answer is which data elements in consumer data (either by themselves or together with other variables) are correlated with specific outcomes. For example, which data elements are correlated with multiple admissions, which elements are correlated with medication adherence, etc. By identifying correlations in advance, specific data elements can be selected to be included in a machine learning-based predictive model as an important variable. For example, household size, income, net worth, education and credit card usage are important predictors of avoidable utilization when used in combination with traditional data. This means that if a member presents with a complex disease and utilization profile, lives alone, doesn’t have credit cards and is low income, the predictive model will likely identify the individual as higher risk for avoidable admissions.
A critical part of using consumer data is differentiating between what data elements are useful in making predictions about an individual’s behavior, and what data elements can be used effectively for segmentation. Segmentation is the practice of finding commonalities with certain groups of members so that communication to these groups is as targeted and personalized as possible. In the example above, the individual is identified as high risk for avoidable hospitalization combined with potentially significant socio-economic barriers (such as low income, no caregiver, etc.). We say “potentially”, because there may be other factors at play that may not be apparent in the data, such as the availability of income from another source (such as a son or daughter), and the availability of a caregiver that doesn’t live in the household, and so forth. Regardless, if these types of member are segmented, we can script a specific communication to the member that focuses on these types of issues with an offer of assistance. This can make for a more engaging and successful conversation with the member.
Some factors in consumer data, by contrast, may be interesting and predictive of future behavior, but are not useful or actionable when segmenting and communicating with members. For example, political affiliation (Republican, Democrat, Independent), NASCAR interest, sports interests, etc. may be interesting, but are not useful or actionable in member communication and segmentation.
Non-Traditional Data – Survey Data
Another excellent source of data that’s underutilized is survey data. Traditionally, survey data has been used to find general trends in the population, but since survey data is generally not available for all members (only for members that responded to the survey), it hasn’t been used consistently to uncover member-specific social determinants of health.
Again, machine learning techniques can make survey data more actionable and usable. In short, machine learning techniques extrapolate the results of the survey to all members, regardless of whether they’ve actually completed the survey. These techniques attempt to identify members that exhibit similar behavior to members that provided an undesirable response in the survey.
Examples of where this can be useful is identifying members that had recent falls, have significant mental and/or physical decline, are dissatisfied with their doctor or plan, have access issues, etc. Pairing these predictions, with predictions on avoidable hospitalizations, or primary care engagement, can deliver significant value in understanding the member and delivering the right resources to improve population health.
Other Potential Non-Traditional Data Sources
FitBit and mobile health apps have become mainstream, and a new floodgate of patient data is about to burst open as mobile devices now saturate the market, putting health and wellness apps and trackers in the hands of every consumer. The health impact of these consumer devices continues to expand with incremental developments like the new blood pressure monitoring capabilities of the Apple Watch. This is just the beginning, as Apple, Google, Microsoft and others invested big money into consumer health in 2017, with ten of the largest U.S. tech companies spending $2.7 billion in healthcare equity deals, according to data from CB Insights research firm.
We’re already seeing a direct health impact, as employers sponsor corporate health challenges that encourage Fitbit-wearing employees to compete, logging the most distance on their lunchtime walks. Consumer devices are also creating a new potential channel of valuable data to payers. The data from these devices can unlock ever more granular insights into member behavior and its impact on their overall health.
This new feed of widespread consumer device data has great potential: it can enable health insurers to introduce personalized population health management programs that augment traditional patient data with real-time, individual health monitoring data to get a better picture of both individuals and populations. Also, the addition of consumer device data promises to unlock even more insights of what motivates member behavior and how to influence it, improving overall health and productivity while reducing absenteeism.
The question is, given the increasingly personal nature of information gathered through these devices, will we ever see a day where this type of information is available and used consistently for decision making?
Given the well-publicized abuses of private information, the key lies in how this data can be made available for use in population health management. Can the data be available in a “blinded” and aggregated format to provide healthcare organization with a semi-granular view of certain types of patients based on age, gender, neighborhood, etc.? For instance, would a 40-year-old man living in a specific Houston suburb be more likely to engage in some kind of aerobic exercise than his counterpart living in downtown Jacksonville? Using machine learning, this type of data can be married with census, consumer and traditional healthcare data to provide insights into specific individuals.
About the Author
Saeed Aminzadeh founded Decision Point Healthcare Solutions with the mission of improving health plan clinical, financial and operational performance through informed, data-driven predictions on strategic decisions. He has more than 25 years of health information technology experience, with a track record of building high-performing organizations designed to solve complex business problems. He has held key senior management positions at Eliza Corporation, Ingenix (currently Optum), IHCIS and ProVentive.