On May 23, 2019, VitreosHealth and PatientBond produced the webinar entitled, Social Determinants and Predictive Analytics: Not Just Fancy Buzzwords. The webinar discussed how to use Social Determinants of Health and hyper-relevant patient/member segmentation for more effective care management and engagement.
Social Determinants of Health (SDoH) and predictive analytics enable a more accurate risk stratification, while psychographic segmentation drives personalization of message and channel mix to amplify desired behavior change across a patient/member population.
Webinar participants discovered:
- How to use Social Determinants of Health (SDoH) and predictive analytics to more accurately segment a patient/member population according to risk
- How to leverage each member’s propensity to engage to offer targeted programs
- What psychographic segmentation is and how it can be used to motivate and activate desired healthy behaviors among patients/members of a health plan.
Kirit Pandit, Co-founder and Chief Technology Officer of VitreosHealth, and Brent Walker, SVP Marketing & Analytics of PatientBond, delivered this 45-minute webinar with 15 minutes of Questions & Answers at the end. As a service to webinar attendees (live or On Demand posted on the PatientBond website), we list the questions and corresponding answers:
Q: What if we already have a member segmentation model – do we need to replace it with your psychographic segmentation model?
A: No, the PatientBond psychographic segmentation model can complement an existing segmentation model and inform our clients’ segment personae with a depth and breadth of further understanding. We constantly do market research to fill the trough of insights – we have more than 100 million data points now on our psychographic segments. We could work with a client to do a segment overlay and identify where, and to what degree, our respective segments overlap. Our clients can keep the segmentation model they have but use our insights to enhance their model.
VitreosHealth can also work with, and piggyback onto, a client’s segmentation strategy to identify and zero in on the “movers,” those who are moving from one level of risk to another, and train predictive models accordingly. Clients may have their own segmentation models (Level 1, 2, 3 or Band 1- 5). Using these, we can create customized predictive models (as opposed to national models), so we work with specific data sets and segmentation models that a client may already have.
Q: What if we have our own member engagement technology (e.g., email CRM system); can we still work with you on psychographics and predictive analytics?
A: Yes, we have an option called “Bring Your Own Technology” and we can work with you to integrate psychographic targeting and communications. The caveat is that the optimal messaging and channel mix and frequency will differ by 5 segment types, so the best results depends on a system’s ability to juggle these variables and adjust when necessary.
Q: Can you use our MCO Quality data to help determine risk factors?
A: Yes, one of the things VitreosHealth does in its predictive models is focus on Gaps in Care, and these are typically based on HEDIS. We do include Quality scores and Quality gaps as a feature set to train the models, and interestingly, depending on which population we look at, there is a different set of Gaps in Care that comes up as being most important in predicting a particular outcome. If a client already has precalculated Quality scores at a member level, that would be a great help.
Q: Do Social Determinants of Health help define the psychographic segments?
A: No, the psychographic segments are determined by a 12-question survey of people’s attitudes, beliefs and behaviors regarding health and wellness. For example, the first question of the survey is how much one agrees or disagrees with the statement, “I believe that I can directly influence how long I live, regardless of my family history.” There are no questions about access to care or financial or social challenges.
Regarding the demographic and socioeconomic characteristics among the psychographic segments, there is really no statistical difference among the segments in household income until you get to over $80,000 per year, which exceeds the national average household income. There’s a higher likelihood that people earning more than $80,000 are Self Achievers or Priority Jugglers, but there’s no statistical difference across the majority of people.
Also, while Willful Endurers and Direction Takers are more likely to only have a high school diploma, we’re only talking around 17-18% of these segments. There is no statistical difference among the segments in terms of the percentage who attended college.
Interestingly, Self Achievers and Willful Endurers – on opposite ends of the wellness and proactivity spectrum - are the most likely to live in urban environments and are overdeveloped among African Americans and Hispanics.
Q: What is the impact on outcomes after incorporating Social Determinants and predictive models?
A: We (VitreosHealth) have seen after the incorporation of Social Determinants – income, education, household composition… basic elements and some of the more advanced attributes if you can get them – depending on the population, can give us from a 5% to 7% boost in accuracy and recall rates. The more social disparity in the population, the more Social Determinants become important.
A recent example involves an analysis we did with a particular physician who wanted us to predict how many cardiovascular admissions his panel of 5,000 patients would have in the year. Without using Social Determinants, we predicted close to 60%, but then after we included Social Determinants this jumped up to 72%. In this case, we got a 12 percentage point boost. At a large population level, we usually see a 5% - 7% boost in accuracy.
Q: How has member engagement improved after incorporating psychographics?
A: From a member engagement perspective, we (PatientBond) do a lot of Test versus Control or A/B testing with our clients. We have them do what they would normally do in terms of engagement and execution for comparison, and we’ve seen increases of 6-7 times typical response rates for diabetes screenings and mammography and 5 times increases in colorectal cancer screenings for a national insurer. Many health systems and insurers are trying to get patients and members to use a wellness portal, and with one major insurer we saw a 17x increase in wellness portal registrations with 58% of those registrants completing a Health Risk Assessment (HRA). Using psychographic insights has been consistently beating Control group performance and we’re having a good impact.
Q: The medically underserved may not have access to technology such as smartphones or laptop computers — how do you reach and engage them?
A: PatientBond can detect whether a participant has a smartphone, flip-phone or landline and can adjust communications accordingly. In a situation where the recipient has no technology at all, we have relationships with print houses, if necessary. We’ve worked with several clients who address the underserved and Medicaid patients, and for the most part, we have not had an issue with that.
Q: Which nonclinical features are most impactful on a person’s predictive risk?
A: It depends on whether you are looking at urban or rural settings or if you’re looking at a much larger geographic area. What we see is income and education probably play the biggest role in terms of the Top 5 social attributes that you can identify with members. Household composition (presence of a caregiver) is also an important predictor. Zip code is highly correlated with education and income. Education and income together are correlated with things like housing/nutrition needs, number of vehicles, and interest in exercising. One of the things we have to be careful about is that we constantly do multi-collinearity tests to make sure we don’t capture redundant features on the Social Determinants side.
Q: Is health literacy factored into digital engagement/interactions?
A: The questions in the psychographic segmentation survey are written at a 6th-7th grade reading level and we tray as best we can to ensure communications and education are written appropriately. There may be instances where our clients must meet certain regulatory requirements and include boilerplate content that has been approved for use, but might not be the most consumer-friendly, but you can’t get around that. However, psychographic language, written to the appropriate literacy level, can act as a “wrapper” around that boilerplate copy to help position it properly.
That said, we have conducted focus group with inner-city and underserved healthcare consumers to understand whether our materials were clearly understood with no sensitivity issues, and they performed very well.
Q: AI (Artificial Intelligence) model are good at predicting and classifying things, but not really good at explaining “Why.” How would you address this problem?
A: As an AI practitioner, the ability to explain the model, or the “black box” as they call it, is becoming more and more important. There are a lot of AI practitioners both in academia and in the commercial space that are putting a lot of effort into this. The answer is multifold. If your predictive model is based on regressions or decision trees, then it is relatively easy to explain why someone is risky. Some of the deep learning or recurrent neural net-type models can get extremely complicated, and the way they work is they try to mimic the brain. So, trying to explain that type of model is like trying to explain how the brain works, which is very, very difficult. There are certain approaches that are very popular in being able to open that “black box.” One approach that we (VitreosHealth) and many others use is trying to fit a decision tree into the output of that model and let the decision tree explain it. Now, it may lead to a certain loss of granularity, so the resulting explanation is relatively simplistic and doesn’t capture all of the complexity that is inherent in that “black box.”
The other important thing is that the challenge is not just in identifying the Top 5 or Top 10 risk factors, what is also important is to understand the sensitivity to each risk factor. For example, cholesterol levels and A1C may both be risk factors, but one patient may be more sensitive to A1C, so reduction in A1C would cause the risk to reduce much faster than a reduction in cholesterol level, whereas for another patient it might be the opposite. There are models that can do a good job understanding this.
Q: How long does it take to set up your system and get it running? What is the typical “time to cost savings?”
A: PatientBond’s engagement technology is cloud-based and API driven and easily integrates with most EMR, CRM or practice management systems. If it’s off-the-shelf use cases, we’re only talking a matter of a few weeks to get it up and running; if it’s more customized it may take longer. A lot of that time is our clients reviewing the materials and approving them, so PatientBond follows a pretty efficient process. Depending on the use case, results can be almost immediate. You can see the difference in kept appointments or increases in screenings very quickly.
Important to keep in mind is, regardless of the population you are going after – the demographic or socioeconomic mix, where they’re located, how big it is – these five segments exist in every population, so having a system that is able to put together the right engagement mix quickly and efficiently will accelerate the time to revenue or cost savings.
With VireosHealth’s predictive modeling – the WHO, the WHY, the WHAT – we can get that end-to-end process in six weeks or so. A lot depends on how fast we can get access to the data sources, but in general it’s a 6 to 8-week process. One example is a client with whom VitreosHealth and PatientBond has worked together, and when we started doing the digital engagement, within a few months we saw the response rate climb up to 60% - 70% for a number of different initiatives and in less than 6 months we saw tremendous improvement in some of the Quality scores and outcomes.