Original article was published on Artificial Intelligence on Medium
OPTIMIZATION REQUIRES UNDERSTANDING BOTH
RISK AND IMPACTABILITY
Over many decades, payers have developed advanced risk scoring models of various types. More recently, and mostly in response to their growing exposure to risk under VBC arrangements, many providers have built comparable risk stratification capabilities by partnering with payers, using third-party resources or creating their own in-house risk tools. They have naturally turned to these methods to help target various value-driving programs.
A practical example of such risk-based intervention targeting was executed by a large value-based care provider that had an ongoing outbound calling programs to encourage adherence for members who were in an unpleasant, extended infusion-based care program which served to forestall later much more serious and expensive medical problems. The provider built and deployed high-quality risk models to target these calls to the members who were most at risk of dropping out before completing the course of treatment. This risk-based targeting was fine as far as it went, and did improve program effectiveness materially.
But what this approach misses is a key dimension of the problem: the propensity of each patient to be impacted by the outbound contact (i.e., ‘impactability’).
For example, there are high risk patients who will not change behavior in response to even a well-timed and well-executed call.
The money spent calling them is entirely wasted. More perversely — but in our experience, always true for a non-trivial minority of patients for any given intervention — there may be patients who would have completed treatment, but drop out because we called them. In that case, we actually pay money to get a worse outcome. Typically, risk of non-adherence and impactability are very weakly correlated, and are in practice independent effects.
We can therefore represent the possibilities in the following matrix:
Ideally, we would only contact the ‘Persuadables,’ and prioritize within this group using our risk models. The most effective approach to do so is to:
- Estimate for each patient the risk (or more formally, the expected value) of non-adherence if not contacted;
- Estimate the probability of change in adherence if contacted for each patient;
3. Select the patients to call based upon maximum projected change in expected value of non-adherence caused by calling them.