Population risk stratification for health systems via accretive predictive modeling

ObjectiveHealth systems rely on multiple approaches for population-level risk stratification/management. However, they can under-represent members with rising risk and complex treatment needs. To address these gaps and broaden the coverage of members at risk, we present an accretive framework of six predictive models across complementary risk measures for population-level stratification/management.Materials and MethodsLogistic regression models were trained/tested for six outcomes across cost (rising and elevated cost), utilization (rising and elevated utilization), and chronic-disease related (multimorbidities and polypharmacy) risk measures in 2016 using claims-based features from 2015 for ∼8.97 million members in a nation-wide administrative claims database. Model performances were validated against a holdout cohort of ∼2.99 million members. The presence/absence of each outcome prediction for members was summed into an accretive predictive risk index (aPRI) for population-level risk stratification evaluation.ResultsIntegrating predictions from the six models enabled member stratification across risk measures including future costs, utilizations, and comorbidities. Each of the risk predictions is represented in aPRI levels 0– 6, and their underlying model probabilities/risk measures increase with increasing aPRI levels. ∼83% of members grouped into a “low risk” (aPRI = 0) or “rising risk” category (aPRI = 1 - 2) and ∼17% into a “high risk” (aPRI = 3 - 6) category. Overlap/correlation analyses of risk predictions and comparison of their drivers further support the complementarity of predictions within aPRI and its enhanced coverage of members at risk.DiscussionBy integrating targeted and complementary risk predictions, aPRI enhances current population-level risk stratification approaches.ConclusionWe have developed an accretive predictive modeling framework for enhanced population-level risk stratification/management.