PREDICTIVE META-MODELING TO QUANTIFY FUTURE HEALTH RISK AND IDENTIFY INDIVIDUALS FOR CLINICAL PROGRAMS AIMED AT IMPROVING HEALTH OUTCOMES AND QUALITY OF CARE

Purpose: Predictive modeling plays an important role at health plans to identify the right individuals for clinical programs aimed at improving health outcomes and quality of care. These models have typically relied on administrative claims data, so there is often a lag in identifying individuals who would benefit most from these programs. Therefore, we sought to integrate a comprehensive set of data sources beyond claims and utilized predictive meta-modeling to more quickly identify high-risk individuals for clinical programs.

Method: Extensive datasets, for example of more than 2 million members in the Medicare population, were leveraged for model building.  A broad set of heterogeneous data sources, including health risk assessments, prior authorizations, consumer data, clinical and pharmacy data, was assembled to form comprehensive risk profiles.  Then, multiple mathematical and statistical functions were applied to create the most optimized and predictive population segments and variables.  Meta-modeling approaches were then used to build a customized model for each population segment such as Commercial or Medicare, new or returning members, and individuals with certain chronic conditions. Each customized model comprised a set of learning algorithms that isolated the most predictive linear and non-linear patterns between individuals’ risk factors and their future health risk.

Result: A meta-model runs more efficiently as each customized model handles only one population segment and the most optimized input variables and algorithms for that segment.  For the new-to-Medicare segment model, which was implemented in early 2013, many high-risk individuals were identified earlier compared to the number of referrals using the predictive model in 2012 (27,200 vs. 1,800 referrals by March; 50,200 vs. 9,400 referrals by June). For individuals newly signed up for health insurance via the Affordable Care Act Health Insurance Exchanges, a model for this segment was launched in early 2014 and has already identified more than 12,500 individuals for referrals to a clinical program. The meta-model has a receiver-operating characteristic (ROC) index of 0.78 suggesting both high sensitivity and specificity.

Conclusion: Predicting health risk is challenging and depends on many factors including clinical conditions, demographics, quality, and access to care. Simply relying on historical claims is not sufficient; therefore, a meta-model utilizing high-dimension data from multiple sources and population segmentation will more effectively identify individuals for clinical programs in a health care setting.