Population Health: A Novel Technique for Predicting Healthcare Costs

sgruber-200 (1)By Sarianne Gruber
Twitter: @subtleimpact

Population Health Management strives for health providers to care for their patients with the overarching goal of improving quality outcomes, coupled with lower costs. Healthcare reform expects the inclusion of all associated costs of care and treatment for a comprehensive analysis as a means to reduce the per capita cost of healthcare. The know-how for attaining the “triple aim” depends on maintaining comprehensive data and applying sound analytics. Regardless if you are assessing health-related behaviors, treatments and/or hospitalizations, the recommended practice for tracking patient outcomes starts with selecting patient subgroups or cohorts such as disease type, treatment regimens or a determinant of high risk. Cohort studies have routinely and continue to be used by epidemiologists to study factors that affect the health and illness of populations over time.  The population health management research led by Dr. Aaron Wells from the Center for Health Research (Healthways, Inc, Franklin Tennessee)  piqued my interest with a coronary heart disease (CHD) study that parses the patient sample into cohorts defined by their prior health cost. Wells and his team have explored and tested a predictive modeling method called a Multidimensional Adaptive Prediction Process. The CHD patient population is first divided into “cost cohorts”; and then several models and covariates are created to optimize future cost prediction for individuals in each cohort. The complete study, titled “Predicting Health Care Cost Transitions Using a Multidimensional Adaptive Prediction Process”, was published in the August edition of Population Health Management.  Here are the key takeaways from the Multidimensional Adaptive Prediction Process (MAPP) study:

  • The method was tested on 3 years of administrative health care claims starting in 2009 for Healthways, Inc. health plan members with at least one diagnosis of Coronary Heart Disease recorded in the first two years (average n=25,143).
  • Administrative medical and pharmacy claims data was collected over a 36-month period (January 2009 through December 2011)
  • A control reference modeling methodology (the standard model) was used on the total annual population as a comparison versus the new model’s performance.
  • MAPP identified that members had contributed $7.9 million and $9.7 million more in 2011 health care costs than standard model in 2011 health costs for cohorts increasing in cost or remaining high cost, respectively.
  • Health Care Claims Expenditures were parsed into four cohorts with the following ranges: Low <=$5000, Medium $5000-$26,000, High $26,000 -$64,000, Very High >$64,000. Cohort thresholds were selected in order to try to achieve equivalent member and cost proportions between 2009 and 2010.
  • The MAPP method delivered a 21% improvement compared to the reference model with an annual difference of $1882 per member.
  • Results demonstrated that an improved future cost prediction is achievable using a novel adaptive method.
  • A major benefit of a MAPP technique is prospectively flagging members with a high likelihood for increased costs.
  • Predicting health care costs helps to efficiently allocate resources.
  • Health care consumer cohorts help to identify the need and quality of care for patients that may otherwise be unmet from analyzing one large diverse sample.

Data Scientists, statisticians and researchers use predictive models in a multitude of healthcare scenarios such as identifying high-risk patients, potential readmits or optimal treatment success.  In standard modeling practice, a single model is created with the intent to predict one outcome of interest (usually referred to as the dependent variable) such as risk of disease, estimate of cost or likelihood of readmission or survival. Wells et al, states the “single-model, single-covariate set approach to predictive modeling is unable to adapt to the multidimensionality of a population because one equation may not be suitable or accurate for all individuals.” Hence, using MAPP (modeling methods include parametric, nonparametric, linear and nonlinear and neural networks) models were estimated separately for individuals in each of the four defined cohorts based on cost levels. A final model was selected based the cost transition path that yielded the most accurate prediction of actual 2010 costs using 2009 data.  The article states that a key advantage of this methodology is that it “allows for more accurate prediction of individuals whose total health costs will escalate over time.”  Full article with details on the MAPP methodology, the cost-component cohort design and the control model comparison tests is accessible at The Center for Health Research, Healthways, Inc. Aaron Wells et al (Population Health Management 2015;18:290-299) DOI:1089/pop.2014.0087.