OBJECTIVES: To describe a predictive modeling technique that was developed at Blue Cross Blue Shield of Louisiana (BCBSLA) and designed to identify plan members who were likely to have a future high cost event (an event that costs $25,000 or more).

METHODS: BCBSLA members from January 1, 2015 through December 31, 2016 were eligible for study inclusion. Data was parsed into development and validation data sets. The outcome variable was binary and defined as members whose annual medical and pharmacy expenses were either (1) $25,000 and higher or (2) less than $25,000. Predictive models were generated for both current low to future high and current high to future high cost members. Model inputs included sociodemographic characteristics, health status and comorbidities, high cost disease related episode indicators (congestive heart failure, diabetes, etc), neighborhood socioeconomic characteristics, healthcare market characteristics, and previous healthcare utilization. All analytics were performed using SAS Enterprise miner version 13.2. A total impact score was generated from the best fitting model. The validation data set was then used to test the performance of the impact score.

RESULTS: A total of 863,323 members met the inclusion criteria. The highest cost members represented 3% of all members and accounted for 40% of the total annual expenses. Model results suggest that a substantial proportion of healthcare expenditures incurred by the future high cost members are preventable/impactable, such as reducing ambulatory emergency department visits, hospital readmissions and conditions that might have been avoided through access to quality of care.

CONCLUSIONS: Predictive modeling and machine learning are excellent tools in the payer-provider-patient realm. By having the ability to identify members in advance, case management and disease management programs can potentially mitigate the clinical event, save lives, and reduce costs.