MODELING HIGH MEDICAL UTILIZATION

Tuesday, October 22, 2013
Key Ballroom Foyer (Hilton Baltimore)
Poster Board # P3-10
Quantitative Methods and Theoretical Developments (MET)

Scott Zasadil, Ph.D., Pamela Peele, Ph.D. and Henry Zeringue, PhD, UPMC Health Plan, Pittsburgh, PA
Purpose: Using medical claims data, we identify patients who have had 50 or more CPT procedures during the previous 12 months. The majority (over 70%) of these patients will utilize fewer services during the following 12 months. Identifying those patients who will have a continued high medical consumption will allow us to target them for enhanced care.

Method: We analyze patients that meet the criteria of: 1) Had 50 or more services during the index period and 2) possessed 12 month continuous enrollment in the Health Plan both during the index period and during the post-identification period. Patients having End Stage Renal Disease, Cancer or a Transplant were removed from this set because it is anticipated that they would already belong to the set of high service utilizers. This left a population of 44,632 meeting these criteria. After dividing this set into a training set and a test set, we developed an algorithm on the training set which predicts whether or not the patient will use more medical services in the following 12 months than they used during the previous 12 months. The patients who are predicted to be future high utilizers from this first stage are then clustered into 5 groups exhibiting different characteristics using a Self-Organizing Map (SOM).

Result: When the algorithm is applied to the hold-out, test set, we find that 45% of the predictions are true high utilizers and the model captures 48% of all high utilizers. One cluster stands out when the predictions are grouped into the 5 distinct SOM clusters. Patients in this cluster have a 40% chance of a hospital admission and a 70% chance of an ED visit during the 12 months following the index period. Discussions were held with our medical management department in order to confirm that these predictions can be acted upon.

Conclusion: We have developed an algorithm which is capable of identifying patients who will utilize a large number of services during the next year. This foreknowledge can be advantageously used to provide better and more coordinated care for these individuals.