A classification model to predict specialty drug use

ObjectivePredicting who is likely to become utilizers of specialty drugs allows care managers to have an early intervention and payers to have financial preparation for the upcoming spending. Our administrative claims-based predictive model is to predict the members who might use specialty drugs. Materials and MethodsA national database* and a commercial health plan claim data were used to select a total 6.5 million people who were not taking any specialty drugs before the Target Prediction Window. There were about 136,700 members who were older than 65 in the study. We extracted 81 features from past history of medical, pharmacy claims, and demographic data to predict the specialty drug use in the following year. Members having at least three-month continuous enrollment either under medical or pharmacy plan in the previous year immediately before the start of the target prediction window and with no specialty drug taking history were eligible for this study. We trained and tuned on 75% of the data using an extreme gradient boosting binary classifier. We used the remaining 25% of the data to predict the outcomes and evaluate the performance. We also recorded the performance for the age group older than 65 years old. ResultsThere were 3% of members who used specialty drugs in the cohort under the current study. The important features for prediction included age, monthly pharmacy payment, monthly medical payment, diseases, procedure, or drug-related codes. On the test data with members of all ages, model performance for the area under the receiving operator characteristics curve (AUROC) was 78.6%. For the test set on members older than 65 (prevalence rate 3.6%), we had an AUROC of 79.3%. DiscussionThere is no similar machine learning model in the field to predict specialty drug use. Our model provides an unparalleled opportunity to allow early intervention for people who might develop diseases that require specialty drug use. It is also important for health plans and providers to know their covered population who might use specialty drugs and predict the increased cost in the next year. ConclusionA predictive model of specialty drug use can be helpful for both payers and providers to prepare for a spending spike or have an early intervention. In return, this helps to improve patients overall satisfaction.