OBJECTIVES: Apply data mining techniques to pharmacy data to identify patients likely to be non-adherent to their medication in the next six months. METHODS: Catalina Health™ receives a nationally representative sample of pharmacy data containing 40% of all U.S. retail prescription volume and 130 million unique patient ID’s.  The data is HIPAA compliant, longitudinal, and not projected.  Select 7 prescription medications representing multiple therapeutic classes.  For each medication, randomly select a 100K patient cohort filling a prescription between June and August 2010.  In order to eliminate patients who switch pharmacies or migrate to mail order, exclude patients who have no history of filling any drug during the prior 18 months, or who have no fill history for any drug during the six month analysis period.  Fit logistic regression models to predict which patients will be non-adherent to their medication in the next six months.  Consider patients non-adherent when Proportion of Days Covered (PDC)<80%.   Model covariates include prior adherence to the medication (if applicable), adherence for co-morbid conditions, days supply of the medication, ethnicity and income variables, patient age and gender, paid with cash, number of refills remaining, medication dose, and  number of co-morbidities.  Model accuracy is assessed using a 20% hold-out of the data.    RESULTS: The models identified 40.3% of the patients as likely to be non-adherent to their medication in the next six months.  This varied by medication (23.2% - 64.6%) and patient type (28.8% Experienced, 38.9% Moderate Experienced, and 77.5% New).  The overall model accuracy rate is 70.4%.  72.3% of the patients predicted to be non-adherent were actually non-adherent (precision), while 69.1% of the patients predicted to be adherent were actually adherent.       CONCLUSIONS: Data mining techniques  applied to pharmacy data can predict patients who are likely to be non-adherent to their medication in the next six months with 72.3% precision.