Abstracts

Rationale: UCB and IBM have established a collaboration to explore the application of machine learning to a large claims database to construct an algorithm to aid antiepileptic drug (AED) choice for an individual patient based on his or her similarity to patients analyzed within the model. Proof of concept analysis supported model validity and here we report on the relative allocation to the model-predicted treatment vs another AED and the impact of mechanism of action (MOA, as defined by US Pharmacopeial conventions 5.0) on the success/failure of various combination therapy regimens. Methods: Medical, pharmacy and hospital data were collected by IMS between Jan 2006 and Sept 2011 for epilepsy patients >16 years of age having ≥1 pharmacy-related record with 80% stability over the entire period and had ≥2 years with 80% continuous eligibility and quarterly eligibility in prescription claims. An index date was set as the earliest valid date of a treatment change (add, new or switch) event. Claims were evaluated for treatment change or epilepsy-related hospitalization after the index date as an indicator of failure of the index date regimen. Treatment change was identified as a key outcome measure due to its propensity to reflect a myriad of potential challenges which could not be specifically measured such as poor efficacy, poor tolerability, lack of compliance, drug fulfillment, etc. Eighty percent of the claims were used to train the model and the remaining 20% were held out to test model validity. Results: The treatment change model supported the proof of concept with ROC curves demonstrating 0.73% accuracy (Figure A). Patients that were given the model-predicted treatment with the greatest chance of success had significantly longer survival rates (time until a failure event) than those who received another treatment (Figure B). The AED predicted by the system as most likely to be successful matched the actual treatment given to the patient in only 17% of the cases when the outcome was defined by treatment change. For patients receiving combinations of AEDs, there was significantly less chance of treatment failure if the AEDs had different MOAs. Conclusions: There is an improved chance of treatment success if given the model-predicted treatment (new AED or add-on/switch). However, 83% of the patients in the test data received treatment that was not predicted by the treatment change model to be the most successful. The model also suggests that combinations of AEDs with different MOAs result in greater success ratios or better treatment outcomes. Using the model's prediction system has the potential to deliver significantly better health outcomes for patients and health savings by applying resources more efficiently and accelerating the match between patients and their ideal therapy. As the project evolves, our goal will be to strengthen the predictive power by integrating diverse data sets and potentially moving to prospective data collection. Funded by UCB Inc. Disclosure information: CD, CC, and PF are employees of UCB Inc. MR-Z, JH and YG are employees of IBM. OD has no disclosures to report.