Atrial fibrillation (AF) is the most common cardiac arrhythmia and is associated with an increased risk of stroke and mortality. Research into the prediction of AF onset has been motivated by the necessity to develop better pacing therapy to reduce the incidence of AF and maintain the heart's normal sinus rhythm. In this paper, we address a similar problem to that posed in the Computers in Cardiology Challenge 2001, but develop patient specific models to distinguish between epochs of ECG located far away from AF rhythms and those located just prior to the onset of those episodes. Our approach is validated using a publicly available dataset.