The study aims to review the mathematical methods developed for the prediction of atrial fibrillation by analysis of surface electrocardiographic records in paroxysmal or post-cardiosurgery patients. A risk stratification based on ECG analysis would be very useful either to optimise the prophylactic antiarrhythmic treatment in high risk patients, or to limit drugs administration in low risk subjects. The works published so far managed to achieve good results in terms of sensitivity and specificity. However, since these methods are not completely reliable yet, their clinical application is still limited. The present study is divided in sections about time domain, frequency domain, premature complexes detection, heart rate variability, and non linear ECG analysis based methods.