Electrocardiography is the method of choice for cardiac electrophysiological evaluation. Arrhythmia is one of the most crucial problems in cardiology. So far, many methods have been developed for arrhythmia detection, recognition and classification. A popular method is ECG modeling using a basis function (such as wavelet, hermite or RBF) and classifying the coefficients of the basis functions. We present a new method based on non-uniform sampling (selecting 7 samples) of ECG signal. It is shown that Left and Right Bundle Branch block arrhythmia as well as normal signals can be better analyzed using a newly introduced method called Finite Rate of Innovation (FRI) and other types of arrhythmia can be better analyzed based on spline modeling. Therefore, a multi-stage algorithm is proposed for diagnosing and compression of ECG signal which is faster and yet accurate.