Atrial fibrillation (AF) is the most common cardiac arrhythmias worldwide. So large are the efforts to implement systems that support the diagnosis of AF. Thus, some authors have attempted to characterize the signal with AF through statistical lower order, which require assumptions of linearity and gaussianity [1], [2]. To work around these assumptions, other authors use methodologies that require more preprocessing steps, complicating the implementation of such methodologies in embedded systems [3], [4]. Therefore, we propose to characterize and classify the signal with AF using the kurtosis. The great difference of this method is that instead of considering the irregularity of the RR time series [5], we consider the irregularities of the FA signal itself. Our methodology is based on kurtosis to be a statistic of high order (Papoulis and Pillai, 2002), due to the characteristic of the ECG signal to be sparse. The proposed methodology also requires small number of preprocessing steps, allowing implementations in embedded systems. To evaluate the proposed methodology, we use signals obtained from the MIT-BIH bases of Atrial Fibrillation data and MIT-BIH Normal Sinus Rhythm referring to patients with AF and with normal heart rhythm, respectively. The results present a specificity and a sensitivity of 94.44% and 78.95%, respectively and accuracy of 86.49%.