Atrial fibrillation (atrial fibrillation, AF) is a common cardiac arrhythmia characterized by rapid and irregular heart rhythm, with the atrium and ventricle operating asynchronously. One of the consequences of atrial fibrillation can be the formation of a clot, consequently a stroke, and even death. That is why it is extremely important to recognize such a situation in a timely and correct manner. It is possible to diagnose this on an electrocardiogram (ECG) based on morphology, because in AF the so-called fibrillation waves (f-waves), but also on the basis of irregularities in heart rhythm, in which case various mathematical and statistical parameters are calculated and analysed based on the RR interval. One of the ways to detect the presence of atrial fibrillation is with the help of artificial neural networks, which is also the way to detect AF in this paper. Two different inputs were used as input to the artificial neural network, depending on whether the detection was based on morphology or heart rate. During the detection, i.e. classification, based on the morphology, the input to the neural network is represented by the values of the ECG signal between the two R teeth, i.e. the RR interval. In heart rate-based detection, the input to the neural network is calculated using mathematical and statistical parameters whose values change according to changes in heart rate. By learning artificial neural network, the network learns how to distinguish different signals, e.g., how to distinguish signals with a normal sinus rhythm from a signal in which atrial fibrillation is present. During the testing of the artificial neural network, the accuracy of the classification is checked with signals that were not used in the learning phase, i.e. on previously unseen data. By performing and implementing artificial neural networks, it can be concluded that the accuracy of detection is higher when signals are classified based on rhythm compared to morphology, also the accuracy is higher when classifying into a smaller number of classes. The accuracy of signal classification when using two classes (NSR and AF) based on morphology is 91.33% at the level of RR interval, and 95.02% on various patients, while the accuracy of classification based on rhythmic parameters is 96.03%.