Atrial fibrillation is defined as a subtype of supraventricular tachycardia with uncoordinated electrical activity in the atrium. The direct consequences of atrial fibrillation on heart function are inefficient atrial contraction and rapid and irregular heart rhythm. During fibrillation, the atrium ceases to be the primary pump thus reducing cardiac pumping efficiency by 20%. However, it does not cause death as in ventricular fibrillation and a person can live for years with symptoms. The estimated number of patients diagnosed with atrial fibrillation worldwide in 2010 was 20.9 million men and 12.9 million women, and an annual exponential growth of five million new cases per year is projected. Age is a high risk of atrial fibrillation. People over the age of 40 have a 25% higher risk of developing atrial fibrillation, and it is more common in men than in women. As the quality-of-life standards increased, the life expectancy of the population also increased. With a higher proportion of the elderly population, the risk of atrial fibrillation increases, and thus the number of patients themselves. In addition to affecting quality of life due to irregular and rapid heart rhythms, atrial fibrillation poses a risk for other heart diseases, such as ischemic heart attack, stroke, thromboembolism, or cardiac arrest. It is present in 3 to 6% of patients hospitalized with an acute condition. Although a large number of patients have pre-existing concomitant heart disease, the clinical picture of atrial fibrillation can range from emergency hospitalization to patients who feel completely healthy or unaware of cardiac abnormalities. The consequence of atrial fibrillation is a declining mechanical function of left atrial contraction during which blood clots form which can lead to serious complications such as stroke or death. It is estimated that 20% of strokes are caused as a result of atrial fibrillation. The characteristics of atrial fibrillation in the ECG signal are absence of P-waves, occurrence of irregular fibrillation waves (f-waves), irregular electrical activity of the ventricles, and unaltered ventricular (QRS) complexes. In more recent literature, the authors used larger databases to classify atrial fibrillation and other heart diseases but noted as a drawback the limitation to single-channel ECG signals. The aim of this doctoral dissertation is to investigate the possibility of improving the detection of atrial fibrillation using a multichannel electrocardiogram. Improving the methods for detecting the QRS complex, i.e. the R wave, would allow more accurate extraction of atrial fibrillation features. The morphology of the ECG signal can be defined as a set of samples between two consecutive R spikes. Thus, it is possible to learn a model based on artificial neural networks precisely on the morphological features of a single cardiac cycle. The first chapter (Introduction) describes the beginnings of computer-aided electrocardiogram analysis. The use of computers in detecting atrial fibrillation is described and an overview of the limitations of known bases and approaches is given. The aim of the research of this paper and the original scientific contribution of the dissertation are presented. The second chapter (Electrophysiology of the heart and electrocardiography) describes the basic anatomy and the electrophysiological processes of the heart. Multi-channel electrocardiogram leads and ECG signal characteristics under healthy conditions are described. The third chapter (Atrial fibrillation) describes the epidemiology, classification, pathogenesis and approaches to the treatment of atrial fibrillation. Also, an overview of atrial fibrillation features obtained from ECG signals based on heart rate and signal morphology are given. The fourth chapter (ECG signal databases) describes the ECG databases that serve as input data for ECG processing and analysis, describing the databases for detecting QRS complexes: MIT-BIH database and QT database with annotations of QRS complex, and databases for training and testing of atrial fibrillation classification models: PhysioNet/CinC 2020 database publicly available from PhysioNet website, synthetic atrial fibrillation database, and Srčana database collected in cooperation with the Polyclinic for the prevention of cardiovascular diseases and rehabilitation. The fifth chapter (Methods for Detecting QRS Complexes) provides a detailed description of the methods used to detect QRS complexes in the ECG signal. The selected algorithms are described through several steps of ECG processing: removing noise and artifacts from ECG signals, ECG signal transformation to highlight QRS complexes, selecting candidates for QRS complexes, and setting criteria and deciding on candidates for QRS complex. The importance of combining processing steps from various QRS complex detection algorithms was highlighted. Finally, the method of weighted fusion of multiple ECG channels is presented and the procedure of evaluating the methods for detecting QRS complexes is described. The sixth chapter (Atrial Fibrillation Detection Model) describes a deep neural network model for atrial fibrillation classification. The procedure for selecting and preprocessing the available data and the majority voting procedure used in the analysis of the multi-channel ECG signal are described. The seventh chapter (Results and Discussion) presents the results of evaluating the accuracy of QRS complex detection and the results of the atrial fibrillation classification model. An overview of previous research and a comparison of the results of this research with the results from the literature is given. Common detection errors of QRS complexes in standard databases when using a weighted multi-channel fusion approach and the advantages and disadvantages of using multi-channel analysis are listed. The testing of the atrial fibrillation detection model was compared with other databases with the detection at the patient level. A discussion of the results was conducted. The eighth chapter (Conclusion) briefly positions this doctoral research in current state of scientific knowledge and emphasizes the scientific contribution together with the most important results of the conducted analysis. The original scientific contribution of this dissertation is the method for detecting QRS complexes using multichannel electrocardiogram for the purpose of extracting morphological features of atrial fibrillation, a model based on deep neural networks for classification of atrial fibrillation and normal sinus rhythm based on multichannel cardiac cycle input data, and validation on holter ECG signal database. Algorithms for detecting QRS complexes have been achieving high accuracy on standard test databases for many years. Nevertheless, new algorithms are published in recent literature with different approaches to the problem. As the literature sometimes lacks a description of the key stages of the algorithm without which the results cannot be reproduced, in this dissertation the algorithms are broken down into four phases. It was investigated how combinations of different phases of algorithms affect the overall accuracy. It has been shown that by combining the IMF filtering method and the Elgendi detection algorithm, the results improved. Improvement was also achieved by using weighted fusion of multiple ECG signal channels. Although the approach gives better results, a step of learning the algorithm is required, which makes its real-life application difficult. Furthermore, a model was realized for classification of atrial fibrillation and normal sinus rhythm. The classification model based on deep neural networks was trained with morphological features described by a single cardiac cycle. The single-channel approach achieved high classification accuracy. The results were improved by the multi-channel approach using the majority voting method. The classification model was evaluated on a publicly available clinical database of 12-channel ECG, a publicly available database of synthetically generated ECG signal records and a database of holter records collected as part of a study in collaboration with the Polyclinic for The Prevention of Cardiovascular Diseases and Rehabilitation Srčana in Zagreb, Croatia. In this research, it has been shown that the binary classification of atrial fibrillation and normal sinus rhythm achieved good results, but that the accuracy decreased with the presence of other diagnoses along with the normal sinus rhythm class. Papers in which high accuracy of classification of several heart diseases has been achieved are mainly studied on large databases and have more complex models with a larger number of parameters compared to the research conducted in this dissertation. One of the main objections to the use of machine learning methods in clinical practice is the "black box" nature of the methods. Current research does not provide insight into the mechanism of heart disease and the process of making a final decision and diagnosis, despite the high accuracy of classification. The computer model implies an optimally recorded signal, without technical or other errors, which, although easily noticeable by the medical staff, can lead the computer based model to the wrong track. Without an adequate explanation of a given diagnosis, it is difficult to detect model bias that is often conditioned by a set of data (e.g., racial model bias). This research shows the importance of database selection and recording methods in the final evaluation and assessment of model accuracy, and thus the importance of standardization of computer interpretation of electrocardiograms. Testing models when examples from the same database are not found in both the training set and the test set achieved lower overall accuracy of the classification. By testing on separate database from training set, the accuracy was 4.27% lower at the heart cycle classification level and 4.79% lower at the subject level. The model trained on a clinical database gave good results when tested on a synthetic database. However, a model learned from a synthetic atrial fibrillation database did not yield good results when tested in a clinical database indicating possible limitations of the synthetic database in generalizing morphological changes in ECG signals during atrial fibrillation episodes. It is also important to note the possibility of malicious changes or attacks when using such systems. Computer analysis of ECG signals is susceptible to such attacks and appropriate safety mechanisms need to be implemented. The number of heart diseases is on the rise, and atrial fibrillation is just one of them. Further research can focus on a larger number of diseases, and the improvement of a model that would more successfully differentiate both diseases from each other and healthy control groups from a group of patients.