Cardiovascular diseases are the main causes of human mortality, with cardiac arrhythmia standing out among then, characterized by the absence of rhythm in a heartbeat. In order to prevent the number of patients with the disease from getting worse, regular heart rate monitoring is desirable. The diagnosis of arrhythmia is made especially using the electrocardiogram (ECG), which is responsible for recording the cardiac activities of an individual. For this, specialists must perform a visual analysis of the biomedical signals generated in the ECG recordings, rendering the diagnosis a long and exhaustive process. Thus, the automatic classification of cardiac arrhythmias can assist specialists in generating faster and more accurate medical reports. Deep learning techniques have been widely used in bioinformatics, since they have shown satisfactory results in pattern recognition and processing large datasets. Therefore, the present work proposes the development of an approach for the automatic classification of cardiac arrhythmias, considering ECG signals, as well as images constructed from these signals, through one-dimensional (1D) and two-dimensional (2D) convolutional neural networks (CNNs). As a result, 109, 446 ECG images were obtained, which were used to train a 2D CNN that achieved an average accuracy of \(82.86\%\) . In contrast, ECG signals were used as input to a CNN 1D, with architecture similar to the proposed CNN 2D, resulting in \(98.55\%\) of average accuracy and an index of 0.98 according to Youden statistics. In this context, the results of the present study indicate that the use of CNNs for the classification of cardiac arrhythmias can present satisfactory performance and can be used as an aid in obtaining medical diagnoses.