The study focuses on the classification of ventricular arrhythmias through classical machine learning and neural network methodologies. Arrhythmias originating in the ventricles of the heart represent a significant threat to life, necessitating prompt diagnosis and intervention. In a clinical environment, the rapid identification of precursors to potentially dangerous rhythm disturbances is crucial. This urgency underscores the importance of analyzing short electrocardiogram segments, which can facilitate timely and accurate detection of these arrhythmias. Using the MIT-BIH Malignant Ventricular Ectopy Database, which comprises 22 recordings featuring various arrhythmia types, the data was segmented into 2-second intervals and annotated. Six different classes of arrhythmias were identified. To tackle the classification challenge, several algorithms were employed, including k-nearest neighbors, support vector machines, and deep neural networks. The neural network architecture was optimized to improve the quality of arrhythmia recognition. In particular, bidirectional layers were used for LSTM networks, which improved the quality of the model. The study was conducted in multiple phases, encompassing both binary classification of arrhythmias into dangerous and non-dangerous categories, as well as experiments involving various class groupings. The results demonstrated that the application of a smoothed electrocardiogram spectrum significantly enhances both sensitivity and classification accuracy.