<p>The identification of abnormal cardiac beats represents the most significant indicators of heart ailments. This paper aimed to find a powerful classification system of ECG arrhythmias based on the conjoint use of the multilayer feed-forward neural network and the particle swarm optimization algorithm. In this work, five predominant categories of of heartbeats from MIT-BIH database are taken as desired output classes and divers kinds of features were computed and employed as inputs of the MLF neural network. First, we have proposed a new particle swarm optimization algorithm (MPSO) to update the weights and bias vectors for optimizing the classification performances. Then, we have added a new stage of features vector reduction based on the PSO algorithm to choose the most relevant features to considered categories of ECG beats for improving classification performances. We have computed a specificity of a 99.76%, a sensitivity of 99.63% and an accuracy of 99.74% with a very low classification error of 0.26%. The obtained results proved a very significant improvement of the multilayer feed forward network convergence ability using the MPSO learning algorithm. We have also demonstrated the dominance of the proposed classification system to distinguish between arrhythmias compared with other last published classification systems carried out on MIT-BIH database.</p>