In this study, we propose a neural network model for the electrocardiogram (ECG) beat recognition. The description of the ECG signals consists of a multi-domain features which contain a set of meaningful and non redundant parameters. The construction of the system is accomplished by a data-driven learning scheme based on a clustering process to find an initial or coarse neuronal structure and a fine tuning hybrid learning algorithm, including gradient descent nonlinear optimization procedure and a least squares optimization step. The salient features of the system are an effective mechanism for variable learning rates and an adaptive metric norm for the distance. The results of experiments show the good efficiency of the proposed solution.