Arrhythmias that can be captured on the electrocardiogram signal, correspond to creation heart diseases. Therefore detection of arrhythmias is an important issue for the diagnosis and treatment follow-up of such diseases. Artificial Neural Networks (ANN) is a commonly used tool for detection and classification of arrhythmias. There exist plenty of studies in the literature, concerning ANN based arrhythmia detection. However, the suggested ANN architectures are quite complex, involving hundreds of inputs and many hidden layers. Consequently, these models with high computational complexity can only be run offline, on relatively expensive processors. This approach is not practical for emergency and intensive care applications which demand high speed online processing and decision making capability. In this study, we aimed at reducing the complexity of the ANN model, while keeping the classification error rate at an acceptable level (e.g., 2-3 %). We used principal component analysis to decrease the number of inputs. The results of the trials that we have carried on MIT BIH data indicate that an ANN with 5 inputs a single hidden layer with 4 neurons is sufficient. This optimal ANN architecture can easily be implemented on a field programmable gate array which offers advantages such as high speed, low power consumption, and suitability for mobile applications.