Aim: Automatic abnormality detection of ECG signals is a challenging topic of great research and commercial interest. It can provide a cost-effective and accessible tool for early and accurate diagnosis, which has been shown to increase chances of successful treatment. The aim of this study is to construct an ensemble model for 12-lead ECG classification as part of the Physionet/Computing in Cardiology Challenge 2020. Methods: The training dataset consisted of 6877 patient records, with each record being 6 to 60 seconds long. For the preprocessing, the signals were filtered to remove baseline wander and high frequency noise, segmented to a length of 30 seconds, and resampled to a rate of 250Hz, to avoid redundant model complexity and reduce training time. The proposed ensemble classifier consisted of three base models: 1) a feedforward neural network, trained on a set of predefined higher-order statistical, spectral and signal processing features, and two deep convolutional neural networks (CNN) combined with bidirectional gated recurrent units (GRU), trained on 2) the processed signals, as well as on 3) extracted median beats from each lead. Model architectures and parameters were evaluated via a 10-fold cross validation process. Results: In Phase I of the challenge, our entry under the name AUTh Team achieved a cross validation Fb score of 78.0%, Gb score of 51.7% and geometric mean of 63.5%. On the hidden test we achieved an Fb score of 77.5%, a Gb score of 51.1% and a geometric mean of 62.9%. Conclusions: Preliminary work was mainly focused on machine learning algorithms, however there is significant margin for improvement in data engineering to address issues such as skewed classes and corrupt records. Different network architectures and modern ensemble boosting methods are also pursued, as further improvement on the performance is expected.