The Multi-label Classification of 12 Channel ECG signals of Patients with Various Diseases

András Bánhalmi, Vilmos Bilicki, Tamás Szépe
University of Szeged


The aim of the Physionet Challenge in 2020 is to classify 12 channel ECG signals into the normal or 8-disease category. For this task we used a data processing chain which is based on state-of-the-art signal classification techniques. The classification task is performed with a deep neural network, which contains mostly convolutional layers and a bidirectional LSVM layer. To prepare data sets that are fed to the neural network, first we did some preprocessing on the original data, namely a band-pass filtering, a resampling, and a normalization for each channel. Then in the second step, a 5-second interval was selected from each ECG signal, whose duration was longer than 5 seconds. For this, a heuristic algorithm was developed that selects the part with the highest variance computed in the low frequency domain. When training and validating our neural network structures, the whole labeled dataset was divided randomly into training and testing parts (with a constant seed to make it reproducible). After computing some statistics on the whole dataset we concluded that the dataset was not balanced, so from the classes with fewer samples we placed more into the training set than those in the more frequent classes. In this way, we got similar results on our test set to those on the official test set. Our best model achieved scores of 0.59 Fβ and 0.39 Gβ, respectively, compared to test scores of 0.56 Fβ and 0.33 Gβ, respectively, measured by the Challenge Organizers.