A large number of deaths are related to cardiac abnormalities. For early diagnosis usually a 12-lead Electrocardiography (ECG) is recorded and analysed. However, the electrode placement for all 12 leads is not always correct as well as there are not always 12-lead ECG systems available. Therefore, the aim of the Computing in Cardiology Challenge 2021 is to correctly classify and label 2, 3, 6 and 12-lead ECG recordings regarding 27 classes of cardiac abnormalities.
For that, a simple feed forward neural network with one input, one output and three fully connected layers for each class together with a one-vs-all classification scheme is developed. Dropout is used after each layer and various activation functions are compared. As a loss function the binary cross Entropy function is used. Features are obtained by an LSTM-based auto-encoder network as well as hand-crafted with emphasis on ECG timing and morphology. For classification, the most important features of both are chosen using a Random Forrest classifier’s feature importance analysis. 5-fold cross-validation is utilised to decide whether the neural network enhances the results of the Random Forrest classifier.
Preliminary results using only hand-crafted statistical features and a single multi-label classification network for all classes (three fully connected layers and Sigmoid activation as well as one Dropout layer) scored -0.373, -0.262, -0.353 and -0.338 for each lead respectively on the Challenge system (team name = AADAConglomerate) which is consistent with the 5-fold cross-validation score (0.441 for each lead) at our system. However, the Random Forrest model achieved a score of 0.129 for 12 leads and unfortunately could not be tested in time on the Challenge system due to technical difficulties.