Study of hyperparameters in the semantic segmentation of heart sounds

Phonocardiograms (PCG) are widely used in the diagnosis of heart diseases. The segmentation of a PCG is a fundamental step in the elaboration of a diagnosis and in the design of automatic recognition systems for these diseases. Furthermore, it is a key step in other types of applications that require knowledge of the positions of heart sounds, such as source separation. In this work, the study of a Convolutional Neural Network (CNN) based on a semantic segmentation architecture is proposed. The incorporation of features used in other types of methods that address this problem is also proposed, to define the combination that offers the best results in conjunction with those used in the state of the art. This network is evaluated using a database of PCGs presented for the heart sound segmentation stage in the context of the 2016 PhysioNet/CinC challenge. In this work, an analysis of the different parameters is performed, such as the input features, length and number of filters per layer, network depth, length of the input signals and number of classes to be recognised. The results of k-fold cross-validation with $k = 10$ for the final network show an accuracy of 92.768 ± 1.190%, sensitivity of 92.734 ± 1.209%, precision of 92.806 ± 1.169% and a F-score of 92. 770 ± 1.189% for the test set.