Novel complex-valued deep learning applied to automatic classification of heart sounds

Abstract Introduction Cardiovascular disease (CVD) is a problem facing governments around the world. Early detection saves healthcare system thousands of dollars by allowing prevention and avoiding costly treatments when has progressed to advanced stages. According World Health Organization (WHO), cardiovascular diseases represent leading cause mortality worldwide (WHO, 2019). That why, we have decided develop novel artificial intelligence models for heart disease, with accuracies that allow us put obtained algorithms into production. The revolution allows application Deep Learning techniques two-dimensional images in domain both real numbers complex value numbers, as classifiers sounds, normality or abnormality functioning heart. Purpose In present work, propose comparison 2D convolutional neural network (CNN) algorithm its counterpart automatic classification sounds normal abnormal. Material database use our research Pascal, which audio files cardiac activity, distributed 351 129 pathological sounds. Methods following steps were applied get objectives work: 1) segmentation single heartbeat, 2) conversion segmented sound associated image scalogram using Hilbert transform, 3) abnormal proposed algorithms, 4) measurement results performing two-tailed t-student hypothesis test cross-validation. Results We comparative table between two models, finding Accuracy, F1 Score, Precision Recall metrics complex-valued convolution networks significant improvements compared valued one. show numbers. For all cases, shows p-values less than 0.05%, giving statistical evidence means are significantly different models. Besides, performance Complex-valued model better Real-valued Conclusion advance learning, since they traditional counterparts based on This proposes an experimental basis construction new learning paradigm, where information another numerical domain, exploited help mathematical transforms. latter health sciences, demand higher, terms Funding Acknowledgement Type funding sources: Other. Main source(s): eVIDA Group