TY - GEN
T1 - Feature Extraction and Classification of Heart Sounds Signals Based on Time-Dependent Entropy and Spectral Entropy Estimation
AU - Rios-Prado, Rosario
AU - Anzueto-Rios, Alvaro
AU - Tovar-Corona, Blanca
N1 - Publisher Copyright:
© 2020 Creative Commons; the authors hold their copyright.
PY - 2020/9/13
Y1 - 2020/9/13
N2 - In this paper, two entropy methods based on Shannon Entropy are exploited, the Time-Dependent Entropy and the Spectral Entropy, calculated in a time domain and frequency domain, respectively. The two calculated entropies together with the Probability Distribution were obtained from a database that contains simultaneous recordings from the four main auscultation areas. These areas are used to test if the probability of detecting the abnormality increases in any of the heart valves and to compare the results in each area respect to signals randomly selected from the database. The parameters obtained from 20 randomly selected signals of the data were used as input features for the K-Nearest Neighbour classifier, obtaining accuracies of 90% and 80% for pathologic and normal sounds classification, respectively. Finally, the features calculated from all the databases were separated and presented in each auscultation area in a 3D-graph where a visible separability is shown. Results suggest that some noise associated with valve dysfunction is reflected in the entropy values. Besides, results show that information in each area is different and the analysis of the four areas might improve the classification when there is a pathology.
AB - In this paper, two entropy methods based on Shannon Entropy are exploited, the Time-Dependent Entropy and the Spectral Entropy, calculated in a time domain and frequency domain, respectively. The two calculated entropies together with the Probability Distribution were obtained from a database that contains simultaneous recordings from the four main auscultation areas. These areas are used to test if the probability of detecting the abnormality increases in any of the heart valves and to compare the results in each area respect to signals randomly selected from the database. The parameters obtained from 20 randomly selected signals of the data were used as input features for the K-Nearest Neighbour classifier, obtaining accuracies of 90% and 80% for pathologic and normal sounds classification, respectively. Finally, the features calculated from all the databases were separated and presented in each auscultation area in a 3D-graph where a visible separability is shown. Results suggest that some noise associated with valve dysfunction is reflected in the entropy values. Besides, results show that information in each area is different and the analysis of the four areas might improve the classification when there is a pathology.
UR - http://www.scopus.com/inward/record.url?scp=85100935636&partnerID=8YFLogxK
U2 - 10.22489/CinC.2020.218
DO - 10.22489/CinC.2020.218
M3 - Contribución a la conferencia
AN - SCOPUS:85100935636
T3 - Computing in Cardiology
BT - 2020 Computing in Cardiology, CinC 2020
PB - IEEE Computer Society
T2 - 2020 Computing in Cardiology, CinC 2020
Y2 - 13 September 2020 through 16 September 2020
ER -