Several methods for automatic heartbeat classification have been developed, but few efforts have been devoted to the recognition of the small ECG changes occurring in healthy people as a response to stimuli. Herein, we describe a procedure for the extraction, selection and classification of features summarizing morphological ECG changes. The proposed procedure is composed by the following stages: 1) extraction of a set of heartbeat morphological features; 2) selection of a subset of features; 3) subject normalization 4) classification. The selection of a subset of features enabled us to summarize ECG changes with only three non redundant features. In addition we performed a comparison between four classificators: k-nearest-neighbors (k-NN), artificial neural networks (ANN), support vector machines (SVM) and naive Bayes classifiers (nB). In order to cope with the possible non linear separation problem, we evaluated two strategies: a subject factor normalization on feature space and the usage of kernel functions for classifiers. The results of the comparison recommended the usage of subject normalization, irrespectively from the classificator: with or without normalization we had the best performance of classification for the linear-SVM and ANN
Comparative study of morphological ECG features classificators: An application on athletes undergone to acute physical stress
LAURINO, MARCO;Piarulli A.;GEMIGNANI, ANGELO;LANDI, ALBERTO;PIAGGI, PAOLO;Menicucci D.
2011-01-01
Abstract
Several methods for automatic heartbeat classification have been developed, but few efforts have been devoted to the recognition of the small ECG changes occurring in healthy people as a response to stimuli. Herein, we describe a procedure for the extraction, selection and classification of features summarizing morphological ECG changes. The proposed procedure is composed by the following stages: 1) extraction of a set of heartbeat morphological features; 2) selection of a subset of features; 3) subject normalization 4) classification. The selection of a subset of features enabled us to summarize ECG changes with only three non redundant features. In addition we performed a comparison between four classificators: k-nearest-neighbors (k-NN), artificial neural networks (ANN), support vector machines (SVM) and naive Bayes classifiers (nB). In order to cope with the possible non linear separation problem, we evaluated two strategies: a subject factor normalization on feature space and the usage of kernel functions for classifiers. The results of the comparison recommended the usage of subject normalization, irrespectively from the classificator: with or without normalization we had the best performance of classification for the linear-SVM and ANNI documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
