Arrhythmia detection based on time–frequency features of heart rate variability and back-propagation neural network

This paper presents a novel method for detection of arrhythmia using adaptive continuous Morlet wavelet transform and back-propagation neural network. The detection is based on extracted features in Time–Frequency (T–F) domain from heart rate variability (HRV) signals. For this, HRV signal is segmented into small length (48 inter beat interval). The T–F analysis methods, namely, adaptive continuous Morlet wavelet transform (ACMWT), adaptive modified Stockwell transform (AMST), and adaptive Stockwell transform (AST), are employed for features’ extraction. The adaptations of these methods are established on energy concentration. The features such as Flatness, flux, Skewness, Kurtosis, Shannon entropy, Renyi entropy, and coefficient of variation are extracted by T–F analysis methods from segmented HRV signals. These features are applied to back-propagation neural network for training and validation of neural networks. The outputs of neural network are applied to three different decision rules as average, vote, and decision vote. The proposed method is validated using the standard database of arrhythmic subjects. Simulated results show that features extracted by ACMWT from HRV signals achieved accuracy (AC) of 95.98%, specificity (SP) of 96.24%, and sensitivity (SE) of 89.5% for low-frequency (LF) band, and AC of 97.13%, SP of 96.13% and SE of 94.38% for high-frequency (HF) band of HRV signal when using decision vote rule. The results achieved by ACMWT are better than AST and AMST.