Robust detection of CHF through new time–frequency features within HRV signals

Purpose The purpose of this study is to develop a high classification–rate detection approach of the congestive heart failure (CHF) through new time–frequency features within heart rate variability (HRV) signals. Methods The study of low–frequency (LF) and high–frequency (HF) bandwidths of the heart rate variability (HRV) signal is helpful in detecting abnormal patterns associated with improper functioning of the sympathetic and parasympathetic components of the autonomic nervous system (ANS). A better characterization of the cardiac activity may be achieved by extending the study of these bandwidths to the time–frequency plane. Indeed, this study uses an approach based on new TF features extracted from time–frequency image processing that further improves TF–based classification of non–stationary signals. Detection has been accomplished over the time–frequency plane through the Otsu image segmentation method. Classification of normal and congestive heart failure (CHF) HRV signals has been accomplished through a support vector machine (SVM). We defined 10 new time–frequency features which have been extracted from both LF and HF bandwidths from extended modified B–distributions (EMBD) of HRV signals, namely the TF–complexity measure (TFCM), the TF–energy concentration (TFEC), the TF–ratio between the area of the detected contours and the whole duration of HRV (TFRAD), the TF–ratio between the energy and the area of the detected contours (TFRAC), and the TF–ratio between the energy of the detected contours and sub–bands energy (TFREC). Feature selection has been carried out through the Laplacian Score (LS) and the Fisher method for comparative purposes. We processed 72 normal sinus rhythm (NSR), and 44 CHF HRV signals collected from 4 ECG databases of the PhysioNet research repository, namely Congestive Heart Failure RR–Interval Database (chf2db), BIDMC Congestive Heart Failure Database (chfdb), and normal subjects from the MIT–BIH Normal Sinus Rhythm Database and the Normal Sinus Rhythm RR–Interval Database. Results Our developed congestive heart failure (CHF) detection approach, which is based on new time–frequency features, outperforms previous studies in the literature. The classification was achieved at an accuracy of 95.65 %, a sensitivity of 100 %, and a specificity of 93.33 %, revealing that time–frequency features extracted with Otsu segmentation method combined with a SVM classifier are effective towards spotting CHF patterns within HRV signals. Conclusion High–frequency (HF) spectral bandwidth which is in relation to parasympathetic autonomic nervous system (ANS) activity is a powerful predictor in patients with CHF.