Identification of patients based on spectral analysis of heart rate variability using artificial neural networks

A new technique for identification of patients with congestive heart failure (CHF) from normal controls and patients with obstructive sleep apnea (OSA) from normal subjects is investigated in this chapter using spectral analysis and neural networks. The identification system consists of two parts: feature extraction part and classification part. The feature extraction part uses the method of approximate spectral density estimation of R-R-Intervals (RRI) data by implementing the soft decision sub-band decomposition technique to obtain the power spectral density of 6 different regions covering the whole spectrum. In the classification part, two different methods of machine learning approaches with neural networks are implemented and compared in their performances. Those approaches are: supervised neural network (back-propagation) and unsupervised neural network (Kohonen self organizing maps). The data used in this work for CHF identification is obtained from Massachusetts Institute of Technology (MIT) databases. A data set of 17 CHF and 53 normal subjects is used as original training data set, while another set of 12 CHF and 12 normal subjects is used as original test data set. A larger training data set, which is obtained by simulating 1000 CHF and 1000 normal subjects according to the spectral features obtained from the original training data, is used to train the neural network. The neural network is used then to test another simulated data set of the same size of the training date set (simulated according to the spectral features obtained from the original test data set). The accuracy of the classification is found to be about 83.65% and 91.43% with supervised neural networks and unsupervised neural networks respectively. The data used in this work for OSA identification is obtained from MIT databases also. A trial data set of 20 OSA and 10 normal subjects is used to obtain the classification features. A larger set of data simulated in a way similar to that done for CHF data is used to train the neural networks. The test phase is implemented on also a large set of data obtained from another MIT data set (test set) consisting of 20 OSA and 10 normal subjects. The accuracy of the classification approaches 84.42% and 98.22% with supervised and unsupervised neural networks respectively.