Abstract- in this paper, classification of the heart diseases using the heart rate variability signals was performed in order to discriminate between normal subjects and patients with low heart rate variability such as patients suffering from congestive heart failure (CHF) and myocardial infarction diseases. A multi- layer feed forward neural network was utilized. For each of the three groups under investigation, three different techniques were used to select the inputs to the proposed classifier. These techniques are time- domain methods, frequency-domain methods, and non- linear methods. Results have shown that using non- linear methods give high rates for classifying heart diseases. Classification rate reaches to 96.36%. In an attempt to improve the classification rate, data fusion at feature extraction level was adopted. A new feed forward neural network was designed. It gives an average classification rate of 98.18%. Keywords– Time-domain features, frequency-domain analysis, non-linear parameters, neural network, and data fusion. I. INTRODUCTION Heart rate variability (HRV) provides a non-invasive measurement of cardiovascular autonomic regulation. Specifically, it is a measurement of the interaction between sympathetic and parasympathetic activity in autonomic functioning [1]. Traditionally, HRV has been quantified by linear time- domain and frequency-domain measures [2]. Statistical time-domain measures can be calculated from a series of normal-to-normal (NN) intervals (that is all intervals between adjacent normal QRS complexes resulting from sinus node depolarization). These may be divided into two classes, (a) those derived from direct measurements of the NN intervals or heart rate signal, and (b) those derived from the differences between NN intervals [2]. These variables may be calculated using smaller segments of the recording period. Frequency–domain analysis provides for the separation of Para-sympathetic (high-frequency range) and sympathetic activity (low frequency range) [3]-[5]. Spectral analysis is the most popular technique used in the analysis of HRV signals. Spectral power in the high- frequency (HF) (0.15-0.4Hz) band reflects respiratory sinus arrhythmia (RSA) and thus cardiac vagal activity. Low-frequency (LF) (0.04-0.15Hz) power is related to baroreceptor control and is mediated by both vagal and sympathetic systems [2]. However, the cardiovascular system is as a non-linear system. The contribution of Makikallio and co-workers serves to give an overview of the various nonlinear measures which are available for the estimation of the risk for cardiac arrhythmias [6]. These include those that attempt to estimate the degree of complexity such as correlation dimension and Lyapunov exponents, the information content such entropy, or the fractal properties reflecting temporal scaling invariance on the basis of detrended fluctuation analysis [6]. Similarly, one measure that has found probably the widest range of application in biological and medical settings is the approximate entropy (ApEn) introduced by Pincus [7]. Pincus describe the concepts on which ApEn is based and presents its definition as a model-independent statistic which quantifies irregularity in time series data. Also, the application of scatter plots in RR interval time series allow autonomic differentiation between supraventricular and ventricular rhythms as well as other forms of arrhythmia [7]. This paper presents fusion at feature extraction level where features extracted from time-domain methods, frequency domain methods, non-linear methods were combined and a new feature vector was formed. II. DATA COLLECTION The heart rate variability (HRV) data used here were extracted from from the PhysioBank Interbeat (RR) Interval Databases extracted from website of RR database [8]. Three groups were chosen which are Normal Sinus Rhythm (NSR), Congestive heart failure (CHF), Myocardial Infarction pre-medication (MI). For normal sinus rhythm, 72 beat annotation files for long-term ECG recordings (35 men, aged 26 to 76, and 37 women, aged 20 to 73) were used. For congestive hear failure, 48 beat annotation files for long-term ECG recordings of subjects aged 22 to 79. Subjects included 19 men and 6 women; gender is not known for the remaining 23 subjects. For myocardial infarction, 100 beat annotation files for long- term ECG recordings of patients with myocardial infraction pre-medication. Gender is not defined for this database. We perform traditional and new complexity methods on heart rate variability signals of length 1024 samples (8 seconds where sampling rate is 128 samples/second). III.TIME-DOMAIN METHODS Statistical time-domain measures were divided into two classes. These are: a. Direct measurements of NN intervals It includes two simple time domain variables that can be calculated these are: 1. Mean of all NN intervals (MNN). Data Fusion for Heart Diseases Classification Using Multi-Layer Feed Forward Neural Network Marwa Obayya, Fatma Abou-Chadi Department of Electronics and Communications Engineering, Mansoura University marwa_obayya @yahoo.com, f-abochadi@ieee.org 67 978-1-4244-2116-9/08/$25.00 ©2008 IEEE