International Journal of Latest Trends in Engineering and Technology (IJLTET) ISSN: 2278-621X http://dx.doi.org/10.21172/1.72.581 530 Vol 7 issue 2 July 2016 First and Second Order Training Algorithms for Artificial Neural Networks to Detect the Cardiac State Sanjit K. Dash Department of ECE Raajdhani Engineering College, Bhubaneswar, Odisha, India G. Sasibhushana Rao Department of ECE College of Engineering, Andhra University, Visakhapatnam, Andhra Pradesh, India Abstract- In this paper two minimization methods for training feedforward networks with backpropagation are discussed. Feedforward network training is a special case of functional minimization, where no explicit model of the data is assumed. Due to the high dimensionality of the data, linearization of the training problem using orthogonal basis function is one of the options. The focus is functional minimization on a different basis. Two different methods, one based on local gradient and the other on Hessian matrix are discussed. MIT-BIH arrhythmia datasets are used to detect six different beats to know the cardiac state. The beats are Normal (N) beat, Left Bundle Branch Block (L) beat, Right Bundle Branch Block(R) beat, premature ventricular contraction (V) beat, paced (PA) beat and fusion of paced and normal (f) beat. Keywords – Backpropagation algorithm, Line search, conjugate gradient algorithm, ECG arrhythmia. I. INTRODUCTION For feed-forward neural network having differentiable activation functions, there exists a powerful and computationally efficient method[1]-[3]. This method is known as error backpropagation, finds the derivatives of an error function with respect to the weights and biases of the network. These derivatives play a central role in the majority of the training algorithms for multi-layer networks. The simplest backpropagation technique involves steepest descent. The backpropagation technique works in two stages. In the first stage the error function is propagated backward through the network to evaluate the derivatives such as Jacobian and Hessian matrices, whereas in a second stage the weight adjustment using the calculated derivatives and different optimization schemes is taken. This particular type of training of the neural network is known as performance learning, apart from other methods for training the network, such as associative learning and competitive learning. There are several optimization schemes other than simple steepest descent such as conjugate gradient, scaled conjugate gradients, Newton’s method, Quasi-Newton methods, Limited memory quasi-Newton methods and Levenberg-Marquardt algorithm. This work is limited to two optimization schemes, i.e. steepest descent and conjugate gradients. ECG gives the first hand information about the health of the heart. Deviation in the electrical conduction in the heart from the normal conduction, known as arrhythmia, is reflected in the ECG. Automatic arrhythmia detection is necessary as manual verification of ECG for long period is a tedious job. In the past a number of authors have developed different techniques to detect arrhythmias. In [4] hermitian bias function and K-nearest neighbor(Knn) is used for classification of arrhythmias. In[5]the authors used principal component analysis(PCA) in the hybrid multilayered perceptron network (HMLP).Dual tree complex wavelet transform(DTCWT)is used in[6] by Thomas, Das & Ari to classify five different types of ECG beats.Fuzzy classifier is used in [7]by the authors in the first stage with an accuracy of 93.34% and then improved to 98.64% in the second stage by applying genetic algorithm. In [8] PCA is applied to the statistical feature extracted from the spectral correlation of ECG data and then support vector machine(SVM) is applied to classify the five different ECG beats with an accuracy of 98.60%. In [9]the accuracy of ECG classification is 96.2%± 3.4% using SVM. In [10]] Yu and Chou applied Independent Component