Detection of Parkinson Decease Using Computational Intelligence Method

Detection of Parkinson D is ease Using Computational Intelligence Methods Elcin Huseyn 1 , Babek Guirimov 2 1 Research Laboratory of Intelligent Control and Decision Making Systems in Industry and Economics, Azerbaijan State Oil and Industry University, 20 Azadlig Ave., Baku, AZ1010, Azerbaijan, [emailprotected] asoiu.edu.az 2 Research Laboratory of Intelligent Control and Decision Making Systems in Industry and Economics, Azerbaijan State Oil and Industry University, 20 Azadlig Ave., Baku, AZ1010, Azerbaijan, [emailprotected] Abstract. Parkinson's disease is a neuro-degenerative movement disorder that causes voice/speech, and behavioral impairments. As a dysfunctional disease, it can be detected by a set of specific symptoms of patients. Such symptoms include both voice/speech and/or physical behavior/movement charac te ris - tics. For better detection both sets of characteristics are used in our research. In this study, as a diagnostic model, we use a system based on multiple-layer (deep) feed-forward neural networks. The networks are trained with Differential Evolution training algorithm using in parallel a pair of data sets (training and validation sets) to avoid overfitting and improve model's generalization ability (performance on untrained data). The applied DE algorithm has allowed avoiding local minima of error function during the training. A third data set is used for testing trained network performance. According to the obtained results, this method demonstrated better results than other existing approaches. Keywords: Parkinson's disease , Artificial Neural Network, Differential Evolution Optimization, Computational Intelligence Introduction Parkinson's disease is a neuro-degenerative movement disorder that causes voice/speech, and behavioral impairments . The disease causes partial or full loss in motor reflexes, speech, behavior, mental processing, and other vital functions [1] . The early detection of disease symptoms is vitally important in order to prevent further disease complications. Using recorded data including voice/speech and physical behavior /movement characteristics from healthy and sick people it is possible to create models, which would allow fast noninvasive diagnostic of the disease. Appropriate models include Support Vector Machines, Rule Based Systems, Artificial Neural Networks and others. Most existing approaches utilize only voice/speech data [2]. In our research to improve detection accuracy, we use also physical behavior/movement characteristics obtained from different subjects. Among all possible methods to create required model, we have chosen multi-layer deep feed-forward neural networks for a number of reasons. First, because they are indeed universal approximators and can be used to reveal any complex relationships in large data sets . Second, because recent developments in the theory and technology have significantly increased efficiency of neural networks. For instance, increased processing power and parallel processing abilities of modern computers allow efficient use of n ew evolutionary training approaches to effectively battle such bottleneck of large multi-layer neural networks as time-consuming parameter adaptation . The global parameter search , which avoids local minima trapping, is now much faster than ever . Third, because, neuron models are not now required to be constrained by smooth differentiable transfer functions, connection weights by simple numerical values, and network arch itecture for large input/output systems by single hidden layer of neurons. Method The used detection model is multi-layer feed-forward neural network with non-linear transfer function based neurons in hidden layers and linear neurons in input and output layers. Given particular values for the neural network parameters, and given values for the inputs, a neural network generates a value for each output: , The operation of an L -layer feed-forward perceptron neural network at each layer l can be described by the following equation: , where is the activation function used at network layer l . In the vector form this can be written more compactly: Or, based on only the network activations as: Matrix will denote weights connecting all neurons of layer with all neurons of layer . Thus for an -layered NN set will contain matrixes . is the weight of connection to neuron at layer from neuron at the previous layer , is the threshold parameter of neuron at layer The total number of connection weights and thresholds (i.e. number of elements in the set W ) for a feed-forward neural network is . The evolutionary algorithm used for training is Differential Evolution [ 5 ] , which is one of the fastest population based algorithms for global search in multi-dimensional v ector space.