This material is available only on bibliotequer
30 Oct 2019 
Viktoria 
A person with Parkinson\’s will progressively lose their physical abilities and become worse if there is no health care or appropriate solution. This disease occurs in all races and spreads in one to two individuals in a thousand. Its spread increases with age. It has been estimated that around 40% of people with this disease may not be diagnosed [LAN 14]. In recent years, the efforts to understand and characterize PD have intensified, a number of data mining and machine learning algorithms are developed to predict the early stage of Parkinson’s disease from biomedical data using voice, gait, and wearable sensors [LAN 14, MA 13, HOS 10, and DAL 13]. Voice signal recording is the earliest, easiest and non-invasive technique for diagnosis of PD [DUF 05].
As most people with PD sufferfrom speech disorders [HO 98], it could be considered a very reasonable way for detecting PD [SAP 07, RAH 07]. In many cases, data mining and machine learning algorithms are used on a very large scale of Parkinsons disease data [PEN 17, SAL 14]. This would create high computational complexity and lower efficiency. In order to overcome this problem, a range of feature selection algorithms are developed to identify the most significant features for predicting Parkinson’s disease [CHE 16, KOT 17].
The most recent attempts at diagnostic improvement consider the optimization of the feature vector of the speech data set as in Kaya et al. (2011) [KAY 11] and A. Tsanas et al. (2012) [TSA 12], the classification methods as in Sakar (2013) [SAK 13] and Das (2010) [DAS 10] or both of them as in (K. Shahsavari et al., 2016; Lahmiri and Shmuel, 2019) [SHA 16, LAH 19]. The results show that feature selection process is very important to improve diagnostic accuracy. Subsequently, this paper presents an attempt to improve the diagnosis of Parkinsons disease.
The main contributions of this work are proposing a new genetic algorithm based feature selection coupled with SVM in the Parkinsons diagnosis problem. In addition, we present a comparison of various types of feature selection algorithms.
Finally, the SVM-based RFE-CBR is a wrapper feature selection algorithm that uses criteria derived from the coefficients in original SVM models to assess features. It recursively removes features that are not informative. Compared to other wrapper techniques, SVM-based RFE-CBR does not use the precision of cross-validation on the training set as a selection criterion. As a result, it is less subject to overfitting and remains fast even though the original feature set is large [YAN 15]. The remainder of the paper is organized in the following manner. Section 2 contains related work. Section 3 introduces the methodology. Section 4 presents the data and the efficiency of our method is compared to various algorithms for selecting optimized characteristics. And finally the conclusion of the work.RELATED WORKUsing the speech samples for the diagnosis of PD has been the subject of several investigations.
For instance, Shahbaba et al. (2009) [SHA 09] used a non-linear model based on Dirichlet mixtures for the diagnosis of PD. An 87.7% classification accuracy was obtained with this method. Little et al. (2009) [LIT 09] conducted a remarkable study about PD identification, they employed a Support Vector Machine (SVM) classifier with Gaussian radial basis kernel functions to predict PD, and also performed feature selection to select the optimal subset of features from the whole feature space, and the best accuracy rate of 91.4% was obtained by the best model. Das (2010) [DAS 10] carried out a comparative study of artificial neural networks (ANN), DMneural, regression and decision trees for the diagnosis of the PD using speech samples.
The experimental results showed that the ANN method achieved a 92.9% general classification performance. Guo et al. (2010) [GUO 10] proposed a hybrid model based on expectation maximization (EM) and a genetic algorithm (GA), and obtained 93.1% classification accuracy. Ozcift and Gulten (2011) [OZC 11] combined the CFS (correlation based feature selection) algorithm with the rotation forest classifiers (RF) of 30 machine learning algorithms to identify the PD, and the best classification accuracy 87.13% was achieved by the proposed CFS-RF system. Chen et al. (2013) [CHE 13] used the feature reduction method to exclude redundant information from the original PD speech signal, embedded in the fuzzy classifier for PD diagnosis.
They achieved an average classification accuracy of 96.07%. In another study (2016) [CHE 16], the authors have also proposed using Extreme Learning Machine (ELM) and Extreme Kernel Learning Machine (KELM) for early diagnosis of PD. Experimental results showed that the proposed KELM method combined with the feature selection method provides very promising classification accuracy with a maximum accuracy of 96.47% and an average accuracy of 95.97% over 10-fold CV. More recently, Peker et al. (2015)[PEK 15] proposed to combine a maximum redundancy maximum relevance attribute selection algorithm with the complex-valued artificial neural network to detect PD; the classification accuracy of 98.12% was obtained by the proposed methodology.
Lahmiri et al (2019) [LAH 19] focused on evaluating the performance of eight pattern ranking techniques, including Battacharyya, GA, ROC, RFE-CBR, Wilcoxon, Entropy, t-test, and MI coupled to a nonlinear support vector (SVM) machine to distinguish patients with Parkinson\’s disease from healthy control subjects. The core parameters of the SVM classifier\’s kernel rbf were optimized using the Bayesian optimization technique.
The results obtained show that the classifier obtained the highest classification accuracy (92.21%) with 14 vocal features identified by the pattern ranking technique based on the Wilcoxon method. The highest specificity (82.79%) when formed with the first 13 significant voice models identified by the ROC based attribute selection technique. The highest sensitivity, 99.63%, with a single voice pattern under the ROC based attribute selection technique.3. PROPOSED FRAMEWORK FOR FEATURE SELECTIONOur goal is to propose an optimization function based on the support vector machine (SVM). This objective function is used in the GA for finding the most important features to get Parkinsons disease.