Efficient Data Mining Techniques for Handling Neurodegenerative Disorder

The neuropsychological battery of scores and Functional Activities Questionnaire (FAQ) are significant to measure the cognitive and functional domain of the patients affected by Alzheimer's Disease. These scores collected from the National Alzheimer's Coordinating Centre are pre-processed to obtain instances with the minimum of four visit times. Thus, we obtain visit-1 to visit-4 data with 5542 instances after pre-processing. The feature selection (Genetic Algorithm) is applied to visit-1 and visit-4 data and evaluated the performance by using Logistic Regression Model before and after imputation and as well as before and after imbalanced data handling technique. The metrics considered for evaluating the performance of Genetic Algorithm (GA) are Area Under Curve (AUC), Mean Squared Error (MSE) and accuracy. The improved performance of GA is obtained for the neuropsychological scores after Miss Forest Imputation and for FAQ scores after subjecting it to Synthetic Minority Oversampling Technique (SMOTE). In the second work, data of visit-1 instances is normalized by z score normalization and these are further subjected to different optimization techniques. All the optimization techniques are evaluated by using Random Forest, Bagging, Naïve Bayes and Logistic Regression classifier models and the performance are evaluated by using Area Under Curve (AUC) as the metric. From the results obtained, we infer the Particle Swarm Optimization shows the better performance than the other optimization techniques. Later in the third work, we proposed a framework for binary and multiclass classification of Alzheimer's Disease (AD) using three-dimensional Structural Magnetic Resonance Images (sMRI) and clinical scores from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. The collected images are subjected to skull stripping, noise removal and tissue segmentation. The three-dimensional Gray Matter tissue, obtained as an output from tissue segmentation, comprises of many two-dimensional slices. But processing and training all the slices requires a lot of computational time. Therefore, our aim is to employ Convolutional Neural Network only on the significant slices and also to report the performance of the model. We propose a slice selection technique based on Shannon entropy to pick anatomically significant slices. Experimental results prove that the proposed slice selection classification framework achieves better performance when fused with clinical score. In the last work, we carried out our studies by collecting the images of Magnetic Resonance Imaging (MRI) from the Open Access Series of Imaging Studies (OASIS) database and categorized MRI images into three different groups based on the size of a ventricular region of the brain and then employed second and higher order statistical methods to extract the textural features from each image. Then the retrieval performance is compared for the extracted texture features from GLCM, Laws Texture Energy Measure and a combination of these two methods. The combination of features from the above methods shows the better precision of 80% and 60 % in the retrieval of Group1 and Group3 images.