Major Depressive Disorder (MDD), is a malady which has perturbed many around the globe. It has become a gigantic health concern for this world and economic burden globally. Traditionally, the clinicians use plethora of treatments to slow down the progression of this disease at early stage. This paper intent to better diagnose the depression; by transforming the Electroencephalogram (EEG) signals while using machine learning and deep learning technique. EEGs of 30 healthy and 34 Major Depressive Disorder (MDD) subjects were analyzed in this study. New features were generated from the preprocessed data and later various classifiers were used to predict the results. The finest accuracy of machine learning algorithm was acquired by K-nearest neighbors (KNN) which were 0.997; Decision tree (DT) acquired an accuracy of 0.984, Support vector machine (SVM) showed 0.957 accuracy while Naïve Bayes (NB) had an accuracy of 0.522. The proposed Deep Learning method which was convolutional neural network (CNN) resulted in the accuracy of 0.996. With these promising results this study proves a feasibility of EEG based Major Depressive Disorder diagnosis.
The automated classification invaded the world of health, but some diseases are still diagnosed manually as is the case of medulloblastoma (MB) cancer, which is a high-risk malignant tumor in the Central Nervous System (CNS). This tumor has many types and is diagnosed by biopsy when examining the histological images, which takes a lot of effort and time. In this paper, we propose to perform an automated Childhood Medulloblastoma classification based on Convolutional Neural Networks (CNN) by exploiting the knowledge learned through different datasets (medical PatchCamelyon (PCam) and ImageNet) with transfer learning. A comparative study is carried out by following two strategies: deep feature extractor and partial fine-tuning, we applied the most popular architectures namely VGG-16, VGG-19, ResNet-50, Inception-V3. Experiments prove that there is an important role of similarity when dealing with deep features extractor or partial fine-tuning. Indeed, when there is a high similarity between the scope application, the deep features extractor is the best, otherwise, partial fine-tuning is more suitable. In our case application, results demonstrate that the process of partial fine-tuning using Inception-V3 pre-trained on ImageNet achieved the best result with an accuracy of 99.16
Handwriting has gained immense popularity amongst the relevant research community as potential biomarker for early detection of Parkinson’s disease (PD). Several online and offline handwriting attributes have been proposed to characterize the presence or absence of PD. Additionally, various tasks have also been designed to capture these features. Nonetheless, it is observed that features extracted from multiple tasks can result in prediction disparities. In this context, determining feature-task relevance can facilitate by selecting the most appropriate subset of features from each task to enhance the overall prediction of an automated system. Most techniques presented in the literature advocate the dominance of either offline or online features over the other. Due to the inherent challenges of directly combining both types of features, we propose a decision fusion based alternative that takes advantage of both online and offline features and improves the overall prediction. By employing a benchmark database PaHaW, we validate the effectiveness of our approach.
Islamophobic hate speech is the indiscriminate negative attitude and behavior towards Muslims and Islam. Speech indicating prejudice against Muslims has negatively impacted the perceptions of Islam. Online platforms like Twitter have carved out policies to stop users from promoting Islamophobic hate speech, however, such content still exists which causes problems for Muslim communities globally. Hence, it becomes pivotal to find solutions to eradicate such speech from social media platforms. This paper presents an effective methodology for Islamophobic hate speech identification in online tweets using deep learning techniques. The proposed technique relies on feature extraction using a one-dimensional Convolutional Neural Network and classification using Long Short-Term Memory network based classifier. The proposed technique is validated on a dataset comprising of 1290 pre-processed online tweets and an accuracy of more than 90
Feature Selection (FS) is an important topic in the domain of machine learning. Support Vector Machine (SVM) is one of the most popular ML models for classification tasks. Efficient feature selection may ensure enhanced classification accuracy. Although there are several feature selection algorithms in practice, they are either separately used, combinedly used with linear SVM or, used with kernel-based SVM. Additionally, there exists another problem in classification methods, which is the selection of correct kernel function. There seems to be as such no general rule for selecting a kernel that maximizes the model’s performance. To handle the issues together, we propose a model called non-linear kernel-free twin quadratic surface SVM with optimized feature selection (LTQSSVM-OFS) which can eventually tackle both feature selection and classification tasks efficiently. We are doing feature selection in a kernel-free way by applying optimization method on a Laplacian Twin Quadratic Surface SVM classifier. We have validated our proposed model using (i) min-max approach-based SVM without FS, (ii) linear SVM, and (iii) radial basis function-based SVM models applied on four datasets, namely ‘Star3642balanced’, ‘Diabetes’, ‘Health care: Heart attack possibility’, and ‘Blood Transfusion’. Experimental results reveal that the proposed model outperforms the other models. Additional worth noting benefit of our model is that it yields better results in case of working with fewer features.
