Development of Efficient Algorithm(s) for Prediction of Neurodegenerative Diseases

Abstract

Neurodegenerative Diseases constitute debilitating and progressive disorders that primarily affect the neural functioning, leading to neuron loss and subsequent decline in cognitive and behavioral activities. The two prevalent diseases affecting the world's significant population falling in the above category are Parkinson's disease (PD) and Alzheimer's disease (AD). These problems burden healthcare systems and society. These diseases are diagnosed by machine learning techniques and demographic traits; however, they possess accuracy limitations. Therefore, this PhD thesis delves into classifying neurodegenerative diseases and their stages, addressing various challenges and opportunities. This research proposes efficient approaches for neurodegenerative disease prediction, catering to the demands of diagnosing the diseases at the earliest stage. This research encompasses a comprehensive review of existing techniques, algorithms, and approaches pertinent to neurodegenerative disease classification while identifying the gaps and limitations that must be addressed. The thesis firstly contributes to the formulation of a deep learning-based multilayered framework using transfer learned Alexnet and LSTM for binary as well as multiclass classification. As deep learning models necessitate a large training dataset to produce better outcomes, Generative Adversarial Network (GAN) are utilized as a data augmentation tool to enhance the classification results and further address the issue of overfitting. Furthermore, this thesis proposes an integrated model for binary and multiclass classification of another neurodegenerative disease, i.e., PD, using Capsule Networks (CapsNet). As CapsNet xiii requires an ample feature set to produce better results, therefore, to extract the features, three manual feature extraction techniques, along with autonomously generated deep learning automated features, are extracted. Further, the features mentioned above are selected and evaluated for irrelevancy and redundancy using the Boruta technique, wherein these selected features are then loaded into the hyper-tuned capsule network to classify and predict PD. In addition to the above models, this thesis also delves into the classification of aforementioned neurodegenerative diseases in a single model utilizing a ranking-based ensemble classifier utilizing weighted strategy of deep learning classifiers. The model's conclusions generalize the classification outcomes by demonstrating the competitiveness of the suggested deep learningbased ensemble technique for AD and PD prediction in multiclass and binary class classification. Further, real-world datasets of MR images from various sources are used in experimental evaluations to verify the effectiveness of the suggested approaches. The results show that the new approaches are superior to the current methods in terms of accuracy, sensitivity, specificity, precision, etc. In summary, by addressing the related issues, this PhD thesis advances the subject of classifying neurodegenerative disorders. The developed methodologies not only improve the lives of those impacted, but also lay the foundation for medical specialists to diagnose the diseases as soon as possible. Further, the research opens avenues for future work in other neurodegenerative diseases to diagnose these diseases accurately.