Detection of neurodegenerative diseases using machine learning-based algorithms

Neurodegenerative diseases are characterized by the death of nerve cells or the deterioration of their structure and function over time. Parkinson's disease, Amyotrophic lateral sclerosis (ALS), and Huntington disease are the most common neurodegenerative diseases. There is no definitive treatment for neurodegenerative diseases. However, there are some treatment methods that slow down the progression of the disease and allow the patient to lead a better quality of life. Early diagnosis of neurodegenerative diseases is of great importance in this respect. One of the earliest and most common symptoms of these diseases is gait disorders. For this reason, this thesis proposes three different pattern classification models to detect neurodegenerative diseases using vertical ground reaction force (VGRF) gait signals obtained through sensors worn on the soles of the feet. For this purpose, an open access dataset of VGRF signals from neurodegenerative diseases and healthy subjects was used. The dataset consists of four different classes: amyotrophic lateral sclerosis (ALS), Parkinson's disease, Huntington's disease, and control subjects. Three different time-frequency analysis methods, namely maximum overlap discrete wavelet transform, variational mode decomposition and continuous wavelet transform, were used to detect these four classes. In the first approach, the VGRF signals were decomposed into fifth-order sub-bands using the maximum overlap discrete wavelet transform (MODWT). The statistical and spectral features of these sub-bands were then extracted to create a feature set of 57 features. The second approach used variational mode decomposition (VMD), a data-adaptive feature extraction method, to decompose the VGRF signals into five intrinsic mode functions (IMF). Statistical and spectral features were extracted from these five IMF signals, resulting in a second feature set of 45 features. The third approach involved creating a third feature set consisting of 23 features. This was done by applying a texture operator called the local binary pattern (LBP) to gait signals transformed into images using the continuous wavelet transform (CWT). KNN and SVM classifiers classified the resulting feature sets separately, achieving a maximum performance of 96.8% for each feature set. Then, to reduce the size of the dataset and save time and memory, the neighbor component analysis (NCA) feature selection algorithm was used to select features for each of the three approaches separately. The feature vectors were divided into 60% training data and 40% test data. The best features were then determined using the training data on the KBA side. Then, the best features identified in the previous step were used to test each model with SVM and KNN using 40% of the test data. For the right foot and right+left foot, 100% performance was recorded in the classification with the test data using SVM and KNN. For the left foot, however, the KNN classification achieved a performance of 96%, while the SVM classification achieved 100%. The performance results demonstrate that the feature extraction methods can accurately detect neurodegenerative diseases based on gait disturbance recordings. Keywords: Machine Learning, Maximum Overlap Discrete Wavelet Transform, Neurodegenerative Diseases, Local Binary Pattern, Variational Mode Decomposition