LOW DIMENSIONAL EMBEDDING SPACES REPRESENTATION FOR MOVEMENT DISORDERS DETECTION.

Abstract

Nowadays, there are no studies about the classification of patients with different diseases according to their gait, and this could be very interesting in the understanding of movement disorders and in the knowledge of the differences between patients movements. This is why in the thesis we have the objective of making a classifier by using Machine Learning, ML, methods to classify two types of subjects in a three dimensional embedding space according to their gait and their type of pathology. These methods used are divided into two types. On one hand there are those based on Manifold Learning: Principal Component Analysis (PCA), t-Stochastic Neighbor Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP). They have unsuper vised learning and they try to maintain the original structure and the most amount of information in the embedding space. On the other hand we use the Autoencoders, AEs, which are based on Neural Networks, NNs. There are Naïve Autoencoders (NAEs), the traditonal AE with a classifier added, Super vised Autoencoders (SAEs), trained in a supervised manner and making a division of groups of data according to their labels, and Variational Autoencoders (VAEs), with some supervised trained part. The data used are provided by the Movement Analysis, Biomechanics, Ergonomics, and Motor Control Laboratory, LAMBECOM from Spanish "Laboratorio de Análisis del Movimiento, Biomecánica, Ergonomía y Control Motor", and we make two different comparisons. The first one is between control (healthy people), and Multiple Sclerosis (ML) subjects, comparing when choosing the dominating leg of the control subjects and the more affected and less affected leg of the people with ML. And the second trial is between people with ML and people with stroke. The results are shown in different graphics. The conclusions are that these methods can clas sify subjects according to the relationship of the data with their labels. The best unsupervised method might be t-SNE as a good indicator of relation between data and labels, although UMAP also has similar results and it is faster than the previous mentioned. In the case of autoencoders, SAEs show a good classification of the two groups of patients. The methods prove a better dif ferentiation in the results of the comparison between people with stroke and people with ML, a good signal that with improvements in the algorithms and more testings with different patients, this classification could be better in the future, with the addition of the interpretation part of the results that explain the differences in a more detailed manner.