Machine learning classification of human gait disorders

• Autor(a) principal: Ventuzelos, João Pedro Dias
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/33934
• topic: Human gait disorders; Ground reaction forces; Time series classification; Multilayer perceptron; Convolutional neural networks

Computerized human gait analysis is commonly used by researchers and physicians to detect disorders, evaluate therapy progress, or improve athletic performance. Advances in instrument and measurement technology has allowed the quantification of human gait characteristics, such as kinematic and kinetic parameters, electromyographic activity and energy consumption. In particular, the quantification of ground reaction forces (GRFs) has proved to be an important tool in the healthcare context. However, the extraction of meaningful features and their interpretation from the amount of complex data is still a challenging task. Consequently, machine learning methods are becoming popular to deal with the high-dimensionality, temporal dependencies, strong variability, and non-linear relationships present in human gait data. This dissertation aims to study the application of machine learning techniques for the classification of human gait disorders, using the annotated GaitRec dataset. The dataset contains bi-lateral 3D-GRF data from healthy individuals, as well from patients with musculoskeletal impairments at the hip, knee, ankle and calcaneus. This work addresses the custom development of classification models capable of differentiating normal vs. abnormal gait patterns (binary problem), as well as classifying pathological gait disorders (multi-class problem). The focus is on the comparison between classical fully-connected models and 1D convolutional neural networks (CNNs), in terms of prediction accuracy. Additionally, pre-processed time series are converted into a two-dimensional input image, which is applied to a 2D-CNN to explore asymmetries in bilateral GRFs. The results obtained show that the fully-connected model outperforms in 1% the 1DCNN model. The binary classifier achieved a prediction accuracy around 99.0%, while the multi-class accuracy score is around 97.2%. The preliminary results achieved with the image-based 2-D CNN are much lower which may indicate that additional efforts will be needed to take advantage of this approach.