< Terug naar vorige pagina

Publicatie

Deep learning for automated movement analysis using motion capture: Case studies on detecting and assessing gait impairments in people with Parkinson's disease and freezing of gait

Boek - Dissertatie

% !TeX root = ../../thesis.tex Parkinson's disease (PD) is a common neurodegenerative disorder, affecting an estimated 6 million people worldwide. PD causes degeneration of dopamine-generating neurons, which leads to progressively worsening gait impairments. One of the most debilitating gait impairments is freezing of gait (FOG). FOG is common in PD and is defined as a 'brief, episodic absence or marked reduction of forwarding progression of the feet despite the intention to walk and reach a destination". PD patients with FOG have a lower quality of life and are at a much higher risk of falling. The current assessment and management of gait impairments in PD are challenging, as existing self-assessment questionnaires are subjective and often inadequate for tracking treatment effects on FOG severity and mild gait impairments. Hence, there is a clear need for a reliable and valid tool that can automatically detect FOG and temporal gait disturbances of people with PD and FOG. Motion capture (MoCap) has proven to be an invaluable tool for the objective assessment of treatment effects and the optimization of therapeutic interventions for people with PD and FOG. In clinical research, MoCap is often performed with marker-based optical MoCap systems. Recently, there has been a shift to marker-less optical MoCap and inertial-based MoCap, due to their potential to assess gait outside of a lab environment. Regardless of the setting, MoCap is a generic term used to describe measurement systems that aim to encode movement as a time series of joint locations and orientations or their higher-order representations. Earlier techniques that aimed to automatically characterize gait from MoCap data relied on hand-crafted representations of the time series. However, it remains challenging to craft representations that generalize due to high inter and intra-subject variability within the acquired time series. Therefore, most recent techniques have shifted to deep learning (DL). DL models employ end to end learning, meaning that DL models learn all the steps required to map a raw input representation to the desired output result. However, the current application of DL in gait assessment of people with PD relies on sliding-window schemes to segment MoCap sequences into fixed-length partitions. Such sliding-window schemes require several data-dependent heuristics that impact the expressivity of the learned representation and raises concerns regarding their generalization to other use-cases or study protocols. This thesis research aims to build automated frameworks for MoCap-based gait assessment that do not require data-dependent heuristics. This would allow them to quickly adapt to various tasks, with minimal human intervention. To accomplish this aim, the research in this thesis exploits the latest advancements in recurrent neural networks, and graph and temporal convolutional neural networks. Specifically, four DL models were defined based on: bidirectional long short-term memory networks (LSTM), temporal convolutional neural networks (TCN), TCNs with multiple refinement stages, and spatial-temporal graph convolutional neural networks (ST-GCN). An additional model was developed that amalgamates the concepts of multi-stage refinement, dilation, and graph convolutions into one novel architecture. The goal of the architectures is to learn discriminative representations that are sufficiently robust to inter and intra-subject variability by optimally exploiting the temporal and spatial dependencies within MoCap sequences. The application of these five DL models was investigated on various relevant case studies in MoCap-based human movement analysis in general, and gait assessment in people with PD and FOG in particular. In this context, this thesis provides two main contributions: (1) The models were applied to assess gait in people with PD and FOG. Objective gait and FOG severity outcomes were extracted and statistically validated versus the expert clinical annotation on a MoCap dataset of fourteen freezers, fourteen non-freezers, and fourteen healthy controls. Experimental results indicated that the best performing model enables automatic and valid gait assessment. (2) It was investigated whether the proposed models have general applicability on use-cases that are not related to PD. The models were applied on four public MoCap-based action segmentation benchmark datasets that feature three different MoCap representations, i.e. marker-based optical MoCap, marker-less optical MoCap, and inertial-based MoCap. Experimental results indicated that the novel proposed model outperforms the four existing state-of-the-art baselines. As all models were applied end-to-end, (i.e., the mapping from raw input data to the output action was learned without human intervention), it can be established that the proposed models are strong baselines for classification tasks using MoCap measurement systems.
Jaar van publicatie:2022
Toegankelijkheid:Open