A Study of Ensemble Machine Learning Model Architectures for Parkinson’s Disease Detection and Freezing of Gait Forecasting

  • Author / Creator
    Ashfaque Mostafa, Tahjid
  • Parkinson's Disease (PD) is a major progressive neurological disorder and is extremely difficult to diagnose PD~\cite{pmid23039866} since there are no defined medical tests for this task. The existing approach involves a combination physical examinations, neuroimaging and demographic analysis performed by expert medical professionals. This process is both time and resource draining as well as being prone to human error and bias. Moreover, the motor symptoms might not appear until the advanced stages of the disease, the diagnosis often does not provide ample time to administer preventive measures. Computer Aided Diagnosis (CAD) systems have been gaining popularity in recent years, but these solutions are not without their own shortcomings. In this work, non-invasive approaches to identify PD and monitor the progression of one of the motor symptoms of PD using deep learning based techniques are analyzed.

    We explore various approaches to discuss PD case from control using Magnetic Resonance (MR) T1 images of the brain, one of the most popular neuroimaging techniques, non-invasive and generating high resolution images in the soft tissue. We experimented with some Convolutional Neural Network (CNN) models of ImageNet Large Scale Visual Recognition Challenge (ILSVRC) with whole brain, extracted Gray Matter (GM) and White Matter (WM) scans. We also propose multiple ensemble architectures combining the ILSVRC models. The detection accuracy increases drastically when we focus on the extracted GM and WM regions from the MR images instead of using the whole brain scans. ILSVRC Deep Learning (DL) models pretrained on the ImageNet dataset perform relatively better than when they are trained solely on the MRI scans. The proposed solutions outperform state of the art existing methods on similar datasets.

    One of the major obstacles in applying learning algorithms to this task is lack of properly labeled training data. So our finding that training on unrelated data might increase the performance of DL models is a possible solution. We also perform occlusion analysis and determine brain areas are relevant in the DL architectures decision making process. This was to further narrow down the regions of interest. Focusing on the identified relevant regions might be helpful in achieving the same performance while reducing the amount of data needed to be processed.

    Freezing of Gait (FOG) is an impairment that affects the majority of patients in the advanced stages of PD, defined as a short period of time when the patient fails to move forward, despite attempting to do so. The patients describe this event as a sudden feeling of their feet being stuck to the ground. FOG can lead to sudden falls and injuries, negatively impacting the quality of life for the patients and their families. Rhythmic Auditory Stimulation (RAS) can be used to help patients recover from FOG and resume normal gait. However, even if FOG is detected in early stages, RAS might not be effective due to the latency between the start of a FOG event, detection and initialization of RAS. In the second section, I propose a system capable of both FOG prediction and detection using signals from tri-axial accelerometer sensors. This approach will be useful in initializing RAS with minimal latency. I compared the performance of several time frequency analysis techniques, including moving windows extracted from the signals, handcrafted features, Recurrence Plots (RP), Short Time Fourier Transform (STFT), Discreet Wavelet Transform (DWT) and Pseudo Wigner Ville Distribution (PWVD) with DL based Long Short Term Memory (LSTM) and CNN. I also propose three Ensemble Network Architectures that combine all the time frequency representations and DL architectures. Experimental results show that our ensemble architectures significantly improve the performance compared with existing techniques on benchmark dataset. Our research group also collaborated with A. T. Still University to collect motion data for a group of PD patients, some of whom experienced FOG during the data collection. I also applied the methods proposed in the second section on this data to identify the instances of FOG.

  • Subjects / Keywords
  • Graduation date
    Fall 2021
  • Type of Item
  • Degree
    Master of Science
  • DOI
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.