This item is under embargo and not available online per the author's request. For access information, please visit http://libanswers.wustl.edu/faq/5640.
Date of Award
Doctor of Philosophy (PhD)
Parkinson disease, the second most common neurodegenerative disorder, is caused by the loss of dopaminergic subcortical neurons. Approximately 50% of people with Parkinson disease experience freezing of gait (FOG), a brief, episodic absence or marked reduction of forward progression of the feet despite the intention to walk. FOG causes falls and is resistant to medication in more than 50% of cases. FOG episodes can often be interrupted by mechanical interventions (e.g., a verbal reminder to march), but it is often not practical to apply these interventions on demand (e.g., there is not usually another person to detect an FOG episode and provide the reminder).Wearable sensors offer the possibility of detecting FOG episodes in real time and thus developing a “closed-loop” treatment: real-time detection can be coupled with on-demand interventions. Objective evaluation methods using wearable sensor technology to monitor and assess FOG have met with varying success. They do not use a signal model that captures FOG patterns explicitly, and they are of limited help in understanding the underlying mechanisms in the structure of the sensor data captured during FOG. In this dissertation, we first develop physically-based signal models for the sensor data, design statistical signal processing methods to detect FOG based on its patterns, and compute the probability of FOG. Then, we proceed to validate the system, using data from experimental gait assessment in a group of people with Parkinson disease.We further develop a modular approach to model, detect, and track FOG in Parkinson disease, using four modules, namely the detection, navigation, validation, and filtering modules. To capture the gait motion, we use an inertial measurement unit (IMU) consisting of a three-axis accelerometer and a three-axis gyroscope. We first build physically-based signal models that describe “no movement” and “trembling motion” during FOG events. In the detection module, we design a generalized likelihood ratio test framework to develop a two-stage detector for determining the zero-velocity event intervals (ZVEI) and trembling event intervals (TREI) that are associated with FOG. However, not all the detected TREI are associated with FOG. Therefore, to filter out the TREI which are not associated with FOG, we consider the fact that the alternating trembling motion in FOG is associated with low foot speeds and small pitch angles. Next, to estimate these gait parameters, we employ a zero-velocity aided inertial navigation system (ZV-INS) in the navigation module. The ZV-INS uses the ZVEI as pseudo measurements, along with a Kalman filter, to estimate the position, velocity, and orientation angles of the foot.To track the degradation of the gait parameters prior to the incidence of FOG, we detect valid gait cycles in the validation module. We first identify the non-stationary segments of the gyroscope signal in the sagittal plane, using ZVEI. Next, we preprocess the non-stationary segments by scaling and interpolating the signal. Finally, we validate the preprocessed non-stationary segment of the gyroscope signal in the sagittal plane as a valid gait cycle, using an optimization framework called sparsity-assisted wavelet denoising (SAWD). In the SAWD algorithm, we simultaneously combine low-pass filtering, multiresolution representations (wavelets), and a sparsity-inducing norm to obtain a sparse representation of the gyroscope signal in the sagittal plane for valid gait cycles, in the form of a discrete wavelet transform coefficient vector. We compute the root-mean-square error between the generated template and the sparse representation of the non-stationary segment of the gyroscope data in the sagittal plane, obtained using the SAWD algorithm. If the root-mean-square error is less than a fixed threshold, then the gait cycle is considered valid.Finally, to detect the onset and duration of FOG, we develop a point-process filter that computes the probability of FOG (pFOG). We model the edges of the TREI as a point-process, then assign weights to the edges, which depend on a participant-specific tunable parameter and the average value of the gait parameters observed in the bin containing the edge. To compute pFOG, we develop a Bayesian recursive filter and integrate the weights assigned to the edges of the TREI over a time window. To adaptively adjust the participant-specific tunable parameter, we develop two novel approaches that assign weights to the edges of the TREI based on the gait parameters extracted from the last valid gait cycle and the foot motion dynamics. We validate the performance of the modular system design using real data obtained from people with Parkinson disease who performed a battery of gait tasks known to trigger FOG. The results indicate improved performance, with an average accuracy greater than 85% and an average false positive rate of less than 14%. Altogether, we not only improve the accuracy of FOG detection but also open new avenues towards the development of low-cost remote health monitoring systems, which will help provide insights into the frequency and patterns of FOG that affect the quality of daily life in people with Parkinson disease.
Gammon Earhart, ShiNung Ching, Neal Patwari, Matthew Lew, Heinz Schaettler
Available for download on Saturday, August 14, 2021