Dictionary and deep learning algorithms with applications to remote health monitoring systems
Date
2017
Authors
Mathews, Sherin Mary
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Dictionary and deep learning algorithms facilitate efficient signal representations, thereby offering tremendous representational power along with achieving good recognition rates in real-world machine learning problems. In this dissertation, we present three dictionary learning approaches and a deep learning framework for classification tasks related to remote health monitoring systems. ☐ This dissertation presents a more robust class specific centralized dictionary learning method to solve the wearable sensor-based physical activity classification problem. Inspired by experiments that achieved high recognition rates using a few representative samples on high dimensional data, we explore the physical activity recognition signals from wearable sensors and propose a dictionary pair learning-based framework for human physical activity monitoring and recognition. The essential strategy involves integrating the class specific centralized regularizer term into the dictionary pair learning objective function and efficiently optimizing the objective function by combining the alternating direction method of multipliers and the l1 -- ls minimization method. Specifically, the class specific regularizer term ensures that the sparse codes belonging to the same class will be concentrated thereby enhancing the classification performance. Experimental results show that the classifiers built in this framework achieve higher recognition rate over four activity recognition tasks and outperforms state-of-the-art methods. ☐ Physical activity recognition involves variations in different walking styles and human body movements which result in the erroneous classification of similar activities. To address this issue, we present a correntropy induced dictionary pair learning framework to achieve improved recognition. In particular, the dictionary pair learning algorithm developed based on the maximum correntropy criterion is much more insensitive to outliers. A combination of alternating direction method of multipliers and an iteratively reweighted method is employed to approximately minimize the objective function. Evaluations are conducted using four activity recognition tasks and results show that the proposed classifier framework achieve enhanced performance compared to the state-of-the-art recognition systems. ☐ Although classification accuracy is enhanced using state-of-art classifiers, actual recognition performance tends to fall off when distinguishing a large number of similar activities. To this end, we propose and evaluate methods for analyzing hierarchical and sequentially structured human activities, designed to scale activity recognition by creating a hierarchical cluster of activity labels. Instead of using a single classifier to distinguish between large numbers of activities, we propose a hierarchy of classifiers, each of which distinguishes between child nodes at a particular location in the hierarchy. We hypothesize that building such a hierarchy of activity will improve recognition performance over that of the at classifier model. We validate the effectiveness of our proposed model by employing it on two standard activity recognition datasets, which include a large set of similar physical activities. The results of hierarchical structure modeling furnish evidence that decomposing the problem leads to more accurate specialized classifiers. ☐ This dissertation also applies deep learning methodology to the classification of single-lead electrocardiogram (ECG) signals. State of-the-art automatic ECG recognition systems often rely on a pattern-matching framework thereby requiring high sampling rates and burdensome computational times to classify arrhythmias. Deep learning networks represent a high level of abstraction showcasing its tremendous representational power. Consequently, to enable implementation in real time, we develop a deep learning framework that includes Restricted Boltzmann Machine and Deep Belief Networks for ECG classification with lower computational time, making it a highly practical option in a clinical setting.