Advancements in wearable technologies have equipped smartwatches with various sensors. Motion and direction sensors, namely accelerometers, gyroscopes, and magnetometers, have broad applications in human activity recognition. While many activity recognition algorithms have been proposed in recent years, most studies emphasize the use of inertial sensors in smartphone devices or other body-worn sensors. By comparison, very few works have evaluated the application of smartwatches for activity monitoring applications. In this study, we present a system to detect five daily activities using smartwatches. Our system identifies activities with 91.9% accuracy based on over 540 minutes of data collected from twelve subjects. We also demonstrate that the opportunity to deploy active learning for activity recognition can be a significant advantage of smartwatches over other devices. By collecting personal data from subjects online, active learning improves the accuracy of classifier predictions, while lowering variance between different users.

In the last decade, machine learning algorithms have been substantially developed and they have gained tremendous empirical success. But, there is limited theoretical understanding about this success. Most real learning problems can be formulated as non-convex optimization problems which are difficult to analyze due to the existence of several local optimal solutions. In this dissertation, we provide simple and efficient algorithms for learning some probabilistic models with provable guarantees on the performance of the algorithm. We particularly focus on analyzing tensor methods which entail non-convex optimization. Furthermore, our main focus is on challenging overcomplete models. Although many existing approaches for learning probabilistic models fail in the challenging overcomplete regime, we provide scalable algorithms for learning such models with low computational and statistical complexity.

In probabilistic modeling, the underlying structure which describes the observed variables can be represented by latent variables. In the overcomplete models, these hidden underlying structures are in a higher dimension compared to the dimension of observed variables. A wide range of applications such as speech and image are well-described by overcomplete models. In this dissertation, we propose and analyze overcomplete tensor decomposition methods and exploit them for learning several latent representations and latent variable models in the unsupervised setting. This include models such as mulitiview mixture model, Gaussian mixtures, Independent Component Analysis, and Sparse Coding (Dictionary Learning). Since latent variables are not observed, we also have the identifiability issue in latent variable modeling and characterizing latent representations. We also propose sufficient conditions for identifiability of overcomplete topic models. In addition to unsupervised setting, we adapt the tensor techniques to supervised setting for learning neural networks and mixtures of generalized linear models.

Wearable sensing systems are becoming widely used for a variety of applications including sports, entertainment, and the military. These systems have recently enabled a variety of medical monitoring and diagnostic applications. Such sensing systems have high potential to significantly improve the quality of life for large segments of the population and enable conceptually new types of applications. However, various research challenges including energy sensitivity and semantic complexity must be addressed before these devices can reach their full potential in improving our lives. The need for multiple sensors, high frequency sampling and constant monitoring leads these systems to be power-hungry and expensive with a short operation lifetime. In addition these systems generate a massive amount of data, where designing efficient and effective data mining algorithms to interpret and analyze the data is very costly due to field experts' involvement.

This dissertation presents a methodology that takes advantage of contextual and semantic properties in human physiological behavior to enable efficient design and optimization of such systems from the data and information point of view. The methodology uses combinatorial modeling and simultaneous minimization to reduce the wireless communication and local processing power consumption. The effectiveness of this technique is shown on an insole instrumented with 99 pressure sensors placed in each shoe, which is used for human gait analysis. Further, the proposed methodology is used to design unsupervised data mining techniques to extract information such as frequent or rare events and patterns from collected multi-dimensional time series data of wearable sensing devices, where high level of noise and uncertainty is inherent. This approach transforms multi-dimensional time series data to combinatorial space and constructs behavior models, which is used for frequent and rare event detection and pattern classification. The effectiveness of this method is demonstrated by applying this technique to discovery and classification of frequent human activities and abnormalities.

The escalating cost of healthcare and the growing prevalence of chronic diseases have created an urgent need for new solutions. Machine learning has the potential to revolutionize healthcare by providing more personalized and efficient care. However, there are unique challenges associated with applying machine learning in healthcare. Privacy concerns prevent data sharing across institutions, which limits available training data, and collecting individual features for patients may be invasive or expensive, as they may involve lab tests or medical imaging. In addition, machine learning models must be explainable so that medical professionals can understand how they arrive at a certain diagnosis. Despite these challenges, machine learning presents new opportunities in healthcare, both in hospitals and in remote health monitoring. In hospitals, machine learning can improve efficiency by assisting medical professionals with patient diagnoses, while in remote health monitoring, the vast quantities of data from personal and wearable devices open new opportunities for preventative care. However, processing and extracting meaningful insights from healthcare data require novel techniques. This dissertation investigates solutions for personalized healthcare in both hospital and remote healthcare settings, including adaptive data acquisition, unsupervised medical image segmentation, and remote health monitoring algorithms and applications. Overall, these solutions have the potential to improve patient outcomes and reduce healthcare costs.

Obstructive sleep apnea (OSA) is the most common type of sleep apnea and the most prevalent sleep disorder in general. Approximately 22 million people in the United States and more than 100 million people worldwide are affected by this serious disorder. The disorder is characterized by repeated involuntary breathing cessations (apnea) or reductions in breathing (hypopnea) during sleep, accompanied by oxygen desaturation (hypoxia). The apneas occur due to the obstruction of the upper airway as a result of the relaxing of throat tissue. The hypoxia triggers a brain activity arousal to restore the normal breathing pattern until the next apneic episode occurs. OSA patients can experience hundreds of apneas per night. The frequent sleep disruptions and the resulting fragmented sleep pattern can have serious physical and psychological consequences and can lead to premature death.

Because the apnea episodes rarely trigger a full awakening, patients are often unaware of having difficulty breathing at night. As a result of this unawareness, OSA is severely underdiagnosed. This problem is exacerbated by the intrusiveness of current diagnosis methods. Polysomnography---the traditional gold-standard diagnosis method---requires close overnight monitoring of patients' body functions, making it an uncomfortable and intrusive diagnostic test. Intrusive diagnosis methods discourage patients from getting tested for OSA until they experience its serious health effects. As a result, many patients remain untreated and OSA continues to be a major public health problem that places a significant burden on health care systems worldwide.

In an effort to increase the level of diagnosis and treatment of the disorder, this work investigates the design and implementation of effective non-intrusive OSA diagnostic methods. We explore two kinds of detection methods: nocturnal methods that rely on diagnosing OSA by detecting apneic episodes from overnight recordings, as well as daytime methods which use features and signals that can be obtained while the subject is awake. This work involves identifying the data to be collected, building the appropriate systems to obtain them, and using machine learning techniques to analyze them.

The effectiveness of our methods was proven using real data from clinical trials. Daytime methods quickly revealed their advantages over the common nocturnal methods and the few existing daytime methods.

Experimental results show that our daytime methods perform the initial diagnosis of OSA non-intrusively without significantly affecting the diagnosis accuracy compared to the current state-of-the-art nocturnal methods.

Pervasive health is an emerging field of research aimed at providing health monitoring to anyone at any time and anywhere. In contrast to existing healthcare systems, this paradigm is centered on prevention, wellness maintenance, and proactive care outside the hospital setting. In recent years, advances in sensing technologies and machine learning, as well as wide adoption of smart devices, have led to the development of successful pervasive health solutions such as daily activity monitoring.

Despite this growth, the full potential of pervasive health is yet to be unlocked. In this dissertation, we make several contributions to the application and methodological sides of pervasive health to advance it closer to its promises. We investigate design, implementation, and deployment of a highly context-aware and real-time solution for the complex application of pediatric asthma management, trying to predict the exacerbation of asthma ahead of time. On the methodological side, we investigate unmet data modeling challenges that arise in building a highly context-aware pervasive health solution and propose novel approaches to address them. In particular, we approach the challenges of integrative and interpretable modeling of heterogeneous health data, transferring and adapting models to new domains/individuals, and tackling sparsity of labels in real-world settings.

This dissertation investigates the limitation and challenges of wrist-worn sensing in activity state recognition including posture and gait quality inference. In spite of all the benefits that come with wrist-worn sensing systems, there are many challenges in inferring activities using them. The most important challenge is the high rate of false positive due to unwanted hand motion when the system is worn on the wrist. Another important challenge is the limitation on the power budget and processing capabilities of a wrist-worn device. Being far from the Center of Mass (COM), inferring posture and gait seem to be a very challenging problem using wrist data. This dissertation tries to provide a deep investigation on these challenges and propose ways to mitigate them. Different time series analysis methods are employed to serve for improvement in both accuracy and power consumption of activity state recognition. Along the way, we introduce some of the systems and algorithms that were built around these ideas and how they can be helpful in tackling the challenges of wrist-worn sensing systems, when it comes to activity state recognition.