Unsupervised Techniques for Time Series Mining and Information Extraction in Medical Informatics

The proliferation of ubiquitous sensing devices along with advances in wireless communication technology will result in extensive use of medical and health monitoring systems. These systems enable constant monitoring of patients vital signs, activities, habits, and environment, which results in generation of tremendous amount of data. The size and the novel nature of such data provides the opportunity to perform fast diagnosis of disease or provide realtime feedback to the patients and caregivers. However, there is a huge gap between the pace that the new sensing technologies are developed and the time required to design efficient and effective data processing techniques, mainly because developing expert knowledge around the new types of data and applications is costly and time consuming. In this thesis we present design, implementation, and evaluation of unsupervised data mining and information extraction techniques which are suitable for medical and health monitoring applications.;The essential idea of unsupervised mining is to discover interesting events and patterns in data according to the frequency of events and the temporal relation between different patterns. Because developing and deploying unsupervised techniques don't require training data, they are among the fastest and most efficient techniques for data processing in the rapidly growing field of medical monitoring. By extending the applicability of the well-known motif discovery algorithm from single-dimensional data to multi-dimensional time series, we show the effectiveness of applying unsupervised techniques to health monitoring systems for discovering most frequent activities and temporal relations between them. The developed technique is especially suitable for human monitoring applications where uncertainty, diversity, and noise is a persistent element of the collected data. We further extend this approach to detect abnormalities in data, which is of special interest in medical community. Our unsupervised activity and abnormality discovery techniques are especially shown to be useful in situations where subjects environment and habits are not known a priori. In addition, the unsupervised activity discovery technique is leveraged to discover hidden information from the data, which enables context and data aware computing. In this regard, a technique to discover the position of sensors on the human body is introduced. Automatic on-body device localization ensures correctness and accuracy of measurements in health and medical monitoring systems. In addition, it provides opportunities to improve the performance and usability of ubiquitous devices. To evaluate the proposed unsupervised data mining techniques, we introduce three health monitoring devices, which are used as our testbed systems. Comparison is made to classical supervised techniques to show the advantages of unsupervised methods in facilitating health care monitoring systems setup phase, accelerating the data processing flow, and decreasing the expert's knowledge requirement in medical monitoring.