Search

Monitoring of engineered systems consists of characterizing the normal behavior of the system and tracking departures from it. Techniques to monitor a system can be split into two classes based on their use of inputs and outputs of the system. Systems-based monitoring refers to the case when both inputs and outputs of a system are available and utilized. Conversely, symptomatic monitoring refers to the case when only outputs of the system are available. This thesis extended symptomatic and systems-based monitoring of biomedical systems via the use of non-stationary signal processing and advanced monitoring methods. Monitoring of various systems of the human body is encumbered by several key hurdles. First, current biomedical knowledge may not fully comprehend the extent of inputs and outputs of a particular system. In addition, regardless of current knowledge, inputs may not be accessible and outputs may be, at best, indirect measurements of the underlying biological process. Finally, even if inputs and outputs are measurable, their relationship may be highly nonlinear and convoluted. These hurdles require the use of advanced signal processing and monitoring approaches. Regardless of the pursuit of symptomatic or system-based monitoring, the aforementioned hurdles can be partially overcome by using non-stationary signal analysis to reveal the way frequency content of biomedical signals change over time. Furthermore, the use of advanced classification and monitoring methods facilitated reliable differentiation between various conditions of the monitored system based on the information from non-stationary signal analysis. The human brain was targeted for advancement of symptomatic monitoring, as it is a system responding to a plethora internal and external stimuli. The complexity of the brain makes it unfeasible to realize system-based monitoring to utilize all the relevant inputs and outputs for the brain. Further, measurement of brain activity (outputs), in the indirect form of electroencephalogram (EEG), remains a workhorse of brain disorder diagnosis. In this thesis, advanced signal processing and pattern recognition methods are employed to devise and study an epilepsy detection and localization algorithm that outperforms those reported in literature. This thesis also extended systems-based monitoring of human biomedical systems via advanced input-output modeling and sophisticated monitoring techniques based on the information from non-stationary signal analysis. Explorations of system-based monitoring in the NMS system were driven by the fact that joint velocities and torques can be seen NMS responses to electrical inputs provided by the central nervous system (CNS) and the electromyograph (EMG) provides an indirect measurement of CNS excitations delivered to the muscles. Thus, both inputs and outputs of this system are more or less available and one can approach its monitoring via the use of system-based approaches.

High-resolution, quantitative data obtained by a human motion capture system can be used to better understand the cause of many diseases for effective treatments. Talking about the daily care of the aging population, two issues are critical. One is to continuously track motions and position of aging people when they are at home, inside a building or in the unknown environment; the other is to monitor their health status in real time when they are in the free-living environment. Continuous monitoring of human movement in their natural living environment potentially provide more valuable feedback than these in laboratory settings. However, it has been extremely challenging to go beyond laboratory and obtain accurate measurements of human physical activity in free-living environments. Commercial motion capture systems produce excellent in-studio capture and reconstructions, but offer no comparable solution for acquisition in everyday environments. Therefore in this dissertation, a wearable human motion analysis system is developed for continuously tracking human motions, monitoring health status, positioning human location and recording the itinerary. In this dissertation, two systems are developed for seeking aforementioned two goals: tracking human body motions and positioning a human. Firstly, an inertial-based human body motion tracking system with our developed inertial measurement unit (IMU) is introduced. By arbitrarily attaching a wearable IMU to each segment, segment motions can be measured and translated into inertial data by IMUs. A human model can be reconstructed in real time based on the inertial data by applying high efficient twists and exponential maps techniques. Secondly, for validating the feasibility of developed tracking system in the practical application, model-based quantification approaches for resting tremor and lower extremity bradykinesia in Parkinson’s disease are proposed. By estimating all involved joint angles in PD symptoms based on reconstructed human model, angle characteristics with corresponding medical ratings are employed for training a HMM classifier for quantification. Besides, a pedestrian positioning system is developed for tracking user’s itinerary and positioning in the global frame. Corresponding tests have been carried out to assess the performance of each system.

For ill defined, time varying, or non-linear systems, adaptive controllers offer an attractive alternative to classical design techniques. These controllers require minimal information. Ironically, the information that is available is often discarded. As a result, control may become unstable. If the available information could be used instead of discarded, the instability might be avoided. We therefore developed a real-time algorithm that imposes linear equality and inequality constraints on a time series model of the process. Thus commonly available information, such as open loop stability, settling time, and steady state gain, can be incorporated into the control. When the information is imposed according to our guidelines, control errors due to mismodeling can be significantly attenuated (in one instance, we reduced the mean squared output error by more than two orders of magnitude). We demonstrate our algorithm in two practical biomedical applications. In the first, we consider second order linear compartmental models (a popular model for pharmacodynamical systems). We develop novel constraints for these models. Then, as an example, we control plasma and tissue concentrations of methotrexate, an antimetabolite used in the treatmen t of certain neoplastic diseases. Our results show that the constraints improve both controller performance and patient safety. In the second biomedical application, we lower mean arterial pressure with a vasodilator. This system is non-linear and time-varying; hence it can be difficult to control. We simulate this system with a series of models that are progressively more complex. For each model, we develop suitable constraints which we then impose during control. As the models become more complex, we show that the constraints become more important for safety and improved control. While pursuing this work, we discovered a simple modification that significantly improves the efficiency and accuracy of positive semi-definite complementary linear programming (a technique for solving quadratic programs). We prove its validity and modify the pivot selection rules to implement least distance programming.