Adaptive trend change detection and pattern recognition in physiological monitoring

Advances in monitoring technology have resulted in the collection of a vast amount of data that exceeds the simultaneous surveillance capabilities of expert clinicians in the clinical environment. To facilitate the clinical decision-making process, this thesis solves two fundamental problems in physiological monitoring: signal estimation and trend-pattern recognition. The general approach is to transform changes in different trend features to nonzero level-shifts by calculating the model-based forecast residuals and then to apply a statistical test or Bayesian approach on the residuals to detect changes. The EWMA-Cusum method describes a signal as the exponentially moving weighted average (EWMA) of historical data. This method is simple, robust, and applicable to most variables. The method based on the Dynamic Linear Model (refereed to as Adaptive-DLM method) describes a signal using the linear growth model combined with an EWMA model. An adaptive Kalman filter is used to estimate the second-order characteristics and adjust the change-detection process online. The Adaptive-DLM method is designed for monitoring variables measured at a high sampling rate. To address the intraoperative variability in variables measured at a low sampling rate, a generalized hidden Markov model is used to classify trend changes into different patterns and to describe the transition between these patterns as a first-order Markov-chain process. Trend patterns are recognized online with a quantitative evaluation of the occurrence probability. In addition to the univariate methods, a test statistic based on Factor Analysis is also proposed to investigate the inver-variable relationship and to reveal subtle clinical events. A novel hybrid median filter is also proposed to fuse heart-rate measurements from the ECG monitor, pulse oximeter, and arterial BP monitor to obtain accurate estimates of HR in the presence of artifacts. These methods have been tested using simulated and clinical data. The EWMA-Cusum and Adaptive-DLM methods have been implemented in a software system iAssist and evaluated by clinicians in the operating room. The results demonstrate that the proposed methods can effectively detect trend changes and assist clinicians in tracking the physiological state of a patient during surgery.Applied Science, Faculty ofElectrical and Computer Engineering, Department ofGraduat