Classification, segmentation, and detection of switching dynamic modes in biological time series

In this dissertation, I develop a method to identify switching dynamic modes in time series, termed Improved Annealed Competition of Experts algorithm (IACE). I utilize systematic approaches based on mutual information and false nearest neighbor to determine appropriate embedding dimension and time delay. Moreover, I obtained further improvements by incorporating a deterministic annealing approach as well as a phase space closeness measure during the training procedure. The application of IACE method to RR interval (Time duration between two consecutive R waves of the electrocardiogram) data obtained from rats during control and administration of double autonomic blockade conditions indicate that IACE algorithm is able to segment dynamic mode changes with pinpoint accuracy.; In the second section of this dissertation, I extend IACE algorithm and use it for detection of linear and nonlinear interactions, by employing histograms showing the frequency of switching modes obtained from the IACE, then examining time-frequency spectra. This extended approach is termed Histogram of Improved Annealed Competition of Experts - Time Frequency (HIACE-TF). With all 10 data sets, comprised of volumetric renal blood flow data, I validated the feasibility of the HIACE-TF approach in detecting nonlinear interactions between the two mechanisms responsible for renal autoregulation.; In the third section of this dissertation, I focus on investigating and classifying inspiratory motor output based on its time-frequency representation (TFR). The dynamic features of inspiratory motor output are obtained by TFR. Visual detection reveals two classes of bursts, whose TFR exhibit concentrated and diverse features respectively. I develop an automatic classification method, based on a fuzzy hybrid neural network, to classify the inspiratory bursts by training TFR features of the bursts. I apply the automatic classification method to inspiratory motor output in anesthetized mice in vivo. The method correctly classifies the burst patterns with high correction rate comparing to the results obtained by visual detection.