Novel Computational Intelligent Technologies And Their Applications In Biomedical Signal Analysis

Nowadays, computational intelligence plays a pivotal role in intelligent science, and it is also considered as one of the cutting-edge information technologies. The methodologies of computational intelligence develop a number of emerging interdisciplinary areas, such as machine learning, data mining, intelligent control, which have been extensively applied to many practical problems. The computational intelligent approaches have been widely accepted to be meritorious in signal processing, pattern recognition, optimization of nonlinear system, data engineering, computer-aided medical diagnosis, etc.This thesis mainly focuses on the fusion strategy in neural network ensemble systems, and the adaptive signal processing, and proposes novel technologies which can be used for computer-aided medical diagnosis and biomedical signal analysis. The major contributions of this thesis are summarized as follows:First, this thesis proposes the use of statistical parameters of vibroarthrographic signals, including the form factor involving the variance of the signal and its derivatives, skewness, kurtosis, and entropy. Then, the radial-basis function network, a typical type of feedforward neural network, was used to analyze vibroarthrographic signals and diagnose knee joint pathology. Later, a novel optimal multilayer perceptron architecture selection method based on regularization and cross-validation is proposed.Second, the thesis summarizes the nongenerative neural network ensemble approaches, such as majority vote, simple average and weigthed average, and proposes several neural network fusion strategies, in particular perceptron average, least-mean-square fusion, normalized weighted average, and adaptive linear combination. The experiments of breast cancer diagnosis, yeast protein subcellular localization sites prediction, breast tissue classification, and function approximation demonstrated the merits of the neural network ensemble systems.Third, the thesis presents two novel adaptive systems to remove the noise and interference in electrocardiographic (ECG) signals. The first system used adaptive filters to eliminate high-frequency random noise, together with some signal processing procedures. The second system is an unbiased and normalized adaptive noise reduction system to suppress random noise. The corresponding noise estimation model is adaptively updated by using the steepest-descent algorithm which minimizes the instantaneous error between the estimated signal power and the desired noise-free signal power. The benchmark ECG databases were used to test the performance of the aforementioned two adaptive systems. The adaptive filter system can provide noise-free ECG signals with lower root-mean-square error and higher normalized correlation coefficient than popular least-mean-square and recursive-least-squares adaptive filers. The adaptive noise reduction system, on the other hand, can provide the higher signal-to-noise ratio improvement in ambulatory ECG recordings over the least-mean-square filter.