Application Of Nonlinear Analysis Methods In Biomedical Signal Processing

Biomedical signals are the random signals generated by the complex mechanics of physiological systems in human body. The biomedical signals commonly exhibit high degree of nonstationary and variability, and long range of frequency bandwidth. Biomedical signal processing plays an important role in several clinical applications, especially for feature-analysis-based computer-aided diagnosis systems that can help physicians make early-stage diagnosis and pharmacotherapy.In this thesis, we use the nonlinear analysis methods to study the phonation records of patients with Parkinson’s disease and knee joint vibroarthrographic signals. We employ the correlation matrix and sequential forward selection methods to select the representative vocal period, fundamental frequency, and perturbation parameters. Then we use the nonparametric kernel density technique to estimate the distributions of different signal patterns in bivariate feature space. We also implement different classifiers to effectively perform phonation pattern classifications. In addition, we apply the detrended fluctuation analysis method to study the self-similarity of vibroarthrographic signals in different time scales, and also propose the averaged envelope amplitude parameter to measure the pathological signal fluctuations. We use the kernel density estimation method to establish the feature distributions models, which help the least-squares support vector machine and Bayesian decision rule achieve excellent diagnostic performance. The experiments demonstrate that the nonlinear analysis methodology is very useful for the analysis of perturbation characteristics of nonstationary biomedical signals, and may provide meaningful information for further feature distribution and pattern classifications.