الفهرس | Only 14 pages are availabe for public view |
Abstract This thesis has been concerned with the development of an automated system for the monitoring, interpreting and classifying of EMG records. The main focus has been to investigate how more information can be extracted from the records by carrying through a detailed study of their signal analysis and interpretation in order to utilize them in further diagnostic stages. The system combines analytical signal processing techniques with feature extraction and pattern recognition features from the available records, which are then used in subsequent pattern classification Three groups of subjects were investigated: normal subjects, neuropathic patients, and myopathy patients. The task has been to identify and apply appropriate techniques whereby the underlying general features of the records would be recognized. In particular, it has been necessary to establish the presence or absence of important significant features in the records to characterize the abnormalities. Three methodologies were adopted for the analysis of the EMG records. These include the decomposition of the EMG signals into their MUAP?s patterns using isolated patterns, spectral analysis using autoregressive modeling, and wavelet transforms. The extracted features from each technique were used them separately to form feature vectors to be applied to the selected classification approach. The dimensionality reduction of the features was made using correlation matrix, Principal component analysis (PCA), and the average energy content of the resulting coefficients. A statistical analysis of the extracted parameters was applied in order to detect significant differences between the three groups of cases. The mean, mode, and standard deviation were calculated. Frequency distributions of each parameter and coefficients were also constructed and student?s ttest was used. The results have shown that there were significant differences in the mean value of some specific parameters. Two types of neural networks were utilized; multilayer feedforward neural network trained with conjugate gradient algorithm, and Kohonen?s selforganizing feature maps. Two topologies for the multilayer feedforward neural network were utilized: feedforward with two layers, and three layers. The chosen network was trained and tested using feature vectors derived from three methodologies. The investigation of two neural network classifiers has revealed that a feed forward multilayer neural network trained using conjugate gradient method with three layers was able to give the highest classification rate when using features derived from the wavelet transform. The heighest correct classification rate reaches 97%. |