Pattern Recognition Methods In Biomedical Magnetic Resonance

In the pharmaceutical industry and drug discovery program, a lot of scientists make great efforts to judge the effect of some candidate drugs at the most possible early stage. In the later 1990's, metabonomic approach was proposed, which provides a rapid and effective method to evaluate the toxicity of the candidate drug. The nuclear magnetic resonance (NMR) and pattern recognition have played a key role in metabonomics. This thesis focuses on the application of pattern recognition to biomedical NMR study. The main content is as follows:1. The development of metabonomics was reviewed. The features of the fluid NMR spectroscopy and how to prepare samples for experiments were described. The most often used experimental NMR techniques were discussed.2. The pattern recognition methods in biomedical NMR experimental data processing were reviewed. The theory and application of unsupervised methods such as Principal Components Analysis, Nonlinear Mapping, Hierarchical Cluster Analysis and supervised methods such as Partial Least Squares, Linear Discriminant Analysis, K-nearest Neighbor Analysis, Neural Networks were discussed.3. K-nearest Neighbor Analysis (KNN Analysis) method was implemented and applied in two groups of liquid NMR data for toxicity study. For one group the predicting accuracy can be improved from 73.3% to 93.3% by increasing the samples of the training set. For the other predicting accuracy of 86% was achieved by adjusting the number of the nearest neighbor. The results also indicated that K-nearest Neighbor Analysis was more accurate than Principal Components Analysis.4. Probabilistic Neural Networks method was implemented and used to analyze two NMR data. The 86.7% prediction accuracy can be achieved by selecting appropriate window width of the window function and increasing the samples of the training set for one group, and 72% for the other.