Given, Douglas George2006-07-282006-07-2819971997http://hdl.handle.net/10012/47The need for rapid, non-invasive techniques to screen specimens is becoming more critical as chemical laboratories enter the automation era. Near-infrared spectroscopic instrumentation is capable of meeting the requirements for specimen screening in the automated environment. However, some measurements from near-infrared spectroscopic instruments yield very low signal-to-noise -ratios. Therefore, the data analysis method used to calibrate such instrumentation must optimise the performance and should also provide a parsimonious solution to maintain a rapid measurement. In this thesis urine, serum, plasma, and plasma anticoagulant spectroscopic data were collected, processed and studied to evaluate the performance of various classification methods, namely, K-Nearest Neighbour and Mahalanobis Distance methods. Wavelengths were transformed into principal component scores, to reduce the number of features. The Mahalanobic Distance method was also optimised using a Genetic algorithm to select the best wavelengths, thus reducing the number of wavelengths required. The conclusion is that the Mahalanobic Distance method is superior to the K-Nearest Neighbour method in terms of predictability. The Genetic algorithm was able to increase predictability even further, while reducing the number of wavelengths required in the Mahalanobis Distance model.application/pdf5154031 bytesapplication/pdfenCopyright: 1997, Given, Douglas George. All rights reserved.Harvested from Collections CanadaMaster Thesis