Pattern recognition has over past decades become a fast growing area of chemometrics. Accurate, user-friendly, and fast pattern recognition methods are desired to accommodate the increased capacity of automated instruments to obtain large-scale data under complex circumstances. It has found significant applications in diverse fields such as environmental monitoring and biomedical diagnostics. In this dissertation, the capabilities of pattern recognition methods in case studies related to environmental remote sensing and biomedical sensing are investigated.
For remote sensing applications, two types of airborne spectroscopic data, passive Fourier transform infrared (FTIR) and gamma-ray, are subject to analysis in order to develop automated classifiers for either ammonia vapor or the radioisotope cesium-137 in the open-air. Support vector machine (SVM) classification is the primary pattern recognition method used in this work. In order to overcome the limitation of available representative patterns associated with airborne data, and provide sufficient patterns presenting the analyte-active class for use in the training set, a spectral simulation protocol is employed to generate abundant patterns bearing both the signature of the target analyte and the background spectral profile. Signal processing procedures including segment selection and digital filtering are further used to extract the information most relevant to the target analyte out the acquired raw data. Also, to ease the computational demand from the SVM, an alternative pattern recognition method, piecewise linear discriminant analysis (PLDA) is applied to optimize signal processing conditions for final SVM classification. Process control techniques are applied to the SVM score profiles of prediction sets to improve pattern recognition performance by incorporating probabilities associated with every SVM score. Ammonia classifiers developed from this methodology result in classification performance with high sensitivity and selectivity, and the cesium-137 classifiers developed from the same concepts exhibit excellent sensitivity to test data with very low signal strengths. Under the case of ammonia classification, the relationship between the concentration profile of the active patterns in the training set and the limit of detection of the corresponding classifier is investigated. Classifiers built to detect low concentrations of ammonia are developed and tested through this work.
For a glucose sensing application, studies are conducted to provide sound performance diagnostics for an established calibration model for glucose from near infrared spectroscopic data. Six-component aqueous matrixes of glucose in the presence of five other interfering species, all spanning physiological levels, serve as samples to be analyzed. A novel residual modeling protocol is proposed to retrieve the residual glucose concentrations, the concentration not being predicted by the calibration model, from the residual spectra, the portion of the raw spectra not being used by the calibration model. The recovered glucose concentration from the residual modeling can be used as a means, combined with process control techniques, to evaluate the performance of the established calibration model. Several modeling techniques are used for residual modeling, including PLS, support vector regression (SVR), a hybrid method, PLS-aided SVR, and an amplified version of the hybrid, amplified PLS-aided SVR. Through this work, a calibration updating strategy is developed which provides an effective way to monitor the established calibration model.