Identificador: TDX:755
Autores: Boqué Martí, Ricard
Resumen:
Limits of detection in multivariate analysisInternational norms concerning the quality of analytical laboratories establish that an analytical method, before being used as a routine method, has to be validated. Validation is defined as the process of verifying that a method is fit for purpose, that is, suitable for its intended use. Among the important performance parameters in method validation, there is the limit of detection (LOD).Analytical methods that use techniques capable of generating multivariate data are more and more frequent. It is therefore important to derive their associated performance parameters. In this Doctoral Thesis we have focused on the study of the LODs. In the first chapter, the historical evolution of the concept of LOD is presented, together with the different approaches to calculate it in analytical methods using univariate calibration. The different factors affecting the calculation of the LOD are also described.. The different existing techniques of multivariate analysis, together with the various mathematical models used, have motivated the development of different approaches for calculating multivariate LODs. In the second chapter all the approaches developed so far are critically reviewed. In chapter 3 an approach is presented to calculate the LOD to data obtained from a hyphenated technique, gas chromatography-mass spectrometry. The approach is based on the use of the calibration line of the scores of the first principal component (obtained by principal component analysis decomposition of the original response matrix) versus the concentrations of the calibration standards. This approach is only applicable if interfering substances are not present and when the first principal component explains a very high percentage of the information in the original data. To overcome this limitation, a LOD estimator was developed to be applied to multivariate calibration using the direct model, in which the responses are modelled as a function of the concentrations. In these models, the calibration step can be carried out either from standards consistent on the pure analytes or standards consistent on mixtures of the different analytes. The LOD estimators derived for both cases, together with an application to real data, constitute the contents of chapter 4.Calibration methods based on the direct model have a big disadvantage: the concentrations of all the analytes contributing to the response or, at least, the spectra of pure analytes and interfering substances, must be known. This condition is seldom met in analytical laboratories, where most of the samples analysed are complex and have composition partially unknown. Also, usually only one specific analyte wants to be determined. With this type of samples one has to resort to multivariate calibration using the inverse model, in which the concentration of the analyte is modelled as a function of the response measured. Inverse models have the advantage that only the concentration of the analyte of interest needs to be known in the calibration samples when building the calibration model. An estimator was derived to calculate the LOD for multivariate calibration using the inverse model, applicable to a wider range of real samples (environmental, foodstuffs,...). In a first work, the LOD estimator is calculated from the confidence intervals of the multivariate calibration model. In a second approach, the estimator is based on the theory of hypothesis testing and uses the uncertainty of the predicted concentrations. These approaches, together with the applications to real data, are included in chapter 5.Finally, in the conclusions, the advantages and limitations of the developed LOD estimators are discussed and a series of guidelines are given on how to improve the LOD estimators in future works. As prospect research, the development of LOD estimators is suggested for analytical methods that generate second-order data.
Repositori URV