Faculty Mentor Information
John Kalivas
Presentation Date
7-2016
Abstract
A goal of chemometric multivariate calibration (modeling) is to predict analyte concentration in a sample using spectral data. Multiple types of modeling methods have been used to predict analyte concentration. However, the samples contain interferents that influence the model and if not fully corrected by the model, analyte concentration prediction errors occur. To reduce the prediction errors caused by interferent species in the system, two new methods were designed to incorporate interferent information. One of the methods uses interferent spectra to require the model to be orthoganol to the interferents. The other method uses interferent spectra to form an orthogonal or oblique model to the interferents. The methods are compared to ridge regression and partial least squares using a near infrared data set. Sum of ranking is used to select models. The new methods have better analyte prediction errors and robustness, but more data sets need to be tested to confirm that both new methods are more effective.
Reducing Spectral Analyte Prediction Error with Penalties on Interferents
A goal of chemometric multivariate calibration (modeling) is to predict analyte concentration in a sample using spectral data. Multiple types of modeling methods have been used to predict analyte concentration. However, the samples contain interferents that influence the model and if not fully corrected by the model, analyte concentration prediction errors occur. To reduce the prediction errors caused by interferent species in the system, two new methods were designed to incorporate interferent information. One of the methods uses interferent spectra to require the model to be orthoganol to the interferents. The other method uses interferent spectra to form an orthogonal or oblique model to the interferents. The methods are compared to ridge regression and partial least squares using a near infrared data set. Sum of ranking is used to select models. The new methods have better analyte prediction errors and robustness, but more data sets need to be tested to confirm that both new methods are more effective.
Comments
Poster #Th53