|
Researchers in Spain have developed a way to eliminate uninformative variables from the multivariate calibration of near-infrared spectra and so allow a more productive determination of heroin in illicit street drugs to be carried out. Javier Moros, Julia Kuligowski, Guillermo Quintás, Salvador Garrigues, and Miguel de la Guardia of the Department of Analytical Chemistry, at the University of Valencia, Spain, explain that their approach allows variable elimination to be carried out. It is, they say, based on the use of the standard deviations of a series of reliability coefficients, that offers an estimate for the cut-off threshold. They have now provided a proof of principle by testing the new criterion's performance on real data from diffuse reflectance near infrared (NIR) spectra of illicit seized street drugs containing heroin and a wide variability of "cutting" agents, additives used to dilute or bulk up street heroin. "Multivariate calibration is one of the most important research topics in the field of statistics and analytical chemistry," Moros and colleagues explain. "It covers a wide range of applications and is frequently used in conjunction with spectroscopic data, being this one of the most common fields of application of chemometric techniques." In the near infrared (NIR) region there lots of confusing factors, such as the typically very broad signals, which complicate the spectra meaning that the assignment of specific spectral features to a given chemical component of a complex sample is, to say the least, often very troublesome. Nevertheless, NIR spectroscopy is proving useful a range of fields particularly as progress in multivariate calibration can disentangle those overlapping signals. "Today, NIR spectroscopy is widely used for the measurement of a wide range of sample features in application areas such as agriculture, food, manufacturing, and pharmaceutical industries, among others," the team adds. The application of partial least squares (PLS) regression to a full-spectrum multivariate calibration allows analytical scientists to determine sample components. Unfortunately, the use of a full spectrum does not always provide optimal data. The spectra can include regions that contain no relevant information and Moros and colleagues hoped to address this deficit that simply confuses the analysis by eradicating the uninformative data from the cheminformatics. Other researchers have demonstrated that better PLS models can be obtained by this elimination approach, which makes interpretation of a spectrum easier and reduces the degree of prediction errors. Indeed, uninformative variable elimination for PLS (UVE-PLS) has been known for more than a decade. "Since its introduction this interesting approach has been extensively used in analytical chemistry," the team says. The team explains that the UVE-PLS approach can classify a variable as informative or previous uninformative depending on a "reliability parameter", which is obtained from the PLS coefficients. Variables below a certain threshold are then considered uninformative and simply removed from the data set before the final calculations are done. By testing various approaches to this UVE-PLS technique, the researchers have found that the application of P1, or quantile 100, criteria for the cut-off threshold with DR-NIR spectra gives them the best prediction capabilities for heroin samples. Used in this way, P1 is more selective than other sampling methods and simplifies the final PLS model by reducing its overall dimensionality, it also eliminates the majority of uninformative variables. Related links:
|
|
