Anzanello M J, Kahmann A, Marcelo M C A, Mariotti K C, Ferrão M F, Ortiz R S
Department of Industrial Engineering, Federal University of Rio Grande do Sul, Porto Alegre, RS, Brazil.
Department of Industrial Engineering, Federal University of Rio Grande do Sul, Porto Alegre, RS, Brazil.
J Pharm Biomed Anal. 2015 Nov 10;115:562-9. doi: 10.1016/j.jpba.2015.08.008. Epub 2015 Aug 18.
Cocaine ATR-FTIR spectra consist of a large number of wavenumbers that typically decreases the performance of exploratory and predictive multivariate techniques. This paper proposes a framework for selecting the most relevant wavenumbers to classify cocaine samples into two categories regarding chemical composition, i.e. salt and base. The proposed framework builds a wavenumber importance index based on the Bhattacharyya distance (BD) followed by a procedure that removes wavenumbers from the spectra according to the order suggested by the BD index. The recommended wavenumber subset is chosen based on multiple criteria assessing classification performance, which are recalculated after each wavenumber is eliminated. The method was applied to ATR-FTIR spectra from 513 samples of cocaine, remarkably reducing the percent of retained wavenumbers and yielding near to perfect classifications in the testing set. In addition, we compared our propositions with other methods tailored to wavenumber selection; we found that the proposed framework, which relies on simple mathematical fundamentals, yielded competitive results.
可卡因的衰减全反射傅里叶变换红外光谱（ATR-FTIR）包含大量波数，这通常会降低探索性和预测性多变量技术的性能。本文提出了一个框架，用于选择最相关的波数，以便根据化学成分将可卡因样品分为两类，即盐类和碱类。所提出的框架基于巴氏距离（BD）建立波数重要性指数，随后按照BD指数建议的顺序从光谱中去除波数。根据评估分类性能的多个标准选择推荐的波数子集，在每次去除一个波数后重新计算这些标准。该方法应用于513个可卡因样品的ATR-FTIR光谱，显著减少了保留波数的百分比，并在测试集中产生了近乎完美的分类。此外，我们将我们的方法与其他针对波数选择的方法进行了比较；我们发现，所提出的基于简单数学原理的框架产生了具有竞争力的结果。
