Algoritmo das projeções sucessivas para seleção de variáveis em calibração de segunda ordem

Abstract

In this work it was developed a new strategy for intervals selection using the successive projections algorithm (SPA) coupled to N-PLS and U-PLS models, both with residual bilinearização (RBL) as a post-calibration step. The new algorithm coupled to N-PLS/RBL models was evaluated in two cases of studies. The first was simulated data for quantitation of two analytes (A and B) in the presence of a single interfering. On the second study was conducted a quantitation of ofloxacin in water in the presence of interferents (ciprofloxacin and danofloxacin) by means of liquid chromatography with diode array detection (LC-DAD) data modeling. The results were compared to the N-PLS/RBL model and the variables selection with the genetic algorithm (GA-N-PLS/RBL). In the first case of study (simulated data) were observed RMSEP values (x 10-3 in arbitrary units) for the analytes A and B in the order of 6.7 to 47.6; 10.6 to 11.4; and 6.0 to 14.0 for the N-PLS/RBL, Ga-N-PLS/RBL and the proposed method, respectively. On the second case of study (HPLC-DAD data) RMSEP value (mg/L) of 0.72 (N-PLS/RBL); 0.70 (GA-N-PLS/RBL) and 0.64 (iSPA N-PLS/RBL) were obtained. When combined with the U-PLS/RBL, the new algorithm was evaluated in the EEM modeling in the presence of inner filter effect. Simulated data and quantitation of phenylephrine in the presence of acetaminophen in water sample and interferences (ibuprofen and acetylsalicylic acid) were used as a case of studies. The results were compared to the U-PLS/RBL and e twell established method PARAFAC. For simulated data was observed the following RMSEP values (in arbitrary units) 1.584; 0.077 and 0.066 for PARAFAC; U-PLS/RBL and the proposed method, respectively. In the quantitation of phenylephrine the found RMSEP (in μg/L) were of 0.164 (PARAFAC); 0.089 (U-PLS/RBL) and 0.069 (ISPA-U-PLS/RBL). In all cases it was shown that variables selection is a useful tool capable of improving accuracy when compared with the respective global models (model without variables selection) leading to more parsimonious models. It was observed in all cases, that the sensitivity loss promoted by variables selection is compensated by using more selective channels, justifying the obtained RMSEP smaller values. Finally, it was also observed that the models based on variables selection such as the proposed method were free from significant bias at 95% confidence.