EXPRESS: Comparison of Linear and Nonlinear Regression Models for Dual ED-XRF Analysis of Trace Arsenic and Lead in...

Applications of energy dispersive X-ray fluorescence (ED-XRF) spectrometry for trace analysis of arsenic (As) and lead (Pb) in foods are limited by spectral interference around characteristic emission lines. To overcome this limitation, various chemometric data modeling approaches can be applied; however, dataset characteristics and model structure can significantly impact these approaches in their ability to accurately relate spectral features to analyte concentrations. This study compared linear and nonlinear multivariate regression models (partial least squares regression (PLSR) and artificial neural network (ANN), respectively) to model the spectral overlap and enable simultaneous analysis of trace As and Pb in rice-based foods. Calibration standards were developed through pelletization of powdered rice (4.0 ± 0.2 g) spiked with As and Pb (∼0-600 µg kg^-1). Standards were analyzed via ED-XRF and resultant spectra were used to develop calibration models utilizing PLSR and ANN modeling. To assess method accuracy and applicability in real-world scenarios, a certified reference material of rice, 11 commercial rice-based products, and 12 Pb-spiked commercial foods were analyzed. PLSR yielded stronger calibration performance (R²_As = 0.98, R²_Pb = 0.93) compared to ANN; however, ANN modeling achieved lower quantification limits for As in samples containing detectable and non-detectable Pb content (53 µg kg^-1 and 54 µg kg^-1, respectively). Both models achieved successful validation in the analysis of the certified reference material with errors of +12.7 % (PLSR) and -19.4 % (ANN) in the prediction of As. PLSR modeling demonstrated strong predictive capabilities in commercial rice-based samples containing non-detectable Pb content, obtaining a median absolute error of 10.7 % in As quantification and 8 of 11 samples yielding less than 20 % error. Conversely, in Pb-spiked commercial samples (76-417 µg kg^-1 Pb), the ANN outperformed PLSR, obtaining an average absolute error of 14.1 % with 8 of 12 samples yielding less than 20 % error in As determinations. PLSR modeling correctly identified low Pb content in samples containing non-detectable Pb content while also yielding 75 % accuracy in the classification of samples containing Pb levels above current regulatory limits, supporting its use as a tool for semi-quantitative screening of Pb content in rice/rice-based foods at levels aligning with current regulations. This study compared PLSR and ANN modeling in their ability to statistically model and overcome As/Pb spectral overlap in ED-XRF, highlighting their advantages and limitations in specific analytical scenarios and demonstrating their potential to enable trace As and Pb screening in rice-based foods.