Synergistic microfluidic processing and feature selection for improved Raman detection of mycotoxins in wheat

To address strong matrix interference and weak signals in ppb-level Raman detection of mycotoxins in wheat, this study developed a synergistic strategy combining the microfluidic "hard processing" with the feature selection "soft processing". A polydimethylsiloxane microfluidic chip was designed to separate and enrich toxins based on solubility differences. Combined with the Monte Carlo sampling - weighted bootstrap sampling - least absolute shrinkage and selection operator feature selection algorithm, specific spectral information was extracted to build a partial least squares regression model. The results demonstrate that this strategy effectively eliminates matrix interference from starch and proteins, and the estimated limits of detection for all four mycotoxins are below the Chinese national regulatory limits (GB 2761-2017), with test set coefficients of determination all exceeding 0.92. Feature dimensionality was reduced from thousands to under one hundred, focusing on mycotoxin core functional group vibrational peaks. This approach balances accuracy, stability, and efficiency. With further validation across multiple scenarios, it holds promise for development into a standardized detection solution for such trace contaminants.