Qualitative and quantitative analysis of microplastics in chicken based on portable near-infrared spectroscopy and interpretable machine learning

Portable near-infrared (NIR) spectroscopy combined with interpretable machine learning was employed for the qualitative and quantitative analysis of polypropylene (PP), polyvinyl chloride (PVC), and polyethylene terephthalate (PET) in chicken. A total of 244 samples were prepared, including pure chicken and PP-, PVC-, and PET-spiked chicken (61 samples per type, mass fractions: 0.01-0.8%). NIR spectra were acquired over 900-1700 nm. After optimal preprocessing and outlier removal via Monte Carlo cross-validation (MCCV), competitive adaptive reweighted sampling (CARS) was selected over the successive projections algorithm (SPA) for characteristic wavelength extraction prior to modeling. For qualitative analysis, a data-driven soft independent modeling of class analogy (DD-SIMCA) model was constructed, achieving sensitivities of 85.71-92.86% and specificities of 82.93-100%. For quantitative analysis, the optimal model varied with microplastic (MP) type: CARS-SVR (support vector regression) for PET (R_p² = 0.8923, RPD = 3.0468); CARS-ElasticNet for PP (R_p² = 0.8267, RPD = 2.4022), and CARS-PLS (partial least squares) for PVC (R_p² = 0.8969, RPD = 3.0397). SHAP-based interpretability analysis was further employed to quantify each wavelength's spectral contribution and identify MP-specific absorption bands. This study innovatively integrates portable NIR spectroscopy with SHAP-based interpretable machine learning for MP detection in chicken and demonstrates the potential of this strategy for rapid on-site screening of MP contamination in poultry products.