Development of portable NIR calibration models for microplastic detection and quantification in corn silage

Microplastic (MPs) feed contamination represents a significant risk to the entire food chain, with potential human exposure as MPs can cross intestinal barriers, enter the bloodstream, or accumulate in animal- derived products [1,2]. Currently, no real-time methods exist that are simple, rapid, and reliable for detecting MPs in animal feed. This study evaluates the potential of Near-Infrared Spectroscopy (NIRS) as a tool for detecting and quantifying MPs in animal feed under field conditions, without MP extraction process. Two types of MPs - white polystyrene (PS) and black low-density pol- yethylene (LDPE) - were used to contaminate corn silage at different concentrations (0, 0.5, 1, 3, 5, 10, and 50.0 mg/g dry matter - DM). MPs of varying sizes were obtained by grinding plastic pieces (< 5mm). A total of 13 grinded corn silage samples (6+6+1) were scanned 6 times each, with 4 scans per session, resulting in 312 spectra. Scanning was performed using the Aurora portable NIR spectrometer (GraiNit S.r.l., Italy) across a spectral range of 950–1650 nm. Calibration models were built using 70% of the dataset and validated on the remaining 30%, us- ing Ucal™ (Unity Scientific, Australia) and grouping the dataset by PS, LDPE and their combination (PS+LDPE). The Standard Normal Vari- ate pre-processing method combined with detrend was applied to five different chemometric approaches: PLS, PLS+PCA, PCR, MLR+PLS, and MLR+PCA [3]. Models were assessed by calculating the coefficient of determinations (R²), Standard Error of Calibration (SEC) and Predic- tion (SEP), Mean Absolute Error (MAE), the distribution of the residu- als and power prediction (PP). The PLS+PCA showed the best results for each MP type. LDPE exhib- ited an excellent calibration fit (R²=0.93, using 11 factors) and a good validation fit (R²=0.89). The standard errors were low and consistent (SEC=4.62 mg/g DM, SEP=5.22 mg/g DM), and the MAE were very low (1.9×10-6 mg/g DM and 0.3 mg/g DM, respec- tively), though calibration residuals showed a non-normal distribution. PS also showed a good fitting calibration (R²=0.92, 12 factors) but slightly lower prediction performance (R²=0.85), with a higher SEP (6.3 mg/g DM) compared to SEC (4.5 mg/g). However, the MAEs remained low (2.7×10-7 mg/g DM in calibration and 0.40 mg/g DM in validation) but the residual distribution was non-normal in the validation phase. The model for the prediction of the contamination by PS+LDPE MPs showed the weakest performance, with a fair calibration fit (R²=0.89, 14 factors) but a lower prediction fit (R²=0.75). Both the standard errors (SEC=5.7 mg/g DM and SEP=8.9 mg/g DM) and MAE (5. ×10-7 mg/g DM in calibration and 1.5 mg/g DM in validation) were higher than those observed for the models developed per single MP type. The three models showed a good PP, always higher than 3.0. In conclusion, portable NIRS effectively detect MP contamination in animal feed, with the PLS+PCA model showing the best results, particularly for LDPE. The findings underscore the possibility of analysing individual MPs (PS and LDPE), as this approach yields better predictive performance compared to combined models (PS+LDPE) and the portable NIRS can be a promising tool for the real-time monitoring directly MPs in corn silage without extraction.