Rapid determination of methamphetamine content using mid-infrared spectroscopy with chemical prior knowledge

This study developed an interpretable mid-infrared (MIR) spectroscopic method for the rapid and non-destructive quantification of methamphetamine in authentic seized samples. A total of 61 methamphetamine powder samples were analyzed, including 51 samples for model development and 10 independent samples for practical validation. After Hotelling's T² outlier screening and spectral inspection, 43 samples were retained as the model-development dataset for repeated calibration and validation. To reduce baseline drift, random noise, and heterogeneous scattering from complex illicit drug matrices, wavelet transform denoising combined with quantile regression-based multiplicative scatter correction (WT-QMSC) was adopted as the preprocessing strategy. Chemical Prior Regions (CPR), including analyte-related functional-group regions and diagnostically relevant matrix-interference regions, were integrated with a CNN-Attention variable-ranking framework, while the final quantitative prediction was performed using partial least squares (PLS). Model robustness was evaluated using 100 repeated stratified random train-test splits with nested internal optimization. The proposed CPR-CNN-Attention-PLS model selected 160 ± 13 variables and achieved an RMSECV of 2.21 ± 0.35% (w/w), R_CV² of 0.98 ± 0.01, RMSEP of 3.52 ± 0.56% (w/w), R_P² of 0.97 ± 0.02, and RPD of 5.72 ± 0.23. Practical validation with 10 independent seized samples yielded a bootstrap RMSEP of 3.28 ± 0.70% (w/w) and a mean bias of +0.54% (w/w). These results indicate that chemically guided MIR-PLS provides a rapid, interpretable, and promising screening approach for methamphetamine quantification, although further validation using larger and more heterogeneous forensic cohorts is still required.