Rapid detection of trace dibutyl phthalate in refined cottonseed oil using surface-enhanced Raman spectroscopy coupled with optimized machine learning models.

Rapid screening of dibutyl phthalate (DBP) in edible oils is complicated by lipid-derived background and refining-stage-dependent spectral variation. We developed a surface-enhanced Raman spectroscopy (SERS)-chemometric workflow using Au-coated Si nanocone arrays and 180 cottonseed-oil samples from 30 production batches at six refining stages. Each spectrum was paired with a sample-specific UHPLC-MS/MS reference concentration. Full-spectrum partial least-squares regression (PLSR) served as the latent-variable baseline and was compared with multiple linear regression (MLR), ridge regression (RRM), and recursive feature elimination with cross-validation coupled to ridge regression (RFECV-RRM). Stage-based leave-one-group-out (LOGO) and production-batch-grouped GroupShuffleSplit (GSS) validation were used, with all supervised preprocessing and feature selection confined to the corresponding outer-training data. RFECV-RRM performed best under LOGO (RMSE = 19.640 ± 2.432 ng g-1, R2 = 0.918 ± 0.013, RPD = 3.569 ± 0.252) and GSS (RMSE = 35.153 ± 4.369 ng g-1, R2 = 0.876 ± 0.030, RPD = 2.953 ± 0.370). Relative to PLSR, RFECV-RRM reduced mean RMSE by 54.9% under LOGO and 24.3% under GSS. Mean matrix-spike recovery was 98.1%. The workflow is suitable for evaluation as an at-line screening approach, subject to independent refinery, substrate-lot and instrument validation.