Multiclass machine learning classification of aflatoxin B1 and ochratoxin A in crude palm oil using SERS with...

Mycotoxin contamination in crude palm oil poses significant food safety challenges, yet conventional detection methods remain time-consuming and resource-intensive. This study presents a rapid analytical framework for the simultaneous detection of aflatoxin B1 (AFB1) and ochratoxin A (OTA) by integrating Au@Ag core-shell nanoparticles as surface-enhanced Raman scattering (SERS) substrates with modified dispersive solid-phase extraction and a statistically validated machine-learning classification model. Monodisperse Au@Ag nanoparticles (15 ± 2 nm cores, 7 ± 1 nm shells) were synthesized and characterized, confirming successful core-shell formation and surface plasmon resonance shifts that enable optimal electromagnetic enhancement at 785 nm excitation. The modified extraction protocol effectively removed lipid matrix interferents while maintaining recoveries of 95.1-110.8%, achieving detection limits of 0.94 pg/mL for AFB1 and 0.93 pg/mL for OTA-3-4 orders of magnitude below regulatory thresholds. Spectral analysis revealed orientation-dependent enhancement mechanisms, with consistently enhanced bands at 1369 cm⁻¹ (AFB1) and 1578 cm⁻¹ (OTA) enabling simultaneous detection in binary mixtures through complementary spectral windows. Four supervised machine learning algorithms (PLS-DA, LDA, k-NN, SVM) were systematically evaluated, with linear projection methods achieving near-perfect test accuracies (0.990-1.000) using only 3-10 principal components. Rigorous nonparametric statistical validation using Friedman and Dunn's tests established that PLS-DA and LDA are statistically equivalent and optimal classifiers, significantly outperforming nonlinear approaches. The validated two-tier screening workflow showed no statistically significant differences relative to LC-MS/MS reference measurements, supporting its potential for high-throughput mycotoxin screening in lipid-rich food matrices.