Explainable machine learning-assisted Raman spectroscopy-based fast Versicolorin A detection in maize for early...

Understanding Versicolorin A contamination levels in agricultural products is crucial for implementing preventive interventions before severe aflatoxin formation. However, traditional Versicolorin A detection methods cannot meet high-throughput, on-site screening requirements of grain supply chains. While near-infrared spectroscopy has been developed for rapid Versicolorin A detection recently, its accuracy is limited by moisture interference and matrix overlap absorption. This study developed a non-destructive Versicolorin A detection method integrating Raman spectroscopy with machine learning algorithms. Using 208 maize samples, the Random Forest model achieved optimal performance with a cross-validated R² of 0.862, root mean square error of 11.78 μg/kg, and residual predictive deviation of 2.69. SHAP analysis identified 681.67 cm⁻¹ as the most influential spectral feature, which may reflect contamination-associated changes in the maize matrix. For classification tasks, XGBoost achieved the highest average precision of 0.919 with a 50 μg/kg threshold. This method enables reliable quantitative detection without chemical pretreatment, providing technological support for food safety monitoring systems.