Predicting and classifying deoxynivalenol in wheat flour using ATR-FTIR spectroscopy and explainable machine learning

Deoxynivalenol (DON) is a major hazardous component in wheat kernels infected by Fusarium head blight. Given the lack of both reliable and cost-effective quantification methods, this study aims to use Fourier-transform infrared spectroscopy with attenuated total reflectance (ATR-FTIR) for predicting DON concentrations. Four machine learning models, including eXtreme Gradient Boosting (XGBoost), Random Forest (RF), Tabular Prior-data Fitted Network (TabPFN), and Categorical Boosting (CatBoost), were firstly developed using key spectral features selected by recursive feature elimination (RFE). TabPFN model outperformed others in quantitative prediction, achieving a coefficient of determination (R²) of 0.86 on hold-out test set. SHapley Additive exPlanations (SHAP) were firstly applied to enhance model interpretability and transparency, identifying wavenumbers at 996 cm^-1, 1083 cm^-1, 1135 cm^-1, and 1574 cm^-1 as having higher influence on model performance. For binary classification of DON levels (<1 or ≥1 mg/kg), Catboost model achieved superior results, yielding a recall of 0.91, accuracy of 0.88, and an F2 score of 0.91. This study highlights the effectiveness of machine learning based analysis of FTIR spectra. Furthermore, the proposed approach herein could be a feasible way for monitoring DON levels in wheat flour, with potential applications at both laboratory and industrial scales to ensure food safety.