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

Area of Science:

• Agricultural Chemistry
• Analytical Chemistry
• Food Science

Background:

• Deoxynivalenol (DON) is a significant mycotoxin contaminant in wheat, posing risks to food safety.
• Current methods for DON quantification lack reliability and cost-effectiveness.
• Fusarium head blight is a primary cause of DON contamination in wheat kernels.

Purpose of the Study:

• To develop and validate a cost-effective method for predicting DON concentrations in wheat using ATR-FTIR spectroscopy.
• To compare the performance of four machine learning models for DON quantification and classification.
• To enhance the interpretability of the predictive models using SHAP analysis.

Main Methods:

• Fourier-transform infrared spectroscopy with attenuated total reflectance (ATR-FTIR) was employed for spectral data acquisition.
• Four machine learning models (XGBoost, RF, TabPFN, CatBoost) were trained on spectral features selected via RFE.
• SHapley Additive exPlanations (SHAP) were utilized for model interpretability.

Main Results:

• The TabPFN model achieved the highest quantitative prediction accuracy (R² = 0.86).
• SHAP analysis identified key wavenumbers influencing DON prediction: 996, 1083, 1135, and 1574 cm⁻¹.
• For binary classification of DON levels, the CatBoost model demonstrated superior performance (Recall=0.91, Accuracy=0.88, F2=0.91).

Conclusions:

• Machine learning analysis of FTIR spectra offers an effective approach for DON detection in wheat.
• The proposed ATR-FTIR and ML methodology provides a feasible solution for monitoring DON levels in wheat flour.
• This approach has potential applications in both laboratory and industrial settings for ensuring food safety.