Rapid and simultaneous detection of Aspergillus and aflatoxin B1 content in maize using data fusion and explainable analysis

Abstract

Maize is highly susceptible to Aspergillus flavus infection, which can subsequently lead to aflatoxin B₁ (AFB₁) contamination during storage, posing serious threats to food safety and human health. This study developed a simultaneous detection method by integrating near-infrared spectroscopy (NIRS) with electronic nose (E-nose) technologies. Random forest (RF), support vector regression (SVR), and feedforward neural network (FNN) models were constructed, with data fusion and feature selection strategies applied. The results demonstrated that the successive projections algorithm (SPA)-FNN model based on mid-level fusion significantly outperformed single-modality models, achieving superior results with $R_P^2$ of 0.9595 and 0.9867, RMSEP of 0.2408 log CFU/g and 24.7404 μg/kg, and RPD of 4.9671 and 8.6701 for fungal colony counts and AFB₁ contents. Crucially, the introduction of Shapley Additive exPlanations (SHAP) analysis transitioned the deep learning model from a “black box” into an interpretable system. It revealed that the predictions are driven by specific biochemical mechanisms, specifically substrate degradation and oxidative stress-induced volatiles, thereby validating the scientific basis of the detection logic. This study provides a rapid, accurate, and mechanistically interpretable framework for maize quality monitoring, laying a solid foundation for the future development of portable, on-site detection devices.