Machine learning-based malachite green and leucomalachite green content in large-mouth bass prediction using Raman spectroscopy

Abstract

Carcinogenic malachite green (MG) and its metabolite leucomalachite green (LMG) in aquatic products present significant food safety risks. Conventional detection methods for MG and LMG, such as the liquid chromatography-tandem mass spectrometry (LC-MS/MS), are often time-consuming and destructive due to their laborious multi-step sample extraction and purification procedures, thereby hindering rapid screening in the aquaculture industry. This study addresses this need by developing a non-destructive method using Raman spectroscopy combined with machine learning to predict MG and LMG concentrations in large-mouth bass muscle tissue. Muscle samples were analyzed at multiple time points (0–480 min) post-MG/LMG exposure. Among evaluated preprocessing techniques (Savitzky-Golay smoothing, multiple scatter correction, standard normal variate, normalization), Savitzky-Golay smoothing proved most effective. Four machine learning regression models were tested: Partial Least-Squares Regression (PLSR), Random Forest Regression (RFR), XGBoost, and Back Propagation Neural Network (BPNN). RFR yielded the best predictions for MG (validation R² = 0.8090), while PLSR was optimal for LMG (validation R² = 0.8435). This Raman/ML approach enables rapid, non-invasive quantification of these banned residues, offering a practical tool for enhanced food safety monitoring in aquaculture.