Machine learning for polycyclic aromatic hydrocarbons analysis in roasted lamb: new insights from spectral and chemical data.

Abstract

Polycyclic aromatic hydrocarbons (PAHs) generated during lamb roasting pose health risks but are difficult to predict due to their low concentrations and complex features. Existing models fail to address data scarcity and low-concentration prediction accuracy in PAHs analysis. A comprehensive PAHs index (CPI) derived via entropy weighting method (EWM) combined with auto-encoders and generative adversarial networks (AE-GAN), may overcome these limitations. We integrated EWM with AE-GAN (for synthetic data augmentation) and developed prediction models using partial least squares regression (PLSR), convolutional neural network (CNN) and Bayesian echo state network (Bayes-ESN). When epoch = 3000 and increasing the amount of data by 750, the Bayes-ESN model worked the best (RP2 = 0.8238), demonstrating AE-GAN's efficacy in mitigating data scarcity. The combination of EWM, AE-GAN and Bayes-ESN provides a robust solution for predicting the levels of PAHs in roasted meat, advancing food safety control through enhanced generative-regression synergy.