AI‐Driven Patulin Detection in Apple Using Machine Learning Coupled With Surface‐Enhanced Raman Spectroscopy

This study presents a comprehensive evaluation of machine learning approaches for surface‐enhanced Raman spectroscopy (SERS)–based prediction of patulin (PAT) concentration in apples, a critical food safety concern due to its toxic effects on health and limitations of conventional laboratory‐based detection methods. Three base algorithms, extreme learning machine (ELM), random forest (RF), and support vector machine (SVM), were implemented in combination with three variable selection methods: Competitive Adaptive Reweighted Sampling (CARS), Uninformative Variable Elimination (UVE), and Genetic Algorithm (GA). The investigation focused on optimizing predictive accuracy while reducing computational complexity through effective variable selection. Results demonstrated that SVM‐based methods achieved superior performance. UVE‐SVM and GA‐SVM exhibited exceptional predictive capabilities with Residual Predictive Deviation (RPD) values of 4.0487 and 4.3036, respectively, and outstanding calibration ( R _c > 0.99) and prediction ( R _p > 0.97) coefficients. This study provided valuable insights for optimizing SERS spectral analysis in food safety applications, recommending UVE‐SVM or GA‐SVM implementations for the highest accuracy in PAT concentration prediction while emphasizing the crucial role of variable selection in enhancing model performance.