A Novel Machine Learning-Based Approach for Adulteration Detection and Quantification in Surimi Using Physicochemical and Elemental Features

Abstract Surimi adulteration poses a significant challenge to food safety. This study developed a machine learning framework that uses 13 physicochemical and elemental features for fish category classification and adulteration-ratio prediction. Four models (RF, XGBoost, LightGBM, and MLP), with logistic regression and ridge regression baselines, were evaluated using 30 source composition records, simulated mixtures and 30 measured validation samples. Five outer record folds and three inner folds separated model selection from evaluation. Principal component analysis (PCA) and SHAP were applied for feature analysis and model interpretation. RF and LightGBM each correctly classified all 30 held-out source records. XGBoost showed the lowest mean internal regression RMSE (0.1061; R² = 0.8799). In the 30 measured validation samples at 0%, 30% and 50%, RF gave the numerically lowest validation error (MAE = 0.0542, RMSE = 0.0791, R² = 0.8517), while the internally selected XGBoost gave MAE = 0.0619, RMSE = 0.0792 and R² = 0.8515. This measured set was not used for model fitting, preprocessing, model selection or noise selection, but its sampling breadth is limited. SHAP analysis identified fat, moisture and zinc as prominent classification features and fat, iron and sodium as prominent regression features. These results support exploratory composition-based modelling; broader independent measured validation is still required for practical regulatory applications.