Two-stage machine learning for predictive assessment of surimi processing suitability

In conventional surimi manufacturing, physicochemical testing is primarily conducted on finished products, restricting the prospective quality assessment of freshwater fish raw materials. This study developed an interpretable two-stage machine learning framework to predict surimi processing suitability. Stage 1 mapped 26 physicochemical features of raw materials to intermediate quality attributes, and Stage 2 classified these intermediate quality attributes into three final suitability grades. Under a stratified 5-fold nested cross-validation (NCV) scheme, the Extra Trees (ET) algorithm exhibited the best performance, enabling the end-to-end framework to effectively predict processing suitability. SHapley Additive exPlanations (SHAP) analysis identified oxidation-prone n-3 polyunsaturated fatty acids (n-3 PUFAs), particularly α-linolenic acid (ALA, C18:3 n-3), docosahexaenoic acid (DHA, C22:6 n-3), and eicosapentaenoic acid (EPA, C20:5 n-3), as the primary physicochemical factors limiting gel formation, while the n-6 PUFA eicosadienoic acid (C20:2 n-6) exhibited a positive contribution. Furthermore, gel strength, water-holding capacity (WHC), and chewiness were the most critical determinants for the final grading. Based on these findings, practical interventions such as targeted antioxidant addition and rapid heating strategies are proposed for high-risk batches. This framework provides a quantitative and mechanism-informed approach for raw material screening and quality control in the surimi industry.