Explainable Machine Learning Enables Quality Prediction and Moderate Drying Optimization of Dried Squid Fillets.

Moderate processing is increasingly required in dried aquatic products to balance various quality attributes. This study developed an explainable machine learning framework integrating a back-propagation artificial neural network (BP-ANN), genetic algorithm (GA), and SHapley Additive exPlanations (SHAP) in dried squid fillets. Four variables were evaluated: lipid-Maillard reaction products (L-MRPs) content, ultrasonic immersion time, drying temperature, and drying time. BP-ANN, support vector regression, random forest, and XGBoost were compared with response surface methodology (RSM). BP-ANN achieved the best overall test performance (R2 = 0.918; RMSE = 0.418), compared with RSM (R2 = 0.853). GA identified 0.62% L-MRPs, 20.18 min ultrasonic immersion, 39.86°C, and 24 h as the optimal conditions. Under these conditions, the experimental TBARS, total color difference, formaldehyde content, and sensory score were 6.76 mg/kg, 25.63, 7.29 mg/kg, and 42.41, respectively. Experimental validation showed the relative errors below 5.76%. SHAP analysis identified that drying temperature and time were dominant for all quality attributes. Finally, the moderately dried squid fillets (MDSF) remained acceptable for 5 days at 25°C, compared with 3 days for dried squid fillets prepared without L-MRPs. Notably, the MDSF maintained acceptable sensory and microbiological quality at least 14 days at 4°C. PRACTICAL APPLICATIONS: Drying conditions strongly affect the quality and shelf life of dried squid products. The machine learning-assisted optimization strategy developed in this study identified processing conditions that reduced lipid oxidation and formaldehyde formation while maintaining desirable sensory quality. This framework enables quantitative and interpretable multi-objective optimization of moderate dried aquatic products.