Interpretable multi-machine learning model for grading chicken wooden breast: Integrating physicochemical analysis and SHAP-driven feature importance

Regarding the issue of wooden breast in chicken meat within the poultry industry, this study systematically assessed the quality differences across varying degrees of wooden breast using multi-index analysis. The evaluation encompassed parameters such as pH, color attributes, water-holding capacity, textural characteristics, and electronic nose data. Additionally, a collaborative strategy involving machine learning models was proposed to construct an effective classification model. The breast meat samples from Kebao white feather broilers were selected and graded as four groups based on tactile and visual criteria: Normal Breast Meat (NB), Mild Wooden Breast (LWB), Moderate Wooden Breast (MWB), and Severe Wooden Breast (SWB). The quality indicators were changed with the degree of wooden breasts ( P <0.05). Among the machine learning models, the BP-ANN model exhibited the highest performance with a classification accuracy of 95.83%, outperforming SVM (91.67%) and PLS-DA (87.50%) models. Through Shapley Additive Explanations (SHAP)-based explainability analysis, in the group of SWB, gumminess and chewiness emerged as dominant factors, which were associated with reduced processing performance due to myofibrillar densification and structural reorganization. The present study focuses on developing rapid detection and grading technologies for chicken wooden breast, providing a theoretical foundation and technical support for its rapid identification and quality control. • The BP-ANN model achieved 96.88% accuracy in classifying wooden breast in chicken meat. • SHAP analysis identified key indicators driving classification, enhancing model interpretability. • Multi-index integration enables robust, non-destructive wooden breast assessment.