Physics-based feature engineering for simultaneous classification of bovine meat type and freshness by bioimpedance spectroscopy

• Identifying the type and freshness level of beef where beef is primarily sold in packaged form. • Impedance spectroscopy enables non-invasive meat assessment, existing methods often rely on raw data or complex feature sets, failing to explain the rationale behind their selection or justify their complexity. • A systematic, physics-based feature engineering methodology that can classify muscle type and freshness in bovine meat simultaneously using a single neural network. • Feature selection is primarily based on significant parameter variation measured over a 14-day period, enabling a classification accuracy of up to 91,5% using both neural networks and SVM models. • The evolution of these parameters aligns with biological fundamentals, confirming the relevance of the selected features and underlying physical principles In this paper, we propose a systematic, physics-based feature engineering methodology that can classify muscle type and freshness in bovine meat simultaneously using a single neural network. We identify a minimal, yet highly informative set of features extracted from bioimpedance spectra based on observed spectral behavior and underlying physical principles. Feature selection in this study is primarily based on significant parameter variation measured over a 14-day period, particularly between days 2 or 3 and day 14 for a binary classification of meat in the two classes safe-to-sell and not-safe-to-sell. This variation serves as a strong indicator of sensitivity and contributes to achieving a high classification rate. The key features selected include parameters from the modified Fricke model and characteristic points in the α and β frequency bands (Rex, Rin, C, Im1, Im2 and Im3). We validate the stability of these features and assess the model’s interpretability. Our analysis highlights the consistent relevance of these parameters, enabling a classification accuracy of up to 91,5% using both neural networks and SVM models. The evolution of these parameters aligns with biological fundamentals, confirming the relevance of the selected features.