Machine learning-assisted prediction of beef spoilage using full-spectrum FTIR gas analysis

A rapid and accurate sensing of complex organic and inorganic gases during the meat spoilage is an essential to ensure food safety and public health. In this study, Fourier-transform infrared (FTIR) spectroscopic gas sensing was combined with machine learning tools to predict and evaluate meat spoilage. Unlike traditional methods, this approach utilizes non-selective gas sensing to capture the comprehensive spectral fingerprint of determining analytes, including ammonia, hydrogen sulfide, and acetone, directly from the sample headspace. Spoilage gases emitted from meat samples stored at 4 °C or 25 °C were monitored for 13 days, and key spoilage indicators, including the concentrations of ammonia, hydrogen sulfide, acetone, ethanol, carbon monoxide, and carbon dioxide, were identified and analyzed using multivariate principal component analysis (PCA) and partial least squares (PLS) regression. PCA and PLS analyses suggested spectral variations associated with storage temperature and storage progression under the tested conditions. These effects were distinguished in the extracted spectral features and supported the classification of spoilage progression under the tested 4 °C and 25 °C storage conditions. K-means clustering and k-nearest neighbor models classified the spectra obtained under the tested conditions into fresh, sub-fresh, and spoiled stages, achieving 99.8% accuracy within the experimental dataset. This nondestructive framework demonstrates the methodological feasibility of FTIR-based headspace gas analysis for real-time meat spoilage monitoring and provides a basis for further validation using independent biological replicates. This approach also supports improved food quality management and contributes to reducing unnecessary food waste in the supply chain.