Predicting beef palatability using conventional and rapid analytical methods: A random forest approach

This study evaluated the ability of traditional and rapid analytical methods to predict consumer beef liking using random forest modelling. Eighty beef samples from four muscles and four animal classes were analysed for intramuscular fat and fatty acid composition (IMF-FA), other meat quality traits (MQ: pH, colour, shear force, cook loss), near-infrared spectroscopy (NIRS), and rapid evaporative ionisation mass spectrometry (REIMS). Consumer sensory evaluation (n = 180) followed Meat Standards Australia protocols, with liking quantified using the composite MQ4 score. Models were developed using these individual datasets and their combinations and evaluated on independent test data. MQ traits provided the highest predictive accuracy of MQ4 (R² = 0.62, RMSE = 7.31), indicating that key meat properties capture most variation in consumer liking. IMF-FA showed moderate performance (R² = 0.43, RMSE = 8.62), reflecting that lipid variables contained some predictive information for MQ4. REIMS achieved R² = 0.55 and RMSE = 8.99 using metabolite and lipid features tentatively associated with post-mortem biochemical processes and flavour precursors. NIRS showed limited predictive performance under the conditions evaluated (R² = 0.17, RMSE = 10.27). No combined model exceeded MQ alone. Relative to IMF-FA alone, adding MQ improved all performance metrics, adding REIMS to MQ + IMF-FA increased R² but not RMSE or MAE. Among the methods evaluated, conventional MQ traits provided the highest predictive performance for consumer palatability. REIMS showed lower predictive ability than MQ traits but warrants further evaluation as a rapid tool for predicting consumer liking. These findings reflect the relative performance of the datasets for predicting MQ4 within the population and experimental conditions evaluated and require validation using larger independent datasets.