Neural network modeling of carcass measurements to predict beef tenderness

Abstract

Neural network (NN) models were developed for predicting and classifying an objective measurement of tenderness using carcass data such as pre-slaughter information (sex, age, kill order), weights, pH, temperatures, lean color readings, lab-determined measurements, grade measurements and organ weights. Tenderness was expressed objectively as Warner-Bratzler shear (WBS) force measured on steaks, aged 6 d, from the longissimus thoracis et lumborum (LTL) muscle. Carcass data from experiments conducted between 1985 and 1995 at the Lacombe Research Centre were combined to form large data sets (n = 775–1177) for modeling. Neural network models to predict actual shear values showed limited potential (R2 = 0.37–0.45) and were only marginally better than a multiple linear regression (MLR) model (R2 = 0.34). Neural network models that classified carcasses into tenderness categories showed better potential (mean accuracy 51–53%). The best four-category (tender, probably tender, probably tough, tough) model classified tender and tough steaks with accuracies of 0.64 and 0.79, respectively. This model reduced tough and probably tough carcasses by 55% in our population. The model required the following 11 inputs, which, except for cooking method, are available by 24 h postmortem: sex, live plant weight, hot carcass weight, 24-h cooler shrink, 24-h pH, 24-h CIE color b*, 24-h CIE lightness L* × hue angle, rib eye area, grader's marbling score (AMSA%), grade, and cooking method. By implementing techniques outlined in this study in a plant situation, the current 23% unacceptable consumer rating for Canadian beef could be reduced to 10–12%. Key words: Neural networks, beef, tenderness, carcass measurements, longissimus muscle