Separating prediction from maturity in cooked cattle hide: confound-partitioned machine learning reveals a bounded thermal-lability axis

Abstract

Small food-quality datasets can yield high machine-learning coefficients when a discrete biological factor separates samples, even if individual-level prediction is weak. Using cooked cattle hide as a collagen-dominated test system, we evaluated whether instrumental models predict trained-panel sensory quality beyond dentition class and whether predictive relationships persist across maturity domains. Neck hides from 24 male Sahiwal-crossbred cattle (two-tooth, n=12; four-tooth, n=12) was processed identically and characterized for pH, CIE L*a*b*, drip and cooking loss, Warner-Bratzler shear force, proximate composition and five sensory attributes. A seven-rung model ladder was evaluated by nested leave-one-animal-out cross-validation against intercept-only and dentition group-mean benchmarks. PC1 explained 78.70% of variance and the five core predictors were strongly collinear (VIF 11.15-43.03). Best cross-validated R² values were 0.980 for juiciness, 0.976 for texture and overall acceptability, 0.962 for color and 0.772 for flavor. For PLSR, fold-safe class centering reduced R² by only 5-27% (Φ=0.05-0.27), but every cross-class transfer had negative R². Exact retraining Shapley attribution across four physical blocks was nearly uniform (22.3-27.6%), precluding a unique instrumental driver. Cooking loss and shear force were correlated in younger hide (r=0.905) but not mature hide (r=-0.146); the correlations differed significantly (Fisher Z=3.49, P=0.00048). Thus, sensory prediction is substantial within the observed maturity structure but is not transportable across it, and the proposed thermal-lability axis is explicitly maturity-bounded.