Machine learning to identify farming practices influencing antioxidant levels and lipid oxidation in meat from diverse European poultry production

This study aimed to identify which farming practices impact antioxidant levels and lipid oxidation in chicken breast and thigh meat. A total of 89 chicken farms, representing diverse production systems, were selected from five European countries involved in the H2020 INTAQT project. On day two post-slaughter, breast and thigh meats were excised from 20 random carcasses per farm, minced, and pooled per farm and meat type. These samples were analysed for α-tocopherol, carnosine, anserine, and thiobarbituric acid reactive substances (TBARS). Twenty farming practices were identified from a farm questionnaire to capture variation in growth performance, feed characteristics, housing conditions, and management strategies, and were used as features in machine learning models. Our workflow combined optimized Random Forests (RF) to predict each trait level from 20 features and rank the feature importance using permutation-based scores, with Conditional Inference Trees (CIT) to derive interpretable decision trees. Models were trained and validated using a train–test split, with hyperparameters tuned via cross-validation. RF models were excellent (R² ≥ 0.90) for carnosine and anserine and fair (0.30 < R² ≤ 0.50) for α-tocopherol and TBARS in breast. RF models were good (0.50 < R² ≤ 0.75) for carnosine, anserine and TBARS, but poor (R² ≤ 0.30) for α-tocopherol in thigh. In breast, age at slaughter was the most important feature for carnosine, and average daily gain (ADG) for anserine; in thigh, ADG was the top feature for carnosine, and age at slaughter for anserine. Vitamin E in feed and ADG were the strongest features of α-tocopherol and TBARS, respectively, in both breast and thigh. In breast, CIT model fits were excellent for carnosine and anserine, and good and fair for α-tocopherol and TBARS, respectively. In thigh, CIT model fits were good for carnosine, fair for anserine and α-tocopherol, and poor for TBARS. In CIT models, outdoor access was the most discriminating practice for breast carnosine and anserine and for thigh carnosine, whereas age at slaughter was the most discriminating variable for thigh anserine. In both breast and thigh, vitamin E in feed was the most discriminating practice for α-tocopherol, and genotype was for TBARS. These findings should be validated using larger and more complex datasets, as the order of importance may change with the inclusion of additional farming practices.