Machine learning to understand relationships between farm practices and milk fatty acids across diverse European dairy farms

ABSTRACT: Milk fatty acid (FA) composition is an indicator of both farm management and the nutritional quality of dairy products. Few studies have linked diverse, multicountry observational farm data to milk FA variation through a validated machine learning workflow. We surveyed 75 European farms representing a broad gradient of production intensity, analyzed seasonally pooled bulk milk samples for 12 FA traits, and examined 29 management practices. A 2-stage workflow combined optimized random forests (RF) to predict FA and rank practices, with conditional inference trees to visualize management synergies and trade-offs. RF models achieved high predictive accuracy (R2 ≥ 0.50) for 8 traits: α-linolenic acid, eicosapentaenoic acid, docosapentaenoic acid, CLA, n-6:n-3 PUFA ratio, linoleic acid, vaccenic acid (VA), and branched-chain fatty acids (BCFA). Conditional inference trees models had comparable accuracy (R2 ≥ 0.50) for all these traits except VA and BCFA. Across models, fresh grass intake, maize silage and concentrate use, stocking rates, herd size, milk yield, and mineral fertilizer were dominant drivers, together explaining most variance in the models. Farms adopting low-input, pasture-based strategies were consistently associated with lower n-6:n-3 PUFA ratios and higher n-3 PUFA, CLA, and BCFA in milk, highlighting synergies alongside trade-offs between production intensity and nutritional quality. Although this profile is associated with favorable health outcomes and contributes to meeting dietary recommendations, further targeted validation is needed to confirm generalizability and adaptability across dairy production contexts.