Explainable artificial intelligence-based analysis of milk composition in subclinical mastitis caused by different bacterial species

Abstract

Background, Context or Rationale Subclinical mastitis induces substantial biochemical and physicochemical alterations in milk, negatively affecting milk quality, processing characteristics and dairy production efficiency. Conventional culture-based diagnostic methods remain labour-intensive and time-consuming, emphasising the need for rapid, data-driven and interpretable diagnostic approaches based on routinely obtainable milk parameters.Aim(s) This study aimed to investigate the biochemical effects of mastitis on milk composition and to identify pathogen-specific alterations associated with Escherichia coli (E. coli), Streptococcus agalactiae (S. agalactiae) and Staphylococcus aureus (S. aureus) using explainable machine learning methodologies.Methods Machine learning models were developed to (i) discriminate infected from noninfected milk samples and (ii) classify mastitis according to the causative bacterial species using fundamental physicochemical milk variables. Several algorithms were comparatively evaluated, and tree-based ensemble approaches demonstrated superior and more robust generalisation performance. Model interpretability was achieved using SHapley Additive exPlanations (SHAP), enabling the association of predictive outcomes with underlying biochemical characteristics of milk.Major Findings The SHAP analyses revealed distinct pathogen-dependent biochemical signatures. E. coli mastitis were primarily characterised by elevated electrical conductivity (EC), accompanied by increased somatic cell count (SCC). In S. agalactiae mastitis, SCC emerged as the dominant discriminative feature, with EC providing additional predictive contribution. In contrast, S. aureus mastitis were mainly associated with alterations in milk solid fractions, particularly solids-nonfat (SNF) and related compositional parameters. When classification was simplified to infected versus noninfected milk, fat content and SCC represented the most influential discriminative variables within the present dataset.Scientific or Industrial Implications The findings demonstrate that routinely measurable milk composition parameters can support rapid and interpretable identification of mastitis and differentiation of likely causative pathogens. By transforming conventional milk quality variables into pathogen-specific biochemical signatures, the proposed framework provides a practical and explainable decision-support approach for subclinical mastitis surveillance and milk quality assessment in dairy production systems.