Next Article in Journal
Molecular Characterization of Avulaviruses Isolated from Mallard Ducks in Moscow in 2008–2024
Next Article in Special Issue
Maternal Tributyrin Supplementation During the Perinatal Period Is Associated with Improved Ewe Milk Quality and Lamb Growth Performance, Immunity, and Antioxidant Status
Previous Article in Journal
Pulmonary Vascular Proliferative Lesions in Wild Korean Raccoon Dogs (Nyctereutes procyonoides): Description of 13 Cases
Previous Article in Special Issue
Good Veterinary Practices for Managing the Colostral Period in Dairy Calves to Improve Their Immune Competence—A Systematic Review
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Veterinary Sciences
Volume 13
Issue 1
10.3390/vetsci13010022
vetsci-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Milk Biomarkers and Herd Welfare Status in Dairy Cattle: A Machine Learning Approach

by
Daniela Elena Babiciu
1,
Anamaria Blaga Petrean
1,*,
Sorana Daina
1,
Daniela Mihaela Neagu
1,*,
Eva Andrea Lazar
2 and
Silvana Popescu
1
1
Faculty of Veterinary Medicine, University of Agricultural Sciences and Veterinary Medicine Cluj-Napoca, 400372 Cluj-Napoca, Romania
2
Horse Welfare Association, 407207 Feiurdeni, Romania
*
Authors to whom correspondence should be addressed.
Vet. Sci. 2026, 13(1), 22; https://doi.org/10.3390/vetsci13010022
Submission received: 1 December 2025 / Revised: 23 December 2025 / Accepted: 24 December 2025 / Published: 25 December 2025
(This article belongs to the Special Issue From Barn to Table: Animal Health, Welfare, and Food Safety)
Browse Figure
Versions Notes

Simple Summary

Evaluating animal welfare on dairy farms is essential for ensuring good management, consumer confidence and sustainable milk production. However, comprehensive protocols such as Welfare Quality® (WQ®) are time-consuming and require trained assessors, limiting their frequent use on farms. This study explored whether commonly recorded milk biomarkers, which are routinely measured on most dairy farms, can reflect herd welfare status. We assessed 43 dairy farms using the full Welfare Quality® (WQ®) protocol and compared welfare outcomes with common milk biomarkers such as somatic cell count (SCC), differential somatic cell count (DSCC), lactose, urea, fat-to-protein ratio and milk yield. Farms with higher welfare scores showed lower somatic cell indicators, higher lactose concentrations and increased milk yield. We also tested six machine learning models to see whether milk data alone could classify farms into welfare categories. The best models correctly identified farm welfare status in about 70% of cases. Across all models, SCC, DSCC, lactose and milk yield were the most important predictors. These findings should be regarded as preliminary and exploratory, providing initial evidence rather than immediate tools for routine welfare assessment or on-farm decision-making.

Abstract

Routine milk-recording data may provide valuable insights into dairy cow welfare, although their ability to accurately reflect herd-level welfare outcomes remains unclear. This study explored the associations between routinely collected milk biomarkers and farm-level welfare status using a comparative machine learning approach. Using the Welfare Quality® (WQ®) protocol, 43 commercial dairy farms were classified as Enhanced, Acceptable, or Not Classified. Farm-level milk variables included somatic cell count (SCC), differential somatic cell count (DSCC), fat-to-protein ratio (FPR), fat, protein, casein, lactose, urea, β-hydroxybutyrate (BHB), acetone, total plate count (TPC), and morning milk yield. Kruskal–Wallis tests revealed significant differences among welfare classes for DSCC, SCC, lactose, and milk yield (False Discovery Rate-adjusted p < 0.05). Six machine learning algorithms were trained using 10-fold stratified cross-validation. The Elastic-Net (ENET) model showed the highest mean performance (Accuracy = 0.72 ± 0.19; Kappa = 0.56 ± 0.31), followed by Random Forest and Multilayer Perceptron (Accuracy = 0.70). Model accuracy exhibited substantial variability across cross-validation folds, reflecting the limited sample size and class imbalance. Across models, the most influential variables were SCC, DSCC, lactose, milk yield, FPR, fat, and urea. Overall, the findings provide preliminary and exploratory evidence that routine milk biomarkers capture welfare-relevant patterns at the herd level, supporting their potential role as complementary indicators within data-driven welfare assessment frameworks.

1. Introduction

Animal welfare is a central dimension of sustainable dairy production and a key driver of consumer confidence in animal-derived foods. According to the World Organisation for Animal Health, animal welfare encompasses the physical and mental state of an animal in relation to the conditions in which it lives [1]. Increasing public concern regarding welfare-related issues has shaped both market expectations and policy frameworks, underscoring the need for transparent, robust, and science-based welfare assessment tools [2]. In this context, Welfare Quality (WQ®) protocol represents a comprehensive and validated framework for evaluating welfare in dairy cattle through the integration of animal-based and resource-based measures [3]. However, the protocol includes more than 30 individual animal-based and resource-based indicators and typically requires approximately 6–8 h per farm assessment. This level of complexity and time investment, combined with the need for trained assessors, substantially limits the feasibility of frequent or routine application on commercial dairy farms [4,5]. These constraints have stimulated interest in identifying complementary and easily obtainable indicators that may support the early detection of welfare issues or serve as preliminary screening tools.
Milk composition represents a particularly promising source of routinely available information, as it is recorded in most dairy production systems through regular milk-recording schemes conducted monthly or even more often, with data collected on each specific recording date (test day), and reflects key aspects of udder health, metabolic balance, and management. Biomarkers such as the fat-to-protein (FPR), milk urea, somatic cell count (SCC), and differential somatic cell count (DSCC) have been linked to energy balance, nitrogen utilisation, and udder health status [6,7,8]. In addition, traits such as milk yield and lactose concentration provide insight into metabolic efficiency and overall herd management quality [9]. Collectively, these routinely measured variables constitute a non-invasive, high-frequency data source with potential relevance for welfare-related monitoring at the herd level.
It is important to emphasise that milk biomarkers primarily reflect physiological and health-related domains of animal welfare, particularly those related to udder health, metabolic status and productivity. As such, they provide indirect information on health- and management-related welfare aspects but cannot directly capture behavioural expressions or affective states, which are essential components of holistic welfare assessment frameworks such as WQ® [3]. Consequently, milk biomarkers should be regarded as complementary indicators rather than substitutes for comprehensive welfare evaluation protocols.
At the same time, machine learning techniques have gained increasing relevance in veterinary and dairy science research due to their ability to model complex, non-linear relationships typical of biological and farm-level data. Machine learning approaches have been applied to classify udder health based on somatic cell indicators [10], explore mastitis dynamics [11], and detect metabolic disorders such as subacute ruminal acidosis using milk mid-infrared spectral profiles [12]. Deep-learning architectures have also been used to forecast milk yield from routine farm records [13], illustrating the expanding role of machine learning in dairy farm monitoring. While these studies demonstrate the capacity of machine learning to extract meaningful patterns from heterogeneous dairy datasets, no previous research has evaluated the extent to which routinely recorded milk biomarkers are associated with farm-level welfare classifications derived from the WQ® protocol, nor has it systematically compared multiple machine learning algorithms within a unified analytical framework.
Addressing this gap, the present study adopts an exploratory approach to examine whether commonly recorded milk biomarkers contain welfare-relevant information at the herd level. By applying and comparing several supervised machine learning methods, the study investigates the capacity of milk composition data to discriminate between farms with different welfare statuses and to identify multivariate patterns associated with welfare classification. Rather than proposing milk biomarkers as stand-alone welfare indicators, this work aims to provide preliminary evidence supporting the integration of milk-based information into scalable, data-driven strategies for precision-welfare monitoring.

2. Materials and Methods

2.1. Study Population and Welfare Assessment

Between April 2023 and March 2025, animal welfare evaluations were carried out on 43 commercial dairy farms located in Transylvania, Romania. Herd size ranged from 28 to 560 cows, and most farms housed Holstein–Friesian or Romanian Spotted cattle. Housing systems included free-stall and tie-stall configurations, with variability in bedding materials, feeding strategies and access to outdoor areas. Management practices reflected typical commercial dairy systems in the region, providing a representative sample of local production conditions.
A total of 4857 lactating cows were assessed with the WQ® Assessment Protocol for Dairy Cattle [3] by two training observers. Because the study spans the period from April to March, the dataset reflects welfare conditions under both warm-season and cool-season environmental conditions, thereby capturing the typical climatic variability of the region.
The WQ® protocol integrates resource-based and animal-based indicators into a comprehensive evaluation structured across four principles (good feeding, good housing, good health and appropriate behaviour), further divided into 12 criteria. These include absence of prolonged hunger and thirst, comfort around resting, ease of movement, absence of injuries and disease, avoidance of pain during management procedures, expression of social and species-specific behaviours, emotional state, and human–animal relationship. Indicators were collected through direct cow observation, behavioural tests (e.g., avoidance distance), resource inspection and analysis of farm records.
Individual-level assessments included body condition scoring, evaluation of lying and rising behaviour, hygiene scoring (lower legs, udder, flanks), lameness scoring (non-lame, moderately lame, severely lame), and clinical signs such as nasal or ocular discharge, coughing, diarrhoea, vulvar discharge, altered respiratory patterns, and integumentary changes (hairless patches, lesions, swellings). Qualitative Behaviour Assessment (QBA) was used to characterise the cows’ emotional expression using predefined descriptors (e.g., relaxed, fearful, agitated, inquisitive, playful, apathetic).
Farm records documented the occurrence of mastitis, dystocia, downer cows and mortality. All indicators were processed using the official WQ® Assessment Tool [14], which calculates criterion-level and principle-level scores and assigns each farm to one of four categories: Not Classified, Acceptable, Enhanced or Excellent.
All welfare assessments were performed with the consent of farm owners and complied with Romanian National Animal Protection Law no. 138/2022 [15].

2.2. Milk Sampling and Laboratory Analysis

To ensure direct comparability between welfare status and milk quality, individual milk samples were collected from all lactating cows at each farm during the routine morning milking. Sampling was performed by certified technicians as part of the official milk recording programme. Because all farms underwent milk recording during the same milking session (morning), this approach ensured methodological standardisation across herds and minimised variability associated with circadian changes in milk composition.
Immediately after collection, milk samples were refrigerated at 4 °C and transported to the laboratory within 24 h. All analyses were performed on the day of arrival, following standard milk recording procedures, to ensure biomarker stability and analytical consistency.
For each farm, individual-cow results were aggregated into farm-level mean values, which served as the analytical unit for all subsequent statistical modelling. Using the mean is consistent to evaluate welfare at the management (farm) scale and reflects the average physiological profile of the herd. This aggregation also reduces the influence of extreme individual values and aligns with the level at which Welfare Quality® scores are assigned. However, aggregating data at the farm level may mask within-herd welfare heterogeneity and dilute signals originating from smaller subgroups of compromised animals. Consequently, the associations identified in this study should be interpreted as reflecting average herd-level conditions rather than individual cow-level welfare variation. Although individual-cow variability is lost through averaging, this approach is appropriate for identifying herd-level patterns relevant for welfare classification and for developing farm-oriented analytical models.
The analytical profile included milk yield (kg per cow at the morning milking), fat, protein, casein and lactose contents, from which the FPR was calculated. Additional routinely monitored biomarkers included urea, acetone, BHB, SCC, DSCC and TPC. All physico-chemical parameters, as well as SCC and DSCC, were determined using a CombiFoss™ 7 FT analyser (Foss Electric, Hillerød, Denmark), while TPC was quantified with a BactoScan™ FC analyser (Foss Electric, Hillerød, Denmark )from the same manufacturer. All analyses were conducted in a laboratory accredited according to ISO/IEC 17025 [16], and the analytical methods used for milk composition, SCC, DSCC, and TPC were fully validated and subject to regular quality control procedures [16].

2.3. Statistical Analysis

All statistical analyses were performed in R software (version 4.5.1, R Core Team, Vienna, Austria). Data management and visualisation were carried out using tidyverse, and descriptive statistics were obtained with psych. Group-level comparisons among the three WQ® categories (Not Classified, Acceptable, and Enhanced) were performed using rstatix. Machine learning workflows were implemented through the caret framework with support from glmnet, randomForest, nnet, e1071, class, MASS, and neuralnet.
All variables were checked for missing values, and no imputation was required. Because the biomarkers were measured on different scales, all predictors were standardised (mean = 0, standard deviation = 1) before model training. Descriptive statistics were calculated using the median and interquartile range (IQR). Group differences were assessed with the Kruskal–Wallis test, and p-values were adjusted using the Benjamini–Hochberg false discovery rate (FDR) procedure [17]. When overall differences were statistically significant, pairwise comparisons between welfare categories were performed using Dunn’s post hoc test with Benjamini–Hochberg FDR correction.
Six supervised machine learning algorithms were evaluated: multinomial Elastic-Net regression (ENET), Random Forest (RF), Multilayer Perceptron (MLP), Support Vector Machine (SVM), k-Nearest Neighbours (k-NN) and Linear Discriminant Analysis (LDA). ENET was implemented via glmnet and combines L1 (LASSO) and L2 (Ridge) penalties to address multicollinearity and perform embedded variable selection [18]. RF constructs multiple decision trees and aggregates their predictions to model complex, non-linear relationships [19]. The MLP captures hierarchical, non-linear patterns through interconnected computational layers [20]. SVM identify the optimal separating hyperplane in high-dimensional or non-linear spaces [21]. The k-NN method classifies observations based on the dominant class among their closest neighbours [22], while LDA maximises between-group separation under assumptions of homogeneous variance and covariance structures [23].
All models were trained using 10-fold stratified cross-validation to preserve class proportions (14 Enhanced, 19 Acceptable, 10 Not Classified) and to limit overfitting. Hyperparameters were optimised using caret’s grid search. Performance was evaluated using Accuracy, Cohen’s Kappa, Sensitivity, Specificity, Precision, and F1-score.
Given the moderate class imbalance, particularly for the Not Classified category (n = 10), no oversampling, undersampling, or class-weight adjustments were applied to avoid altering the natural class distribution. Stratified cross-validation was used to maintain representation across folds. Performance metrics for the minority class should therefore be interpreted with caution, as reduced sensitivity is a common challenge in multi-class classification with imbalanced datasets.

3. Results

3.1. Milk Biomarker Variation Across Welfare Categories

In this study, none of the evaluated farms reached the Excellent category. The final distribution was: 19 Acceptable, 14 Enhanced and 10 Not Classified farms.
Median values and interquartile ranges (IQR) for each milk biomarker across the three WQ® categories are presented in Table 1, which summarises both global differences following false discovery rate (FDR) adjustment and post hoc pairwise comparisons indicated by superscript letters.
Somatic cell indicators exhibited the most pronounced differences across welfare categories. SCC increased progressively from Enhanced to Not Classified farms (Table 1). This corresponds to an approximately threefold increase between the highest and lowest welfare categories (FDR-adjusted p = 0.028). A similar gradient was observed for DSCC, which increased progressively from Enhanced to Not Classified farms, with statistically significant differences between higher- and lower-welfare groups following false discovery rate adjustment. Post hoc comparisons indicated that Enhanced farms differed significantly from the other categories for both indicators, while Acceptable and Not Classified farms did not differ from each other.
Lactose concentration also differed significantly between welfare categories. Enhanced farms showed the highest median lactose content, followed by Acceptable farms, whereas Not Classified farms exhibited lower values. The overall difference between Enhanced and Not Classified farms was approximately 0.25 g/100 g (FDR-adjusted p = 0.006).
Milk yield at the morning milking followed a similar pattern, decreasing with lower welfare classification. Enhanced farms differed significantly from both Acceptable and Not Classified farms, whereas no significant difference was detected between the latter two categories. The milk yield difference was approximately 3 L per cow between the highest and lowest welfare categories.
Acetone showed a modest but statistically significant difference across categories, with slightly higher median values in Not Classified farms compared with Enhanced and Acceptable farms. Casein content displayed a decreasing trend from Enhanced to Not Classified farms; however, this pattern did not reach statistical significance after correction for multiple testing.
No significant differences between welfare categories were observed for milk fat, protein, FPR, urea, BHB, or TPC following FDR adjustment (p > 0.05).

3.2. Performance of Machine Learning Models

Six supervised machine learning models were trained to predict the WQ® classification of dairy farms based on milk biomarkers: ENET, RF, MLP, SVM, k-NN and LDA. All models were evaluated under identical 10-fold stratified cross-validation conditions, and their predictive performance metrics are summarised in Table 2.
Among the evaluated algorithms, ENET showed the highest overall classification performance, with the strongest agreement between predicted and observed classes as reflected by its mean Kappa values. RF and MLP achieved similar average accuracy levels; however, their agreement metrics exhibited greater variability across cross-validation folds. SVM and k-NN produced intermediate performance with moderate agreement across classes, while LDA showed the lowest performance and weakest agreement among the tested models. Notably, the relatively large standard deviations observed for both accuracy and Cohen’s Kappa indicate limited model stability across cross-validation folds, particularly for RF and k-NN, highlighting sensitivity to data partitioning.

3.3. Class-Wise Classification Performance

A detailed evaluation of the class-wise predictive performance of the machine learning models is presented in Table 3. This subsection reports the ability of each algorithm to correctly identify the three WQ® categories based on milk biomarkers.
Across most algorithms, the Enhanced category was identified with the highest and most consistent performance, with sensitivity and F1-score values showing the strongest classification results for this group. The Acceptable category showed intermediate performance across models, with moderate sensitivity and F1-score values. The Not Classified category showed the weakest and most variable classification performance. Sensitivity values ranged widely between algorithms, and in some cases, such as k-NN, the models did not correctly identify any farms from this category. Specificity values for this class were consistently high across all models.
Among the evaluated algorithms, ENET and RF achieved the most balanced class-wise performance, correctly identifying farms across all three categories with relatively consistent metrics. SVM and k-NN performed well for the Enhanced category but showed limited ability to identify the minority class. LDA and MLP produced moderate but stable values across all classes.

3.4. Variable Importance Across Machine-Learning Models

Figure 1 presents the relative importance of the milk biomarkers in predicting the WQ® category across the six machine learning algorithms (ENET, RF, SVM, k-NN, LDA, and MLP). Because each algorithm generates importance values on different internal scales, the figure displays model-specific importance rankings rather than directly comparable magnitudes. For ENET, variable importance was derived from standardised regression coefficients. For RF, importance was based on impurity reduction, and for SVM, k-NN, LDA, and MLP, importance was calculated using permutation-based decreases in accuracy. This approach ensures that the importance values reported for each model reflect the internal decision process of the respective algorithm.
Across all models, SCC and DSCC were consistently ranked as the most important predictors of welfare status. Enhanced farms showed lower SCC and DSCC values, while Not Classified farms showed higher values, with the Acceptable category in between.
Indicators such as the FPR, lactose, and milk yield also received high importance scores across several models. Their contribution was most pronounced in ENET and LDA, which assign higher weights to variables with strong linear separation between classes.
The RF model ranked SCC, DSCC, FPR, fat and urea among its most influential variables. SVM and MLP assigned high importance to DSCC, lactose, and urea. The k-NN and LDA placed greater importance on FPR and lactose.
Acetone, BHB and TPC showed lower but consistent importance scores across models.
Overall, SCC, DSCC, FPR, lactose and milk yield were the variables most frequently ranked among the top predictors across the six machine learning algorithms.

4. Discussion

This study adds to the existing literature by demonstrating that routinely recorded milk biomarkers are coherently associated with multidimensional, standardised welfare classifications at farm level, rather than with isolated health or management outcomes. Consistent associations were observed for several routinely recorded milk biomarkers, most notably SCC, DSCC, lactose, and milk yield, in relation to farm-level welfare classifications derived from the WQ® protocol. Importantly, these associations were detected at the herd level using aggregated routine milk-recording data, indicating that welfare-related physiological signals remain observable beyond individual animal assessments. While previous studies have linked individual milk traits to specific health or management conditions [3,4,5,24,25,26], the present findings extend this knowledge by showing that coherent welfare-related gradients in milk biomarkers correspond to standardised, multidimensional welfare classifications at the farm level.
The observed welfare gradients for SCC and DSCC corroborate their established role as indicators of mastitis and inflammatory stress [8,27]. The magnitude of these differences is biologically meaningful, with median SCC values approximately threefold higher and milk yield approximately 25% lower in Not Classified farms compared with Enhanced farms, indicating that inflammatory burden and reduced productivity co-occur at lower welfare levels. However, the present study extends existing knowledge by demonstrating that these indicators remain informative when aggregated at the herd level and evaluated against multidimensional welfare classifications rather than single health outcomes. Specifically, the alignment between SCC- and DSCC-based patterns and Welfare Quality® categories suggests that chronic udder health challenges contribute measurably to overall welfare differentiation among farms. These findings indicate that routinely collected milk data can capture cumulative welfare-related effects of management practices, such as hygiene, milking routines, bedding quality, and equipment maintenance, which are otherwise difficult to quantify at the herd scale.
Elevated SCC and DSCC likely reflect cumulative effects of hygiene, milking routines, bedding management and equipment maintenance, all of which influence welfare outcomes. From a mechanistic standpoint, inadequate management conditions promote inflammatory responses and increased somatic cell recruitment in milk. Comparable associations were documented in machine learning studies focused on detecting mastitis or udder inflammation [10,11]. Additional work has shown that predictive performance can be further improved by incorporating milk yield, rumination time, and electrical conductivity [28], and that automated milking system data can support machine learning-based detection of udder disorders [29]. Ensemble algorithms have also shown superior predictive performance for subclinical mastitis relative to single-model approaches [30].
Lactose and milk yield also differentiated welfare categories. Lactose synthesis is closely linked to mammary epithelial integrity and systemic physiological stability, and reductions in lactose content are commonly associated with inflammatory processes, increased epithelial permeability, and stress-related alterations in glucose partitioning [9]. Accordingly, higher lactose levels observed in Enhanced farms are consistent with better udder health and physiological stability at the herd level. The higher productivity recorded in these farms reinforces the established link between good welfare and sustained metabolic performance, a relationship previously documented in herd-level studies [31,32,33]. These findings are consistent with reports indicating that milk yield can reflect overall welfare conditions [24,26]. Furthermore, deep learning techniques have demonstrated the feasibility of accurately forecasting milk yield from sensor-derived variables [13], supporting the inclusion of production metrics in data-driven welfare prediction frameworks.
In contrast, metabolic indicators such as the FPR, acetone and BHB did not differ significantly between welfare categories. Although these biomarkers are valuable for detecting energy imbalance or subclinical ketosis, their predictive behaviour is known to vary across farms, parities and feeding systems, as shown in machine learning studies on ketosis detection [34]. Previous work has also demonstrated that the relationship between FPR and energy balance depends on lactation stage and parity [35] and that nitrogen metabolism can be affected by health disorders such as lameness [36]. Their weaker association with welfare classification in the present study may therefore reflect heterogeneity in ration composition and lactation stage, a pattern also noted in earlier metabolic prediction studies [12,37]. Similarly, urea values were relatively uniform across categories, suggesting broadly adequate nutritional management and limited discriminatory power for welfare classification.
This study adds novel methodological insight by showing that consistent welfare-related biomarker patterns emerge across different machine learning algorithms, despite limited sample size and class imbalance. The machine learning analysis provided additional insight into the predictive value of milk biomarkers. The ENET model achieved the highest mean accuracy, followed closely by MLP and RF, indicating that both regularised linear approaches and non-linear architectures can capture multivariate patterns linking milk composition to welfare categories. However, the relatively large variability of accuracy estimates across folds reflects the limited sample size and the known instability of machine learning models trained on small datasets. Comparable issues have been reported in other veterinary machine learning applications involving disease detection or physiological prediction [38,39]. The class imbalance observed in this study also influenced model behaviour, particularly the reduced sensitivity for the Not Classified category. Similar challenges have been documented in machine learning studies working with highly imbalanced datasets, including displaced abomasum prediction [40].
Despite these challenges, the convergence of results across models represents a key novel finding of this study. The consistent prominence of SCC, DSCC, lactose and milk yield supports the biological relevance of these biomarkers and their utility for welfare-related prediction, rather than reflecting model-specific artefacts. Variable importance profiles emphasised the dominant contribution of udder-health biomarkers, with SCC and DSCC ranking highest in nearly all algorithms. Lactose, FPR and milk yield also appeared frequently among the top contributors, indicating that welfare classification is influenced by a combination of inflammatory, metabolic and production-related signals. These findings are consistent with multi-biomarker machine learning approaches in dairy cattle [41,42] and highlight the potential for integrating milk composition analysis into precision-welfare monitoring. The increasing availability of sensor-based behavioural data further suggests that future predictive systems could combine milk biomarkers with variables such as activity, rumination or movement patterns, an approach successfully used for calving prediction [43] and estrus detection [44]. Reviews summarising more than two decades of machine learning research in dairy systems highlight the rapid expansion of algorithmic monitoring tools and their suitability for welfare-oriented applications [45,46,47]. Additional machine learning studies focusing on lameness prediction reinforce the value of combined data sources for early detection of welfare-relevant disorders [48,49,50].

Limitations and Future Directions

Several limitations of this study should be acknowledged when interpreting the findings. First, the sample size of 43 farms restricted the statistical power and increased the risk of overfitting in the machine learning models, despite the use of cross-validation. The substantial variability in accuracy across folds indicates limited model stability. In addition, nested cross-validation was not applied; therefore, the reported performance metrics may be optimistically biassed. Moreover, the cross-sectional design of the study precludes any causal inference between milk biomarkers and welfare status, and the observed relationships should be interpreted as associative and exploratory rather than causal. Second, farm-level aggregation obscures individual cow-level variation and may mask within-herd welfare heterogeneity. Third, the study was conducted within a specific regional production context, and management practices may differ from those in other dairy systems. In addition, the absence of farms classified as Excellent limited generalizability across the full WQ® spectrum and restricted differentiation at the upper end of the welfare scale. Fourth, unmeasured factors such as milking routines, hygiene, staffing, or equipment maintenance may have contributed to the SCC and DSCC gradients observed. Finally, the study relied exclusively on milk biomarkers; integrating behavioural, environmental, and health-sensor data could enhance predictive precision.
Future research should expand the dataset, incorporate individual-level and longitudinal measurements, investigate hybrid modelling frameworks such as mixed-effects models combined with machine learning, apply nested cross-validation, and explore explainable machine learning approaches. Combining milk biomarkers with sensor-based behavioural monitoring may ultimately enable high-frequency, automated welfare surveillance systems aligned with the One Welfare paradigm.

5. Conclusions

This study shows that several routinely measured milk biomarkers, particularly SCC, DSCC, lactose, and milk yield, are consistently associated with herd-level welfare classifications derived from the WQ® protocol. Farms with higher welfare status tended to exhibit more favourable milk profiles, whereas lower-welfare herds displayed patterns indicative of udder health challenges and reduced productivity.
A key novel contribution of this work lies in the comparative machine learning approach, which revealed a stable set of influential biomarkers across multiple algorithms, despite limited sample size and class imbalance. Rather than aiming to demonstrate high predictive accuracy, the analysis highlights the robustness of specific milk variables as drivers of welfare differentiation and transparently illustrates the methodological constraints of applying ML in real-world dairy datasets. These findings position milk biomarkers as complementary, non-invasive indicators that can support exploratory welfare screening at the herd level, rather than as stand-alone assessment tools.
Overall, this study advances current knowledge by integrating welfare scoring, milk-recording data, and machine learning to explore scalable approaches for welfare monitoring. Further research incorporating larger and more balanced datasets, individual cow-level information, and behavioural and environmental measures is required before such approaches can be considered for robust application in precision-welfare monitoring.

Author Contributions

Conceptualisation, D.E.B. and S.P.; methodology, D.E.B., A.B.P. and S.P.; validation, D.M.N., S.D. and E.A.L.; formal analysis, D.E.B., D.M.N., E.A.L. and S.P.; investigation, D.E.B., A.B.P., S.D., D.M.N., E.A.L. and S.P.; data curation, D.E.B., A.B.P., E.A.L. and S.P.; writing—original draft preparation, D.E.B., A.B.P., S.D. and D.M.N.; writing—review and editing, D.E.B., A.B.P., S.D. and E.A.L.; supervision, S.P. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

Ethical review and approval were waived for this study because the animal welfare assessment protocol does not imply touching the animals, nor does it cause any stress to them. All procedures involving animals were conducted in accordance with the Romanian National Animal Protection Law (Law no. 138/2022) and EU Directive 2010/63/EU.

Informed Consent Statement

Written informed consent was obtained from the owner of the animals involved in this study.

Data Availability Statement

The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding authors.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. World Organisation for Animal Health. Introduction to the Recommendations for Animal Welfare. In Terrestrial Animal Health Code; OIE: Paris, France, 2008; pp. 235–236. [Google Scholar]
  2. Clark, B.; Stewart, G.B.; Panzone, L.A.; Kyriazakis, I.; Frewer, L.J. A systematic review of public attitudes, perceptions and behaviours towards production diseases associated with farm animal welfare. J. Agric. Environ. Ethics 2016, 29, 455–478. [Google Scholar] [CrossRef]
  3. Welfare Quality® Consortium. Welfare Quality® Assessment Protocol for Cattle; Welfare Quality® Consortium: Lelystad, The Netherlands, 2009. [Google Scholar]
  4. Andreasen, S.N.; Sandøe, P.; Forkman, B. Can animal-based welfare assessment be simplified? A comparison of the Welfare Quality® protocol for dairy cattle and the simpler and less time-consuming protocol developed by the Danish Cattle Federation. Anim. Welf. 2014, 23, 81–94. [Google Scholar] [CrossRef]
  5. Heath, C.A.E.; Browne, W.J.; Mullan, S.; Main, D.C.J. Navigating the iceberg: Reducing the number of parameters within the Welfare Quality® assessment protocol for dairy cows. Animal 2014, 8, 1978–1986. [Google Scholar] [CrossRef] [PubMed]
  6. Antanaitis, R.; Džermeikaitė, K.; Januškevičius, V.; Šimonytė, I.; Baumgartner, W. In-Line Registered Milk Fat-to-Protein Ratio for the Assessment of Metabolic Status in Dairy Cows. Animals 2023, 13, 3293. [Google Scholar] [CrossRef]
  7. Broderick, G.A.; Clayton, M.K. A statistical evaluation of animal and nutritional factors influencing concentrations of milk urea nitrogen. J. Dairy Sci. 1997, 80, 2964–2971. [Google Scholar] [CrossRef]
  8. Ruegg, P.L. A 100-year review: Mastitis detection, management, and prevention. J. Dairy Sci. 2017, 100, 10381–10397. [Google Scholar] [CrossRef]
  9. Coignard, M.; Guatteo, R.; Veissier, I.; Lehébel, A.; Hoogveld, C.; Mounier, L.; Bareille, N. Does milk yield reflect the level of welfare in dairy herds? Vet. J. 2014, 199, 184–187. [Google Scholar] [CrossRef]
  10. Bobbo, T.; Biffani, S.; Taccioli, C.; Penasa, M.; Cassandro, M. Comparison of machine learning methods to predict udder health status based on somatic cell counts in dairy cows. Sci. Rep. 2021, 11, 13642. [Google Scholar] [CrossRef] [PubMed]
  11. Hyde, R.M.; Down, P.M.; Bradley, A.J.; Breen, J.E.; Hudson, C.; Leach, K.A.; Green, M.J. Automated prediction of mastitis infection patterns in dairy herds using machine learning. Sci. Rep. 2020, 10, 4289. [Google Scholar] [CrossRef]
  12. Huot, F.; Claveau, S.; Bunel, A.; Warner, D.; Santschi, D.E.; Gervais, R.; Paquet, E.R. Predicting subacute ruminal acidosis from milk mid-infrared estimated fatty acids and machine learning. J. Dairy Sci. 2024, 107, 9504–9515. [Google Scholar] [CrossRef]
  13. Alwadi, M.; Alwadi, A.; Chetty, G.; Alnaimi, J. Smart dairy farming for predicting milk production yield based on deep machine learning. Int. J. Inf. Technol. 2024, 16, 4181–4190. [Google Scholar] [CrossRef]
  14. INRAE. Welfare Quality® Scoring System. Available online: https://www1.clermont.inrae.fr/wq/?simu=on (accessed on 7 March 2025).
  15. The Parliament of Romania. Law No. 138/2022 for Amending and Supplementing Law No. 205/2004 Regarding the Protection of Animals. Monitorul Oficial al României, Part I, No. 484, 17 May 2022. Available online: https://legislatie.just.ro/public/DetaliiDocument/292029 (accessed on 1 November 2025).
  16. ISO/IEC 17025:2017; General Requirements for the Competence of Testing and Calibration Laboratories. International Organization for Standardization: Geneva, Switzerland, 2017.
  17. Benjamini, Y.; Hochberg, Y. Controlling the false discovery rate. J. R. Stat. Soc. B 1995, 57, 289–300. [Google Scholar] [CrossRef]
  18. Zou, H.; Hastie, T. Regularization and variable selection via the elastic net. J. R. Stat. Soc. B 2005, 67, 301–320, Erratum in J. R. Stat. Soc. B 2005, 67, 768. https://doi.org/10.1111/j.1467-9868.2005.00527.x. [Google Scholar] [CrossRef]
  19. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  20. Taud, H.; Mas, J.F. Multilayer Perceptron (MLP). In Geomatic Approaches for Modeling Land Change Scenarios; Camacho Olmedo, M.T., Paegelow, M., Mas, J.-F., Escobar, F., Eds.; Springer: Cham, Switzerland, 2018. [Google Scholar] [CrossRef]
  21. Cortes, C.; Vapnik, V. Support-vector networks. Mach. Learn. 1995, 20, 273–297. [Google Scholar] [CrossRef]
  22. Cover, T.; Hart, P. Nearest neighbor pattern classification. IEEE Trans. Inf. Theory 1967, 13, 21–27. [Google Scholar] [CrossRef]
  23. Zhao, S.; Zhang, B.; Yang, J.; Zhou, J.; Xu, Y. Linear discriminant analysis. Nat. Rev. Methods Primers 2024, 4, 70. [Google Scholar] [CrossRef]
  24. de Vries, M.; Bokkers, E.A.M.; van Schaik, G.; Engel, B.; Dijkstra, T.; de Boer, I.J.M. Improving time efficiency in identifying dairy herds with poorer welfare in a population. J. Dairy Sci. 2016, 99, 8282–8296. [Google Scholar] [CrossRef]
  25. Otten, N.D.; Rousing, T.; Houe, H.; Thomsen, P.T.; Sørensen, J.T. Comparison of animal welfare indices in dairy herds based on different sources of data. Anim. Welf. 2016, 25, 207–215. [Google Scholar] [CrossRef]
  26. Lutz, B.; Zwygart, S.; Thomann, B.; Stucki, D.; Burla, J.B. Common data-based indicators and welfare in Swiss dairy herds. Front. Vet. Sci. 2022, 9, 991363. [Google Scholar] [CrossRef] [PubMed]
  27. Kirkeby, C.; Toft, N.; Schwarz, D.; Farre, M.; Nielsen, S.S.; Zervens, L.; Hechinger, S.; Halasa, T. Differential somatic cell count as an additional indicator for intramammary infections in dairy cows. J. Dairy Sci. 2020, 103, 1759–1775. [Google Scholar] [CrossRef]
  28. Tian, H.; Zhou, X.; Wang, H.; Xu, C.; Zhao, Z.; Xu, W.; Deng, Z. The Prediction of Clinical Mastitis in Dairy Cows Based on Milk Yield, Rumination Time, and Milk Electrical Conductivity Using Machine Learning Algorithms. Animals 2024, 14, 427. [Google Scholar] [CrossRef]
  29. Faris, D.N.; Gad, A.M.; El-Tarabany, M.S.; Ramadan, S.I.; Afifi, G.G.; Manaa, E.A. Machine learning algorithms for clinical mastitis prediction. J. Adv. Vet. Res. 2024, 14, 975–981. [Google Scholar]
  30. Satoła, A.; Satoła, K. Performance comparison of machine learning models used for predicting subclinical mastitis in dairy cows: Bagging, boosting, stacking, and super-learner ensembles versus single machine learning models. J. Dairy Sci. 2024, 107, 3959–3972. [Google Scholar] [CrossRef]
  31. de Vries, M.; Bokkers, E.A.M.; Dijkstra, T.; van Schaik, G.; de Boer, I.J.M. Invited review: Associations between variables of routine herd data and dairy cattle welfare indicators. J. Dairy Sci. 2011, 94, 3213–3228. [Google Scholar] [CrossRef] [PubMed]
  32. De Vries, M.; Bokkers, E.A.M.; Van Schaik, G.; Engel, B.; Dijkstra, T.; De Boer, I.J.M. Exploring the value of routinely collected herd data for estimating dairy cattle welfare. J. Dairy Sci. 2014, 97, 715–730. [Google Scholar] [CrossRef] [PubMed]
  33. Mota, L.F.; Pegolo, S.; Baba, T.; Peñagaricano, F.; Morota, G.; Bittante, G.; Cecchinato, A. Evaluating the performance of machine learning methods and variable selection methods for predicting difficult-to-measure traits in Holstein dairy cattle using milk infrared spectral data. J. Dairy Sci. 2021, 104, 8107–8121. [Google Scholar] [CrossRef]
  34. Satoła, A.; Bauer, E.A. Predicting Subclinical Ketosis in Dairy Cows Using Machine Learning Techniques. Animals 2021, 11, 2131. [Google Scholar] [CrossRef]
  35. Cabezas-Garcia, E.H.; Gordon, A.W.; Mulligan, F.J.; Ferris, C.P. Revisiting the Relationships between Fat-to-Protein Ratio in Milk and Energy Balance in Dairy Cows of Different Parities, and at Different Stages of Lactation. Animals 2021, 11, 3256. [Google Scholar] [CrossRef]
  36. Necula, D.-C.; Warren, H.E.; Taylor-Pickard, J.; Simiz, E.; Stef, L. Associations of Lameness with Indicators of Nitrogen Metabolism and Excretion in Dairy Cows. Agriculture 2022, 12, 2109. [Google Scholar] [CrossRef]
  37. Xu, W.; van Knegsel, A.T.; Vervoort, J.J.; Bruckmaier, R.M.; van Hoeij, R.J.; Kemp, B.; Saccenti, E. Prediction of metabolic status of dairy cows in early lactation with on-farm cow data and machine learning algorithms. J. Dairy Sci. 2019, 102, 10186–10201. [Google Scholar] [CrossRef]
  38. Megahed, A.A.; Bommineni, R.; Short, M.; Galvao, K.N.; Bittar, J.H. Using supervised machine learning algorithms to predict bovine leukemia virus seropositivity in dairy cattle in Florida: A 10-year retrospective study. Prev. Vet. Med. 2025, 235, 106387. [Google Scholar] [CrossRef] [PubMed]
  39. Becker, C.A.; Aghalari, A.; Marufuzzaman, M.; Stone, A.E. Predicting dairy cattle heat stress using machine learning techniques. J. Dairy Sci. 2021, 104, 501–524. [Google Scholar] [CrossRef] [PubMed]
  40. Asgari, Z.; Sadeghi-Sefidmazgi, A.; Pakdel, A.; Shahinfar, S. Machine Learning Approaches for the Prediction of Displaced Abomasum in Dairy Cows Using a Highly Imbalanced Dataset. Animals 2025, 15, 1833. [Google Scholar] [CrossRef]
  41. Lasser, J.; Matzhold, C.; Egger-Danner, C.; Fuerst-Waltl, B.; Steininger, F.; Wittek, T.; Klimek, P. Integrating diverse data sources to predict disease risk in dairy cattle—A machine learning approach. J. Anim. Sci. 2021, 99, skab294. [Google Scholar] [CrossRef]
  42. Sabbadin, S.; Magro, S.; Costa, A.; Lorenzi, V.; Fusi, F.; Bertocchi, L.; De Marchi, M. Bulk milk quality differs across welfare levels in Italian dairy cattle farms with different housing systems. Ital. J. Anim. Sci. 2025, 24, 910–922. [Google Scholar] [CrossRef]
  43. Borchers, M.R.; Chang, Y.M.; Proudfoot, K.L.; Wadsworth, B.A.; Stone, A.E.; Bewley, J.M. Machine-learning-based calving prediction from activity, lying, and ruminating behaviors in dairy cattle. J. Dairy Sci. 2017, 100, 5664–5674. [Google Scholar] [CrossRef] [PubMed]
  44. Wang, J.; Bell, M.; Liu, X.; Liu, G. Machine-Learning Techniques Can Enhance Dairy Cow Estrus Detection Using Location and Acceleration Data. Animals 2020, 10, 1160. [Google Scholar] [CrossRef]
  45. Cockburn, M. Review: Application and Prospective Discussion of Machine Learning for the Management of Dairy Farms. Animals 2020, 10, 1690. [Google Scholar] [CrossRef]
  46. Shine, P.; Murphy, M.D. Over 20 Years of Machine Learning Applications on Dairy Farms: A Comprehensive Mapping Study. Sensors 2022, 22, 52. [Google Scholar] [CrossRef]
  47. Grzesiak, W.; Zaborski, D.; Pluciński, M.; Jędrzejczak-Silicka, M.; Pilarczyk, R.; Sablik, P. The Use of Selected Machine Learning Methods in Dairy Cattle Farming: A Review. Animals 2025, 15, 2033. [Google Scholar] [CrossRef] [PubMed]
  48. Shahinfar, S.; Khansefid, M.; Haile-Mariam, M.; Pryce, J.E. Machine learning approaches for the prediction of lameness in dairy cows. Animal 2021, 15, 100391. [Google Scholar] [CrossRef] [PubMed]
  49. Neupane, R.; Aryal, A.; Haeussermann, A.; Hartung, E.; Pinedo, P.; Paudyal, S. Evaluating machine learning algorithms to predict lameness in dairy cattle. PLoS ONE 2024, 19, e0301167. [Google Scholar] [CrossRef] [PubMed]
  50. Dervić, E.; Matzhold, C.; Egger-Danner, C.; Steininger, F.; Klimek, P. Improving lameness detection in cows: A machine learning algorithm application. J. Dairy Sci. 2024, 107, 11550–11562. [Google Scholar] [CrossRef]
Vetsci 13 00022 g001
Figure 1. Variable importance rankings for all milk biomarkers across the six supervised machine learning algorithms used in this study: Random Forest (RF), Elastic-Net Multinomial Regression (ENET), k-Nearest neighbours (k-NN), Linear Discriminant Analysis (LDA), Multilayer Perceptron (MLP), and Support Vector Machine (SVM). Bars represent the relative contribution of each biomarker to predictive performance, quantified as permutation-based accuracy decrease (% for RF and ENET; mean accuracy drop for k-NN, LDA, MLP, and SVM). Higher values indicate stronger influence on welfare classification according to the Welfare Quality® categories.
Figure 1. Variable importance rankings for all milk biomarkers across the six supervised machine learning algorithms used in this study: Random Forest (RF), Elastic-Net Multinomial Regression (ENET), k-Nearest neighbours (k-NN), Linear Discriminant Analysis (LDA), Multilayer Perceptron (MLP), and Support Vector Machine (SVM). Bars represent the relative contribution of each biomarker to predictive performance, quantified as permutation-based accuracy decrease (% for RF and ENET; mean accuracy drop for k-NN, LDA, MLP, and SVM). Higher values indicate stronger influence on welfare classification according to the Welfare Quality® categories.
Vetsci 13 00022 g001
Table 1. Median (IQR) values of milk biomarkers across the three Welfare Quality® (WQ®) categories and FDR-adjusted p-values.
Table 1. Median (IQR) values of milk biomarkers across the three Welfare Quality® (WQ®) categories and FDR-adjusted p-values.
BiomarkerWelfare Classp (FDR-Adjusted)
EnhancedAcceptableNot Classified
Fat (g/100 g)3.99 (0.75)3.85 (0.40)4.12 (1.12)0.26
Protein (g/100 g)3.53 (0.440)3.55 (0.46)3.43 (0.47)0.27
FPR1.14 (0.09)1.08 (0.14)1.11 (0.59)0.26
Casein (g/100 g)2.85 (0.32)2.83 (0.38)2.56 (0.37)0.062
Lactose (g/100 g)4.84 (0.14) a4.82 (0.15) b4.59 (0.25) c0.006
Urea (mg/dL)22.28 (16.73)23.54 (13.43)23.00 (28.97)0.802
Acetone (mmol/L)0.00 (0.00) a0.00 (0.00) a0.01 (0.05) b0.043
BHB (mmol/L)0.02 (0.03)0.02 (0.03)0.03 (0.05)0.340
DSCC (%)52.02 (18.08) a64.64 (14.30) b81.82 (19.28) b0.003
SCC (×103 cells/mL)284.60 (160.89) a502.15 (302.51) b838.20 (329.250) b0.028
TPC (×103 CFU/mL)42.50 (78.50)150.00 (164.00)106.00 (37.00)0.103
Milk yield (L)12.50 (2.75) a10.00 (2.00) b9.50 (2.00) b0.010
FPR: fat-to-protein ratio; BHB: β-hydroxybutyrate; DSCC: differential somatic cell count; SCC: somatic cell count; TPC: total plate count; FDR: false discovery rate. Values are presented as median (IQR). Within a row, values with different superscript letters (a–c) differ significantly between welfare categories, whereas values sharing the same letter do not differ significantly (Dunn’s post hoc test with Benjamini–Hochberg false discovery rate correction, p < 0.05). Bold values indicate statistically significant differences (p < 0.05).
Table 2. Predictive performance of the six machine learning models.
Table 2. Predictive performance of the six machine learning models.
ModelAccuracy_meanAccuracy_sdKappa_meanKappa_sd
ENET0.7210.1920.5610.314
RF0.7050.2850.4660.522
MLP0.7050.1670.5270.266
SVM0.6960.2270.4960.396
KNN0.6860.2520.5240.377
LDA0.6210.1720.4010.315
Accuracy_mean = average classification accuracy; Accuracy_sd = standard deviation of accuracy across 10-fold cross-validation; Kappa_mean = average Cohen’s Kappa; Kappa_sd = standard deviation of Kappa. Higher accuracy indicates better model performance. Cohen’s Kappa values were interpreted according to commonly used conventions (0.41–0.60 = moderate; 0.61–0.80 = substantial; >0.80 = almost perfect). ENET = Elastic-Net Multinomial Regression; RF = Random Forest; MLP = Multilayer Perceptron; SVM = Support Vector Machine; KNN = k-Nearest neighbours; LDA = Linear Discriminant Analysis.
Table 3. Class-wise performance metrics for all machine learning models.
Table 3. Class-wise performance metrics for all machine learning models.
ModelClassSensitivitySpecificityPrecisionF1-Score
ENETNot Classified0.700.960.870.77
Acceptable0.680.790.720.70
Enhanced0.780.790.640.70
RFNot Classified0.501.001.000.66
Acceptable0.840.660.660.74
Enhanced0.780.890.780.78
k-NNNot Classified0.001.00NANA
Acceptable0.840.540.590.69
Enhanced0.850.860.750.80
SVMNot Classified0.401.001.000.57
Acceptable0.680.700.650.66
Enhanced0.850.750.630.72
LDANot Classified0.501.001.000.66
Acceptable0.730.700.660.70
Enhanced0.780.790.640.70
MLPNot Classified0.600.960.850.70
Acceptable0.630.750.660.64
Enhanced0.780.750.610.68
Sensitivity = true positive rate; Specificity = true negative rate; Precision = positive predictive value; F1 = harmonic mean of precision and recall. NA = metric could not be computed because the model did not predict any farm in that class. Higher values indicate better class discrimination. ENET = Elastic-Net Multinomial Regression; RF = Random Forest; k-NN = k-Nearest neighbours; SVM = Support Vector Machine; LDA = Linear Discriminant Analysis; MLP = Multilayer Perceptron.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Babiciu, D.E.; Blaga Petrean, A.; Daina, S.; Neagu, D.M.; Lazar, E.A.; Popescu, S. Milk Biomarkers and Herd Welfare Status in Dairy Cattle: A Machine Learning Approach. Vet. Sci. 2026, 13, 22. https://doi.org/10.3390/vetsci13010022

AMA Style

Babiciu DE, Blaga Petrean A, Daina S, Neagu DM, Lazar EA, Popescu S. Milk Biomarkers and Herd Welfare Status in Dairy Cattle: A Machine Learning Approach. Veterinary Sciences. 2026; 13(1):22. https://doi.org/10.3390/vetsci13010022

Chicago/Turabian Style

Babiciu, Daniela Elena, Anamaria Blaga Petrean, Sorana Daina, Daniela Mihaela Neagu, Eva Andrea Lazar, and Silvana Popescu. 2026. "Milk Biomarkers and Herd Welfare Status in Dairy Cattle: A Machine Learning Approach" Veterinary Sciences 13, no. 1: 22. https://doi.org/10.3390/vetsci13010022

APA Style

Babiciu, D. E., Blaga Petrean, A., Daina, S., Neagu, D. M., Lazar, E. A., & Popescu, S. (2026). Milk Biomarkers and Herd Welfare Status in Dairy Cattle: A Machine Learning Approach. Veterinary Sciences, 13(1), 22. https://doi.org/10.3390/vetsci13010022

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop