Collar-based prediction of individual dry matter intake in dairy cows: validation across international research farms and evidence of familial variation in commercial herds

Introduction

Feed is the largest cost in dairy production, accounting for over 50% of total expenses (Martin et al., 2021; Nascimento et al., 2024; Siberski-Cooper and Koltes, 2022), and individual dry matter intake (DMI) both determines that cost at the cow level and is the denominator of feed efficiency (FE), an increasingly important heritable trait (VandeHaar et al., 2016) defined as energy-corrected milk (ECM)/DMI (Connor, 2015; Guinguina et al., 2020). Yet DMI has remained largely unmeasurable outside research settings.

Combining DMI with economic parameters such as feed costs and milk prices enables the calculation of income over feed, a direct measure of cow-level profitability (VandeHaar et al., 2016). These metrics can support both culling decisions and breeding strategies, supporting improved herd efficiency and genetic progress (Siberski-Cooper and Koltes, 2022). In addition, optimizing FE contributes to environmental sustainability, as individual DMI is a key driver of enteric methane emissions (Connor, 2015; Guinguina et al., 2020; Kennedy et al., 2024).

The primary barrier to evaluating individual cow efficiency in commercial systems is the lack of a practical method for measuring DMI at scale. Direct measurement systems are costly and labor-intensive, confining them to experimental facilities (Brown et al., 2022; Nascimento et al., 2024). As a result, DMI estimation on commercial farms typically relies on standardized equations (de Souza et al., 2019; National Research Council, 2001). These equations require inputs that are not always practically available and often exhibit reduced accuracy at the extremes of production, particularly overestimating in high-producing cows (de Souza et al., 2019).

Precision livestock technologies, particularly collar-based sensors, are now widely used on commercial farms and provide continuous behavioral data. Sensor-derived measures such as rumination, eating, and activity have been shown to improve DMI prediction when combined with traditional performance variables (Brown et al., 2022; Martin et al., 2021). These behavioral traits are also associated with feed efficiency, suggesting that sensor data may capture biologically meaningful variation in intake and efficiency (Nascimento et al., 2024). However, until recently, these data streams remained largely unvalidated for routine commercial estimation of DMI and FE.

Accordingly, the objectives of this study were to: 1) validate a collar-based pipeline for predicting individual DMI across independent international research farms; and 2) characterize familial variation in predicted DMI and derived FE at commercial scale as an indicator of their biological and genetic relevance.

Materials and methods

Data collection

To train and validate the DMI prediction pipeline, individual DMI and performance data were collected from 320 lactating dairy cows across five international research farms. The participating institutions included Michigan State University, Università Cattolica del Sacro Cuore, the University of British Columbia, Penn State University, and the Institute of Animal Science – Prague (VUZV). Data from Michigan State University came from three different experiments, for which dietary data are available (Krogstad et al., 2024; Naughton et al., 2024; Mills et al., 2024). Dietary data from Penn State are available in Martins et al. (2025), and only data from the control group of this experiment were included in the analysis. The available dietary composition at each farm, including main forage sources, forage-to-concentrate ratio, and nutrient and energy profile, is summarized in Table 1. The initial dataset comprised 21,550 daily data points. Reference DMI was measured manually for cows housed in tie-stalls at Michigan State University or automatically using electronic feeding systems: Calan Broadbent Feeding System (American Calan Inc., Northwood, NH) was used in Penn State University and the Insentec roughage intake control system (Insentec B.V., Marknesse, the Netherlands) was used at VUZV, Università Cattolica del Sacro Cuore and the University of British Columbia. To ensure baseline physiological responses and uniformity in feeding practices, the dataset was rigorously filtered. Specifically, the analysis was restricted to cows fed a total mixed ration (TMR); any cows receiving more than 1.5 kg of dry matter (DM) per day in additional concentrate supplements alongside the TMR were removed from the dataset. In addition, daily data were ignored in cases of partial milk or collar data; collar data comprising less than 23 hours in a specific day were considered partial (to ensure behavioral classification was based on a near-complete diurnal cycle). Moreover, cows identified as eating less than 1 hour per day, suggesting abnormal behavior or potential measurement artifacts, were removed. Finally, to focus on steady-state lactation rather than the metabolic volatility of the immediate transition period, only cows with greater than 9 days in milk (DIM) were included in the final analysis. After filtering, the dataset comprised 18,350 valid daily data points which were aggregated to 16,888 rolling average weekly data points. A weekly observation was considered eligible if it was based on at least 3 valid daily data points. Data availability differed substantially between farms due to differences in experimental design and filtering criteria. Table 2 presents data availability by research farm.

Table 1. Dietary composition of total mixed rations fed to lactating dairy cows at each research farm.

—, not reported or not available;

F:C, forage-to-concentrate ratio on a DM basis.

*Estimated from the measured TMR composition using the dedicated formula from National Research Council (2001).

**CP, NDF, starch, and NEL are formulated (planned) diet targets, applied to both diets. DM and F:C are from delivery records.

Table 2. Data availability and descriptive statistics by research farm.

DIM, days in milk; DMI, dry matter intake. Weekly observations represent rolling averages calculated from at least three valid daily records. Descriptive statistics are based on the filtered dataset used to develop the underlying DMI-prediction algorithm.

Bold values in the Overall row indicate pooled totals and weighted descriptive statistics across all five research farms.

Sensor technology

All cows in the study were equipped with an AfiCollar (Afimilk Ltd., Afikim, Israel), a wearable neck-tag sensor containing a 3D accelerometer that continuously monitors animal activity and head movements. These signals are used to classify and quantify time spent eating, rumination, and panting behaviors (Leso et al., 2021).

Algorithm development and validation

Individual DMI was estimated using a machine learning pipeline applied to collar-derived behavioral outputs, cow-status features (parity, days in milk, and pregnancy status), and milk production data (Katz and Yair, 2025a, b). The prediction methodology underlying this pipeline is described in these granted patents.

Model performance was evaluated using a leave-farm-out cross-validation approach, in which the model was iteratively trained on data from four farms and tested on an independent fifth farm. This approach ensured that predictions were evaluated under conditions independent from the training data. To account for variability between farms, feed composition, and management conditions, the model incorporated normalization relative to cohort-level intake. This normalization was applied as a post-prediction calibration step and did not form part of the prediction model itself. The leave-farm-out validation was first performed independently, producing raw individual DMI predictions with no reference to cohort-level intake. The raw predictions were subsequently rescaled so that, within each held-out cohort, the mean predicted DMI matched the cohort mean intake. During cross-validation this cohort mean was the measured DMI mean of the held-out farm, a single aggregate that sets the level of the predictions but leaves their relative ranking across cows unchanged, so it cannot inflate the model’s ability to discriminate between individual animals. Following completion of cross-validation, a final production model was trained on the complete dataset from all five research farms.

To characterize the relative contribution of the main input categories, feature importances from the final production model were aggregated into the three groups defined above: behavioral, milk production, and cow-status features, and expressed as a share of total importance.

Inherited variation in predicted DMI and derived feed efficiency

The validated pipeline was applied to commercial farm data to evaluate evidence of inherited variation in predicted DMI and in the FE derived from it. For this assessment, data were collected from five commercial dairy farms in Israel, entirely independent of the research farms and animals used for pipeline validation.

The commercial dataset encompassed 2,629 cow-lactation records from 2,096 Holstein cows across the five farms (per-farm descriptive statistics are provided in Supplementary Table S1; overall data span: May 2022 to April 2026). The dataset included 617 dam–daughter pairs from 541 unique dams; 64.0% of the daughters were primiparous, 29.0% were in second lactation, and 7.0% were in third lactation or higher. Of 542 sires represented in the dataset, 44 had ≥10 qualifying daughters and were retained for the sire analysis, contributing 853 daughters (mean 19.4 per qualifying sire, median 15.5, SD 10.3, range 10–51). Of these daughters, 47.2% were primiparous, 27.7% were in second lactation, and 25.1% were in third lactation or higher.

ECM was calculated according to National Research Council (2001) using individual milk fat, protein, and lactose values determined continuously by inline milk analyzers (AfiLab, Afimilk Ltd., Afikim, Israel). For each cow and lactation, the average ECM and predicted DMI were calculated across the available lactation period, specifically between 21 and 305 DIM. An average was considered valid if it was based on at least 75% data coverage across this period. Individual cow FE was defined as the ratio of ECM to the predicted DMI.

To assess familial resemblance, average predicted DMI of dams was evaluated in relation to that of their daughters using linear regression; the same approach was then applied to the derived FE phenotype. For each phenotype, pairs were included where both dam and daughter had qualifying records, the daughter’s lactation was ≤ 4, with a maximum lactation difference of 3 between pair members. Where multiple qualifying lactations were available, the most recent qualifying lactation of both the dam and the daughter was used to form the pair. To account for potential confounding, sequential models were fitted by progressively adding farm as a fixed effect, a temporal covariate (farm × year-month mean of the phenotype under analysis, to control for herd-period environment), and daughter parity category (1, 2, or >2). Because feed efficiency requires a qualifying ECM record whereas predicted DMI does not, the predicted DMI analysis drew on a larger set of qualifying pairs than the FE analysis (n = 665 vs. n = 617). To further examine variation among sire groups, daughters were grouped by sire; sires with fewer than 10 qualifying daughters were excluded. A subanalysis was additionally restricted to sires with at least three daughters in each of at least two farms to verify that the sire effect was not driven by farm composition. Sire × farm combinations represented by a single daughter were excluded from the analysis as they provide no within-farm replication for estimation of the sire effect, resulting in a final analyzed sample of n = 553. Of the five commercial farms, three contributed qualifying daughters to the cross-farm subanalysis; sire representation was concentrated in two of those farms, with a smaller contribution from the third (Supplementary Table S2).

Statistical analysis

Statistical analyses were performed in Python; scikit-learn (sklearn) was used for all analyses, except for the among-sire analysis of variance (ANOVA), which was carried out with statsmodels. Model performance was evaluated using the coefficient of determination (r²) and root mean square error (RMSE) between predicted and measured DMI. All reported accuracy metrics were derived exclusively from the leave-farm-out cross-validation and not from the final production model. Slope bias was assessed by regressing measured DMI on predicted DMI and testing whether the slope differed significantly from unity (St-Pierre, 2003). Mean square prediction error (MSPE) was decomposed into mean bias, slope bias, and random error components, each expressed as a proportion of total MSPE. Farm-level prediction accuracy was additionally summarized using CV-RMSE, calculated as RMSE expressed as a percentage of farm mean DMI.

For the analysis of inherited variation, dam–daughter resemblance was evaluated using sequential linear regression models as described above. Among-sire variation was assessed using a linear model with sire, farm, temporal covariate, and parity as fixed effects, with the sire effect evaluated by Type I sequential ANOVA decomposition. An indicative heritability value was approximated from the dam–daughter regression slope for each phenotype (predicted DMI and derived FE) after full adjustment for farm, temporal, and parity effects, which provides a rough approximation of half the narrow-sense heritability (h² ≈ 2 × slope); this is reported as an indicative benchmark rather than a formal genetic parameter estimate. Statistical significance for all tests was declared at p < 0.05.

Results

Algorithm performance for DMI prediction

The relationship between measured and predicted DMI is presented in Figure 1. The model achieved an r² of 0.77 and an RMSE of 1.9 kg. Per-farm prediction accuracy is presented in Table 3. CV-RMSE was consistent across farms, ranging from 6.9% to 8.2%, with a pooled value of 7.3%. Farm-level r² ranged from 0.45 to 0.78, and RMSE ranged from 1.69 kg to 2.49 kg. MSPE decomposition indicated that prediction error was almost entirely attributable to random variation (99.97%), with negligible mean bias (0.00%) and slope bias (0.03%). The calibration regression of measured on predicted DMI yielded an intercept of 0.184 kg and a slope of 0.992 (95% CI: 0.984–1.000), the latter deviating marginally from unity (p = 0.04), consistent with the negligible slope-bias contribution (0.03%) identified above in the MSPE decomposition. This indicates that while a marginal tendency toward regression to the mean was detectable, it had no meaningful effect on prediction accuracy across the observed intake range.

Table 3. Leave-farm-out cross-validation prediction accuracy by research farm.

r², coefficient of determination; RMSE, root mean square error; CV-RMSE, coefficient of variation of RMSE, calculated as RMSE expressed as a percentage of farm mean DMI; Prediction accuracy was evaluated using a leave-farm-out cross-validation approach, in which the model was trained on data from four farms and tested on the held-out fifth farm.

In the final production model trained on the complete five-farm dataset, all three input categories contributed substantially and comparably to predictive performance: collar-derived behavioral features (eating, rumination, and panting) accounted for the largest share of total feature importance (36.9%), followed by milk production data (34.0%) and cow-status features (29.1%).

Evidence of familial variation in predicted DMI and derived feed efficiency

The familial variation analysis of the predicted DMI phenotype included 665 dam–daughter pairs from 581 unique dams. The unadjusted relationship between predicted DMI of dams and their daughters is illustrated in Figure 2. The unadjusted dam–daughter association was significant (r = 0.33, slope = 0.36, p = 1.2 × 10⁻¹⁸, r² = 0.11). After sequential adjustment for farm (slope = 0.18, p = 0.0001), addition of a farm-period mean predicted DMI temporal covariate (slope = 0.13, p = 0.001; farm × temporal interaction not significant, p = 0.31), and further adjustment for daughter parity (slope = 0.075, p = 0.01), the dam–daughter association remained significant at every step of the sequential adjustment. Daughter parity had a large influence on predicted DMI (multiparous daughters predicted to consume approximately 4 kg/day more than primiparous daughters). The fully adjusted slope yields an indicative h² approximation of ~0.15; this marked attenuation from the farm-adjusted value (~0.35) reflects strong confounding of predicted DMI with parity, and the fully adjusted estimate lies toward the lower end of published heritabilities for measured DMI in dairy cattle (Berry et al., 2014; Negussie et al., 2019).

The corresponding analysis of the derived FE phenotype included 617 dam–daughter pairs from 541 unique dams. The unadjusted relationship between FE of dams and their daughters is shown in Figure 3. A significant positive association was observed in the unadjusted analysis (r = 0.40, slope = 0.36, p = 8.8 × 10⁻²⁵). The association remained significant after progressive adjustment for farm (slope = 0.096, p = 0.010), temporal herd effects (slope = 0.112, p = 0.002), and daughter parity (slope = 0.106, p = 0.003). The consistent dam–daughter slope across the adjusted model specifications indicates that the association is not attributable to shared farm environment, seasonal herd trends, or parity structure. This fully adjusted slope yields an indicative h² approximation of ~0.21, consistent in magnitude with published estimates for related feed efficiency traits, though formal heritability estimation was beyond the scope of this study.

Figure 4 presents raw (unadjusted) FE of daughters grouped by sire. A significant effect of sire was detected after adjustment for farm, temporal, and parity effects (F = 8.3, p = 1.2 × 10⁻⁴⁰; n = 853 daughters, 44 sires). The sire effect was confirmed in a subanalysis restricted to 25 sires with at least three daughters in each of at least two farms (F = 6.8, p = 3.1 × 10⁻¹⁹; n = 553), demonstrating that the among-sire variation in daughter FE is not an artifact of farm-composition differences across sire groups. The distribution of qualifying daughters per sire and farm for these sires is provided in Supplementary Table S2.

Discussion

Accuracy and applicability of the underlying DMI prediction

The principal contribution of this study is the validation of collar-based prediction of individual DMI as a scalable measurement obtainable without individual intake-measurement infrastructure, together with evidence that both predicted DMI and the derived FE phenotype exhibit familial variation. The accuracy of that prediction, addressed below, underpins both findings. This prediction relies on sensor-derived behavioral data drawn from data streams increasingly available in commercial dairy systems. Unlike conventional prediction equations, which lose accuracy at production extremes, do not incorporate behavioral inputs and often require inputs that are difficult to obtain on a continuous basis, the integration of wearable sensor data enables scalable estimation of individual feed intake, addressing a key limitation in the practical application of feed efficiency metrics. These results support previous work showing that behavioral traits such as eating and rumination are associated with feed intake and can enhance prediction when combined with conventional production variables (Martin et al., 2021).

The prediction accuracy achieved in the current study is consistent with or better than previous data-driven DMI prediction approaches, with the key distinction of being validated across independent farms rather than within a single site. Brown et al. (2022) reported an r² of 0.67 and RMSEP of 2.83 kg for their best-performing model; Martin et al. (2021) reported r² = 0.78 and RMSEP = 1.85 kg for weekly DMI prediction using behavioral data, though under a less stringent, single-farm evaluation design (n = 124 cows). Compared with equation-based approaches, de Souza et al. (2019) reported a concordance correlation coefficient of 0.72 and RMSEP of 2.89 kg.

Although the model was validated using data collected in research environments, the required inputs (behavioral metrics from wearable sensors, milk production data, and basic cow descriptors) are routinely captured in modern dairy management systems. Feed delivery and intake are typically managed at the group level on commercial farms, and the cohort-level mean intake required for the model’s normalization step is readily obtainable from routine group feed records (total dry matter offered minus refusals, divided by group size), so the approach remains applicable without individual intake-measurement infrastructure. Together, these factors support the applicability of the model to commercial settings.

Farm-level r² ranged from 0.45 to 0.78, a span that largely reflects differences in within-herd intake variation rather than differences in prediction error. Because r² expresses the proportion of observed variance explained, it falls in herds whose cows differ little from one another even when absolute error is low. Both extremes illustrate this. At the Institute of Animal Science – Prague, the narrowest intake distribution in the study (22.3 ± 2.2 kg/day) coincided with the lowest r² (0.45) despite the lowest absolute error (RMSE = 1.69 kg). At Michigan State University, the widest distribution (27.7 ± 4.3 kg/day) coincided with the highest r² (0.78) despite a larger absolute error (RMSE = 2.00 kg). On an absolute scale, performance was considerably more uniform than the r² values suggest, with CV-RMSE ranging from 6.9% to 8.2% across five herds differing in country, housing, diet, and management. The upper end of that range was observed at Penn State University, where the higher RMSE (2.49 kg) likely reflects the higher mean DMI of cows at this site (mean DMI = 30.5 kg/day); the CV-RMSE of 8.2% was only marginally above the pooled value of 7.3%, suggesting no systematic model failure.

Biological relevance of the predicted DMI and derived feed efficiency phenotypes

Predicted DMI, the phenotype the collar delivers directly, showed familial variation that survived all adjustments, including parity, which exerts a large influence on intake. This indicates that the collar-predicted trait itself carries familial information, rather than the resemblance arising from herd structure or lactation stage.

The familial variation observed for the derived FE phenotype in the dam–daughter regression agreed with that from the complementary among-sire linear model. As farm was modeled as a fixed effect in both the among-sire and dam–daughter analyses, and the sire effect remained significant in the cross-farm subanalysis, any imbalance in farm representation does not confound the results. Together, these facts support the interpretation that the phenotype captures biological variation consistent with a heritable component.

The indicative h² for the derived FE phenotype (~0.21) is consistent with published heritability estimates for related feed efficiency traits in dairy cattle; approximately 0.20 to 0.32 for gross feed efficiency (ECM/DMI; Spurlock et al., 2012) and 0.15 to 0.18 for residual feed intake (Tempelman et al., 2015). These results are also in line with evidence linking behavioral traits to feed efficiency and its genetic architecture (Nascimento et al., 2024). Because the analysis is based on phenotypic relationships rather than pedigree- or genomic-based animal models, it does not constitute formal heritability estimation; precise genetic parameter estimates would require such methods. Nevertheless, the alignment with the published range suggests that the derived FE phenotype retains sufficient biological signal for a meaningful genetic component to be detectable. FE is, however, a ratio trait, and direct selection on a ratio distributes selection pressure across its component traits disproportionately, so the resulting genetic change in those components is difficult to predict (Gunsett, 1984; Islam et al., 2020). Because FE combines collar-predicted intake with independently measured milk output, part of its familial signal originates in the milk component rather than in the collar-predicted intake; the FE estimates reported here should therefore not be read as evidence that the collar captures efficiency directly. Predicted DMI, by contrast, is a linear trait and is not subject to this behavior.

Taken together, these results provide converging evidence that collar-predicted DMI, and the FE derived from it, capture biologically meaningful familial variation, at a magnitude broadly consistent with published genetic parameters for related traits, and that they can be evaluated at scale even in the absence of direct intake measurements.

Implications for management and breeding

Predicted DMI, and the FE derived from it, also have direct practical value beyond their candidacy for genetic evaluation: because it is estimable at the individual level from data already collected in commercial systems, it can support cow-level decisions such as identifying lower-efficiency animals and informing culling strategy, alongside related economic metrics such as income over feed.

The observed familial patterns also suggest that the derived FE phenotype could, following formal genetic evaluation, complement existing selection criteria. Such evaluation would require applying pedigree- or genomic-based animal models to data from herds with both collar-based sensor deployment and genotyped animals, enabling estimation of genomic heritability and genomic correlations with relevant trait complexes.

Future perspectives and model development

The model was trained and validated on housed cows offered a total mixed ration; performance in grazing systems is unknown and would require separate validation. Further improvements to the current approach may be achieved by expanding the diversity of training data to include a broader range of production systems, diets, and environmental conditions. Because the commercial dataset comprised Holstein cows exclusively, extending validation to additional breeds and to crossbred populations is a priority for establishing broader generalizability. Incorporating more granular inputs, such as milk fatty-acid profiles and individual component concentrations from in-line analysis, alongside body weight and health events, may further improve prediction accuracy and robustness.

A further direction is the integration of environmental and dietary measurements to extend the predicted phenotype toward enteric methane estimation, linking scalable efficiency phenotyping to the sustainability objectives that motivate it.

As precision dairy technologies continue to evolve, increasing availability of high-frequency data is expected to provide opportunities to refine DMI prediction models and integrate them into automated decision-support systems for real-time herd management.

Conclusions

In conclusion, this study demonstrates that collar-based behavioral data can be used to predict individual DMI, a measurement otherwise requiring research-scale intake infrastructure, at commercial scale using data streams already present in modern dairy systems, and that this predicted intake exhibits familial variation. Feed efficiency, the ultimate commercial objective motivating this work, can in turn be derived from the collar-based intake, and the derived FE phenotype likewise exhibits familial variation on a commercial scale, providing preliminary evidence to justify further genetic evaluation for breeding applications. Together, these findings support the transition toward scalable, data-driven decision-making in precision dairy farming.

Correction note

This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Claude Sonnet 4.6 and Claude Opus 5.0 (claude.ai, Anthropic) were used to assist with editing and improving the English language and flow of the manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fanim.2026.1897979/full#supplementary-material

Descriptive statistics of the five commercial Israeli dairy farms used for the inherited variation analysis. Farm labels are anonymized. Qualifying sires: bulls with ≥10 daughters in the commercial dataset. Dam–daughter pairs are those qualifying for the feed efficiency analysis (n = 617); the predicted DMI analysis used a larger set of 665 pairs, as it does not require a qualifying ECM record. Supplementary material at the source

Distribution of qualifying daughters by sire and commercial farm for the 25 sires included in the cross-farm subanalysis. Values represent the number of qualifying daughters per sire per farm. Only sires with at least three qualifying daughters in each of at least two commercial farms are shown, ordered by total daughters (descending). Sire × farm combinations represented by a single daughter (n = 8, all Farm A) were excluded from the adjusted model; analyzed n = 553. Number of farms is the number of farms at which the sire had any qualifying daughter, including single-daughter combinations excluded from the adjusted model. Farm labels are anonymized. — = absent from this farm. Supplementary material at the source

Footnotes

Edited by: Nicolas Lyons, DairyNZ, New Zealand

Reviewed by: Jan Lassen, AU Foulum, Denmark

Alejandro Palladino, National Scientific and Technical Research Council (CONICET), Argentina