Development of prediction model for body weight and energy balance indicators from milk traits in lactating dairy cows based on deep neural networks

1 Introduction

The monitoring of the energy balance (EB) of high-yielding dairy cows during the lactation period is important (Nigussie, 2018) because it is directly related to milk production and reproductive performance (Heuer et al., 2001) and, ultimately, the profitability of dairy enterprises. Several methods have been proposed to estimate EB using body weight (BW) changes, body condition scores (Friggens et al., 2007), and analysis of metabolites in blood and milk (Moore et al., 2005). However, these methods are difficult to apply to large herds (Alvarez et al., 2018), making monitoring of the individual EB of cows in the field challenging. EB can also be evaluated as the difference between the measured energy intake (feed intake) and demand (milk production and maintenance); however, this requires measurements of milk yield and composition, BW, dry matter intake (DMI), and energy density of feedstuff (Mäntysaari and Mäntysaari, 2015), which are not broadly available on commercial farms (Yan et al., 2009).

Recording the daily BW, milk yield, and milk composition is possible through modern automatic milking systems (Mäntysaari and Mäntysaari, 2015). Although data from these automatic milking systems is used in modeling studies to predict and evaluate BW, DMI, EB, and milk yield (Caixeta et al., 2015), to date, these models require very detailed information, which has limited their adoption in commercial dairy farms (Vanrobays et al., 2015). Additionally, a range of other factors, such as the stage of lactation, parity, and a cow’s individual characteristics, which also need to be considered by the prediction model, affect BW and EB.

Multiple linear regression (MLR) is one of the most widely used modeling approaches for agricultural applications (Basak et al., 2020a). Although it is a powerful modeling technique, it assumes that the relationship between independent and dependent variables is linear. This assumption of linearity may not always be correct, and can lead to biased results that fail to provide satisfactory prediction accuracy (Chen et al., 2022). Alternatively, local regression (LOESS), which is a nonparametric local regression model for performing nonlinear predictions, is used to address this limitation (Shamim et al., 2016).

Recently, machine learning algorithms, such as artificial neural networks or deep neural networks (DNN), have become popular as powerful learning methods that are particularly beneficial for modeling nonlinear and complex relationships between variables (Chen et al., 2022). A DNN, which is an extension of an artificial neural network, tends to outperform the latter in direct comparisons using the same dataset (e.g., Guo et al., 2021). Further, DNN models also have better predictive performance than traditional methods (Ruchay et al., 2021). DNN models have been used in studies of dairy cows, including the estimation of body condition scores and BW through image processing, animal identification, breeding classification, and heat detection (Chowdhury et al., 2016; Shen et al., 2020).

In this study, we compared the prediction accuracy of BW using three different models based on MLR, LOESS, and DNN and presented a decision-making support system to evaluate daily EB for individual cows. Automatic milking systems were used to record the daily BW and milk yield during the lactation period. Additional information readily available on commercial dairy farms, such as milk traits, days in milk (DIM), and parity was also included in the models.

2 Material and methods

2.2

2.2 Input variable selection

In each modeling method, the selection of input variables plays a crucial role in determining a suitable model structure (Basak et al., 2020b). In the present study, principal component analysis was conducted to identify the main variables in the automatic milking system data (Fig. 1).

Close

Fig. 1

Milk production variables in the first and second principal components. DIM, days in milk; DMI, dry matter intake; EB, energy balance; ECM, energy corrected milk; FE, feed efficiency; FPR, milk fat-to-protein ratio; SCC, somatic cell counts.

Export to PPT

From these, energy-corrected milk (ECM), DIM, fat-to-protein ratio, and parity, which are readily available on commercial dairy farms, were selected as input variables for all models. ECM provides a more precise representation of cows' energy output compared to milk yield alone, as it accounts for milk yield adjusted to the ratio of milk solids (Knob et al., 2021). The lactation phase, parity, milk production, and fat-to-protein ratio exhibit significant predictability for EB (Heuer et al., 2000). The ECM (Shirley, 2006) and fat-to-protein ratios (Lee et al., 2017) were calculated as follows:

(3)

$$\mathrm{ECM}(\mathrm{kg}/\mathrm{d})=0.327\times \mathrm{milk}\mathrm{yield}\left(\mathrm{kg}\right)+7.2\times \mathrm{milk}\mathrm{protein}\left(\mathrm{kg}\right)+12.95\times \mathrm{milk}\mathrm{fat}\left(\mathrm{kg}\right)$$

(4)

$$\mathrm{Milk}\mathrm{fat}‐\mathrm{to}‐\mathrm{protein}\mathrm{ratio}=\frac{\mathrm{milk}\mathrm{fat}\left(\%\right)}{\mathrm{milk}\mathrm{protein}\left(\%\right)}$$

3 Results

3.1

3.1 Prediction of BW

To obtain an optimal prediction model for BW, predictive performance was tested and compared using three different methods. As shown in Table 2, the DNN model had the lowest RMSE (32.92) and MAE (25.65) when compared to the MLR and LOESS models in the tenfold cross-validation. As the DNN model had a higher accuracy in this study than the other models, we used the BW predicted by this model to calculate the energy demand. The ECM and DMI patterns during lactation are shown in Fig. 2. The ECM increased sharply during the first 8–9 weeks (DIM 61) of lactation, after which it decreased. In contrast, the DMI increased slowly until approximately 14 weeks (DIM 100).

Table 2 Predictive performance of different modeling approaches for prediction of body weight (BW) using ten-fold cross validation.

Models	RMSE	MAE
Multiple Linear Regression (MLR)	50.94	38.15
Local Regression (LOESS)	40.93	32.73
Deep Neural Network (DNN)	32.92	25.65

The features were days in milk (DIM), energy corrected milk (ECM), fat-to-protein ratio, and parity. MAE, mean absolute error (obtained using ten-fold cross validation); RMSE, root mean square error (obtained using ten-fold cross validation).

Close

Fig. 2

Relationship between days in milk (DIM) and (A) the mean of energy corrected milk (ECM) and (B) dry matter intake (DMI) in Holstein cows (mean ± SD).

Export to PPT

3.2

3.2 Calculation of energy balance

The estimated energy demand, energy intake, and EB during lactation are shown in Fig. 3. The mean daily energy demand was high after parturition up to DIM 61, after which it began to decline (Fig. 3A). The mean daily energy intake was low after parturition and peaked at a DIM of 100. These patterns of changes in energy intake and demand during lactation were similar to those of the DMI and ECM results (Fig. 2). This was expected, as daily energy intake was computed from DMI, and the net energy of the total mixed ration and daily energy demand were determined as the sum of energy demand for milk production and maintenance. As shown in Fig. 3C, the daily mean EB was negative after parturition and then increased to a positive value.

Close

Fig. 3

Relationship between days in milk (DIM) and (A) the mean of energy demand (ED), (B) energy intake (EI), and (C) energy balance (EB) in Holstein cows (mean ± SD).

Export to PPT

4 Discussion

Monitoring the EB of individual cows is essential for their proper management and breeding (Mäntysaari et al., 2019). It assists farmers in recognizing cows that might be prone to metabolic stress and production diseases, while also verifying the adequacy of existing management and nutritional approaches. Changes in the EB throughout a cow's lifespan might serve as a valuable prospective selection objective due to the genetic differences in EB profiles observed among bull daughter groups in their initial lactation (Coffey et al., 2001). Although several methods have been proposed to estimate EB using BW changes, body condition scores (Friggens et al., 2007), and analysis of metabolites in blood and milk (Moore et al., 2005), these are difficult to apply to large herds (Coffey et al., 2001; Alvarez et al., 2018), making it challenging to monitor the individual EB of cows in the field. Therefore, an easy and effective method for monitoring EB in cows is required. The daily EB can be determined by subtracting the measured energy intake from the demand (GfE, 2001). However, this calculation requires BW and milk yield and composition measurements, which can be challenging to acquire at the farm level (Mäntysaari et al., 2019).

This study predicted BW in lactating cows based on milk trait (ECM, DIM, and fat-to-protein ratio) data using automatic milking systems and parity information, which are more broadly available on commercial farms. In our study, we utilized daily measurements of milk yield and composition instead of relying on monthly evaluations. Frequent measurements enabled us to smooth the milk production data prior to the modelling analysis. The predicted BW was used to calculate the energy required for maintenance (GfE, 2001). Using predicted BW has the advantage of being simple to integrate with automatic milking systems and sensor-based monitoring systems, allowing for almost continuous BW monitoring. The BW of cows decreases sharply during the first 3–5 weeks of lactation and then increases at the end of lactation (Vanrobays et al., 2015; Mäntysaari et al., 2019). During early lactation, insufficient feed intake triggers the mobilization of energy from body reserves, ultimately causing a decline in BW (Mäntysaari et al., 2012). In contrast, lost body reserves are restored later in lactation with elevated feed intake and reduced milk yield, leading to an increase in BW.

During the initial stages of lactation, the energy demands of high-producing cows are rarely met by their feed intake (Mäntysaari et al., 2012), which results in energy mobilization from their body reserves to make up for the energy deficit, causing NEB during the early lactation period. Notably, at least 80 % of dairy cows undergo NEB during early lactation (Nigussie, 2018). In general, when NEB occurs in the early lactation of dairy cows, EB reaches zero during mid-lactation and becomes positive in late lactation. Cows experiencing body tissue and energy loss in the early lactation typically reach PEB around DIM 40–80 (Coffey et al., 2001). However, several cases of negative EB during the mid- and late lactation periods in high-yielding dairy cows with relatively high milk yields were observed. In this study, cows on average achieved PEB at DIM 66, 105, and 200 for EB indexes 1 (early lactation period), 2 (mid-lactation period), and 3 (late lactation period), respectively. Coffey et al. (2001) reported that the cumulative body energy loss in the first lactation period was fully regained at approximately DIM 200. Further, the more delayed the transition from NEB to PEB, the higher the ECM and feed efficiency. This was due to the dairy cows mobilizing the necessary energy requirements from body fat to produce large amounts of milk, leading to the cows remaining in the NEB state until the mid-lactation period (Table 3). In addition, for EB index 3, the daily EB remained negative until mid-lactation, which suggests that milk productivity increases but reproductive efficiency may decrease. NEB leads to decreased fertility and metabolic disorders, such as ketosis and mastitis. (Puangdee et al., 2016), and severe NEB postpones early ovulation and recuperation of postpartum reproductive function and exerts carryover effects that diminish fertility during the breeding period (Nigussie, 2018). Moreover, postpartum reproductive activity may resume only once the nadir of NEB is reached (Coffey et al., 2002), indicating that the transformation from NEB to PEB could serve as a valuable sign of the restoration of reproductive activity.

These results suggest that the present model is an appropriate method for evaluating EB on a commercial farm without measuring BW daily. Monitoring the EB of individual cows has clear benefits from the perspective of using EB as a diagnostic tool for nutrition and reproduction. In addition, EB indexes can be used as indicators for farm management decision-making. These advanced modeling techniques offer concrete benefits to dairy farmers in real practice. The precise anticipation of BW and EB has a pivotal role in guiding decisions related to feed management, allowing for meticulous adjustments in the dietary plans of individual cows. By integrating readily accessible information, such as milk traits, parity, and DIM, the devised models can provide tailor-made recommendations for the specific nutritional requirements of each EB index group. This customized approach enhances feed utilization efficiency and enables economically efficient milk production, considering the reproductive efficiency of the next parity. Further, it facilitates the early identification and prompt intervention of metabolic disorders. The outcomes highlighted in this study underscore the potential significance of advancing dairy farming practices, thereby contributing to progress in sustainable livestock management.

Our research had some limitations that should be taken into account when interpreting the findings. The data used for model training and testing were derived from only one farm, which could have contributed to an unbalanced distribution of BW values for model training, and farm-specific BW and EB patterns may exist. Therefore, models must be trained on data from a larger number of farms to ensure the robustness of the predictions.

5 Conclusion

We developed a BW prediction model for individual cows using milk production traits and parity information and estimated their daily EB based on the predicted BW. In this study, milk production traits readily available on commercial dairy farms were used as input variables for BW prediction. The DNN model demonstrated the highest predictive accuracy during the lactation period, outperforming the LOESS and MLR models in the ten-fold cross-validation. This investigation highlighted variations in the transition of EB status from negative to positive among cows, leading to the classification of cows into three EB index groups based on DIM, which captured different EB transition patterns. Notably, EB index 3 exhibited the slowest transition from negative to positive EB, accompanied by the highest FE and ECM values. The EB of individual cows during lactation can be estimated and monitored with moderately high accuracy using EB indexes. In conclusion, EB indexes could be used as indicators for individual and herd management. Future work will aim to validate these models on multiple dairy farms.

Ethics approval

This study was approved by the Institutional Animal Care and Use Committee of Jeonbuk National University, Korea (No. 2020–1958).

CRediT authorship contribution statement

Eunjeong Jeon: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Visualization, Writing – original draft, Writing – review & editing. Sangbuem Cho: Conceptualization, Data curation, Formal analysis, Software, Investigation, Methodology, Visualization, Writing – review & editing. Seongsoo Hwang: Investigation, Methodology, Software, Validation, Writing – review & editing. Kwanghyun Cho: Investigation, Methodology, Resources, Software, Writing – review & editing. Cedric Gondro: Investigation, Methodology, Software, Validation, Writing – review & editing. Nag-Jin Choi: Conceptualization, Data curation, Investigation, Methodology, Project administration, Funding acquisition, Writing – review & editing.