Transformation of feed energy ingested by ruminants into milk is accompanied by energy losses via fecal and urine excretions, fermentation gases and heat. Heat production may differ among dairy cows despite comparable milk yield and body weight. Therefore, heat production can be considered an indicator of metabolic efficiency and directly measured in respiration chambers. The latter is an accurate but time-consuming technique. In contrast, milk Fourier transform mid-infrared (FTIR) spectroscopy is an inexpensive high-throughput method and used to estimate different physiological traits in cows. Thus, this study aimed to develop a heat production prediction model using heat production measurements in respiration chambers, milk FTIR spectra and milk yield measurements from dairy cows.

Heat production was computed based on the animal’s consumed oxygen, and produced carbon dioxide and methane in respiration chambers. Heat production data included 168 24-h-observations from 64 German Holstein and 20 dual-purpose Simmental cows. Animals were milked twice daily at 07:00 and 16:30 h in the respiration chambers. Milk yield was determined to predict heat production using a linear regression. Milk samples were collected from each milking and FTIR spectra were obtained with MilkoScan FT 6000. The average or milk yield-weighted average of the absorption spectra from the morning and afternoon milking were calculated to obtain a computed spectrum. A total of 288 wavenumbers per spectrum and the corresponding milk yield were used to develop the heat production model using partial least squares (PLS) regression.

Measured heat production of studied animals ranged between 712 and 1470 kJ/kg BW^{0.75}. The coefficient of determination for the linear regression between milk yield and heat production was 0.46, whereas it was 0.23 for the FTIR spectra-based PLS model. The PLS prediction model using weighted average spectra and milk yield resulted in a cross-validation variance of 57% and a root mean square error of prediction of 86.5 kJ/kg BW^{0.75}. The ratio of performance to deviation (RPD) was 1.56.

The PLS model using weighted average FTIR spectra and milk yield has higher potential to predict heat production of dairy cows than models applying FTIR spectra or milk yield only.

The conversion of feed energy (gross energy) ingested by ruminants into human-edible food energy such as milk and meat is accompanied by energy losses in form of fecal and urine excretions, fermentation gases (e.g. methane) and heat. For dairy cows, the proportion of feed converted to milk is defined as feed conversion ratio (FCR) which is only one of numerous measures of efficiency. High-feed efficient cows were described to have a better organic matter digestibility thus providing more digestible energy relative to less efficient counterparts [

The measurement of HP from ruminants can be performed by indirect calorimetry. To this end, the daily amount of oxygen (O_{2}) consumption, carbon dioxide (CO_{2}) production and methane (CH_{4}) emission along with urinary nitrogen (N_{u}) excretion is measured in respiration chambers (RC) [

Infrared spectroscopy is a high-throughput method and its broad application in livestock sector became more popular in the last few years. For the analysis of milk components e.g., fat, protein, lactose and urea, Fourier transform mid-infrared (FTIR) spectrometry has been developed and refined, and nowadays is frequently used routinely to control milk quality and energy [_{4} emission [

The obtained data for this study originated from four different experimental projects, which have been conducted at the Leibniz Institute for Farm Animal Biology (FBN) in Dummerstorf, Germany (Additional file

Animal handling during all projects was carried out according to the instructions for the use of animals as experimental subjects of the State Government in Mecklenburg-Western Pomerania. Experimental protocols were confirmed by the local animal ethics committee (Landesamt für Landwirtschaft, Lebensmittelsicherheit und Fischerei Mecklenburg-Vorpommern). Animals were transferred into RC with a gas recovery rate of 99.9% ± 0.96% and a light cycle of 06:00 to 19:00 h at 15 °C [

Ingredients, nutrient composition and energy content of the diets fed to the animals during the respiration chamber experiments

Item | Minimum | Maximum | Mean | Median | SD |
---|---|---|---|---|---|

Ingredients, g/kg of DM | |||||

Grass silage | 500 | 915 | 692 | 680 | 81 |

Corn silage | 238.0 | 452.5 | 360.0 | 375.0 | 93.3 |

Barley straw | 42 | 423 | 214 | 259 | 93 |

Concentrate | 4.6 | 27.9 | 17.7 | 17.8 | 3.4 |

Nutrient composition | |||||

DM, g/kg | 353 | 941 | 529 | 388 | 235 |

NDF^{a}, g/kg of DM | 275 | 497 | 377 | 377 | 51 |

ADF^{b}, g/kg of DM | 146 | 254 | 197 | 197 | 23 |

Crude protein, g/kg of DM | 122 | 171 | 152 | 155 | 13 |

Crude fat, g/kg of DM | 24 | 33 | 30 | 30 | 2 |

Ash, g/kg of DM | 38 | 135 | 71 | 69 | 19 |

ME^{c}, MJ/kg of DM | 8.5 | 11.6 | 10.3 | 10.5 | 0.7 |

^{a}Neutral detergent fiber

^{b}Acid detergent fiber

^{c}Metabolizable energy

The airflow through the chamber was roughly 30 m^{3} per hour and measured by a differential-pressure type V cone flow meter (McCrometer, Hemet, CA). Concentrations of CO_{2} and CH_{4} in the chamber were analyzed by infrared-absorption and the concentration of O_{2} was analyzed paramagnetically (SIDOR SICK AG, Waldkirch, Germany) every 6 min. The body weight (BW) of the animals was determined directly before and after the stay in the RC to calculate the mean metabolic body weight (mBW):

Heat production was computed using the Brouwer [_{2} consumption and CO_{2} and CH_{4} production and normalized to mBW:

The N_{u} excretion was estimated to be 150 g/d based on comparable diet compositions and feed intake levels [^{− 1}. From these, selected wavenumbers (^{− 1}. The spectra regions were shown to contain crucial information on different milk constitutes [

To analyze the predictive ability of milk yield alone on HP, we first applied a univariate linear model using the “lm” function in R [

,where Y_{i} is the HP of animal as response variable, ß_{i} is the regression coefficient, and ε_{i} is the random residual. Secondly, the FTIR spectra obtained from the morning and afternoon milking were averaged to estimate HP without considering milk yield (M1). In a third approach, the average of two absorption spectra from the evening and morning milking was calculated and subsequently multiplied by the corresponding daily milk yield (M2). Finally, the milk yield-weighted average of the spectra was applied to, in which absorption spectra from morning and afternoon milking were multiplied with the respective milk yield of the morning and afternoon milking (M3). Partial least squares (PLS) regression is widely accepted as the preferred method to analyze the potential relationship between the predictor, i.e. spectral data, and the related physiological outcome [

where _{i} and _{i} depict the observed and predicted HP, respectively. The square root of the MSEP (RMSEP) indicates the overall error of the prediction. We carried out a random cross validation in order to observe the performance of the developed prediction models with 10 splits and 10 iterations and the outcome was shown as root mean squared error of cross validation (RMSECV) and the coefficient of determination of cross validation (R^{2}CV). This approach allows the observations to be utilized for both calibration and validation, hence each observation used for validation exactly once. Both RMSECV and R^{2}CV values were the average of a 10-fold cross validation [^{2} of external cross-validation (R^{2}V) values were computed based on the average of 4 iterations of the external cross-validation. The optimal number of Latent variables (i.e., PLS components) for the prediction model was chosen based on visual observation of the cross-validation RMSE plot against the PLS factors and determining the lowest RMSE.

The concordance correlation coefficient (CCC) was computed using the “epiR” package implemented in R [

The descriptive analysis of the animal performance, daily gas measurement and HP (kJ/kg BW^{0.75}) is shown in Table ^{0.75}, respectively.

Descriptive analysis of animal performance, gas exchange measurements and computed heat production (

Item | Minimum | Maximum | Mean | Median | SD |
---|---|---|---|---|---|

Animal performance | |||||

BW, kg | 500 | 915 | 692 | 680 | 80 |

mBW^{a}, kg^{0.75} | 105.7 | 166.4 | 134.8 | 133.3 | 11.7 |

Lactation number | 1.0 | 10 | 2.9 | 3.0 | 1.6 |

Days in milk, d | 42.0 | 423.0 | 213.5 | 259.0 | 92.8 |

DMI, kg/d | 4.6 | 27.9 | 17.7 | 17.8 | 3.4 |

Milk yield, L/d | 5.2 | 51.4 | 25.7 | 24.0 | 10.3 |

Gas exchange measurements | |||||

O_{2}, L | 4686 | 9238 | 6786 | 6740 | 720 |

CO_{2}, L | 4615 | 9313 | 7072 | 7099 | 908 |

CH_{4}, L | 299 | 792 | 570 | 579 | 81 |

Heat production, kJ/kg BW^{0.75} | 712 | 1469 | 1067 | 1068 | 136 |

^{a}Metabolic bodyweight (mBW = BW^{075})

The three spectral regions retained (968–1577, 1720–1808, and 2564–2965 cm^{− 1}) for the prediction model are those, which are typically used for quantifying milk fat, protein as well as lactose contents. This was confirmed by the outcome of the loading value plot (Additional file ^{− 1} to be most important for estimating HP. The wavenumbers around 1175 cm^{− 1} stretching the triacylglycerol ester C–O linkage, C=O stretching (approx. 1750 cm^{− 1}), and acyl chain C–H symmetric and asymmetric stretching (2800–3000 cm^{− 1}) are generally important to determine milk fat content [^{− 1}) are utilized to assess milk protein content [^{− 1}) can be used to evaluate carbohydrates such as lactose [

The linear model L1 resulted in the coefficient of determination of 0.46. The M1 prediction model, without consideration of milk yield, resulted in a RMSEP value of 99.9 kJ/kg BW^{0.75} and a R^{2} value of 0.23 (Table ^{2}CV value of 86.7 kJ/kg BW^{0.75} and 0.25, respectively. The RMSEV and R^{2}CV values of the external validation of M1 were 114.1 kJ/kg BW^{0.75} and 0.18, respectively. The PLS prediction model M2, involving the average of the morning and afternoon spectra as well as milk yield, resulted in a RMSEP and R^{2} of 93.2 kJ/kg BW^{0.75} and 0.52, respectively (Table ^{2}CV values were 89.4 kJ/kg BW^{0.75} and 0.55, respectively, and for the external validation 84.0 kJ/kg BW^{0.75} and 0.48, respectively. The results for the prediction model M3, considering the weighted average milk spectra and milk yield, were even slightly better with a RMSEP and R^{2} of 91.2 kJ/kg BW^{0.75} and 0.54, respectively (Table ^{2}CV of 86.5 kJ/kg BW^{0.75} and 0.57, respectively. However, the external validation approach of the M3 model showed RMSEV and R^{2} values of 95.5 kJ/kg BW^{0.75} and 0.47, respectively. These results show that the involvement of FTIR spectra improves the prediction accuracy of HP as compared to the L1 model considering milk yield only.

The statistics of partial least square regression approach for the milk Fourier transform mid-infrared spectrometry-based estimation model for heat production of dairy cows

Trait | Prediction model | Calibration | LV^{c} | Cross Validation | External Validation | |||
---|---|---|---|---|---|---|---|---|

R^{2} | RMSEP^{b} | R^{2}CV | RMSECV^{d} | R^{2}V | RMSEV^{e} | |||

Heat production, kJ/kg BW^{0.75} | M1^{a} | 0.23 | 99.9 | 14 | 0.25 | 86.7 | 0.18 | 114.1 |

M2^{a} | 0.52 | 93.2 | 4 | 0.55 | 89.4 | 0.48 | 84.0 | |

M3^{a} | 0.54 | 91.2 | 5 | 0.57 | 86.5 | 0.47 | 95.5 |

^{a}Model M1 was developed using the averaged morning and afternoon spectral data. The prediction model M2 was developed by averaging the morning and afternoon spectral data and subsequent multiplication with daily milk yield. The prediction model M3 was computed by weighted averaging, where each morning or afternoon absorption spectra was multiplied to the respective milk yield

^{b}The square root of the mean squared error of prediction

^{c}Latent variables; i.e. the partial least square regression components for the prediction model

^{d}Root mean squared error of cross validation

^{e}Root mean squared error of external validation

The predicted against observed and residuals versus predicted HP of the FTIR spectra- and milk yield-based estimation models M2 and M3 are shown in Fig. _{4} estimation model than utilizing FTIR wavenumbers only. These earlier findings agree with our results showing that applying milk yield in the final transformation of the spectra as a factor slightly improved the performance of models than using milk yield as a predictor variable. Furthermore, previous studies indicated that the use of milk yield-weighted average compared to the average of FTIR spectra is more biologically relevant [^{2} between calibration and validation (< 10%) between M2 and M3 show the fair robustness of both prediction models. The concordance correlation coefficient (CCC) analysis pinpointed a substantial predictive ability of our FTIR-based models (CCC of 0.67 and 0.71 for M2 and M3 models, respectively) [_{4} emission or body energy status, yet clear disparities between model performances can be observed [_{4} production estimation model using observations obtained from a non-invasive CH_{4} measurement approach using the sniffer method [_{4} production from milk FTIR wavenumbers resulted in a R^{2} of validation of 0.13 only [_{4} emission in dairy cows by generating a prediction PLS model for CH_{4} measurements acquired from RC experiments. Their prediction model showed a R^{2}CV of 0.30 for CH_{4} production [^{2}CV of our present models, particularly M2 and M3 models. On the other hand, Smith et al. [^{2} of internal and external cross-validation of 0.77 and 0.60, respectively [^{2} values of our models. Vanlierde et al. [_{4} measurements executed in RC of different European cattle research centers, including different breeds and feeding managements. The predictive ability of the model for CH_{4} emission indicated R^{2} of calibration and cross-validation of 0.65 and 0.57, respectively [^{2} values of FTIR-based models of this work. The modest R^{2} values obtained from the present estimation models call for a need to enlarge the current data set, but this can only be achieved in future studies. Besides, CH_{4} production is only one component of HP, and the variation of O_{2} consumption and CO_{2} production as further variables of HP certainly contribute to the lower R^{2} as well. Another explanation may be that we considered multiple measurements on the same animal as independent due to changes in diet or physiological state. On the other hand it has been shown that the R^{2} value is improved when repeated CH_{4} emission measurements i.e., recorded during a 7-day measurement period, were incorporated in the prediction model [_{4} production from milk fatty acids analyzed by gas-chromatography. The same limitation might be applicable for the HP estimation in the present study, as spectra wavenumbers associated with milk fat content is apparently among the important regions for predicting HP level of dairy cows.

Predicted against observed measurements (

Predicted against observed measurements (

The coefficient of determination greatly relies on the range and variability of the observations. Despite considering observations from experiments containing different dietary compositions and quite a range of DMI, BW and milk yield, the current FTIR-based models seem still not to have covered the high-enough variability to reach a higher rate of HP predicting accuracy. The higher predictive ability of models generated by Dehareng et al. [^{2}CV of 0.68–0.79 and 0.77, respectively, is likely due to the wide range of CH_{4} measurements, which were between 218–653 g/d and 180–950 g/d, respectively. A recent study combined the CH_{4} production data obtained from RC and the sulphur hexafluoride (SF_{6}) tracer technique, thereby enlarging the dataset to 1089 observations from 299 cows, and thus the variability of CH_{4} production (286–546 g/d) [^{2} of validation of 0.64 vs. 0.57), again underscoring the need of more data sets for improving the HP prediction model in future.

We also determined the applicability of our models by analyzing the ratio of performance to deviation (RPD). The RPD value is based on the relation between the standard error of prediction to the standard deviation of the referenced observation, in which the higher RPD values pinpoints the more suitability of the model for future screening, quality control as well as any other applications [_{4} production (RPD = 1.19). Conversely, the RPD of our models was lower than the FTIR-based model by Dehareng et al., who reported a RPD of 2.19 [

The current work was the first to show the potential use of milk FTIR spectra to predict HP of dairy cows. The outcome of this study, at least at preliminary stages, revealed that milk FTIR spectra together with milk yield can potentially be used to identify dairy cows with different HP and thus feed efficiency levels. Selection for animals that utilize feed more efficiently would be ideal for dairy managers in order to benefit the financial returns. The FTIR-based models were robust with evident predictive ability; however, they can only describe a moderate part of the observed HP variations. The applicability of the prediction models remained relatively poor, which implies the need to enlarge the current data set. Future work exploiting higher number of observations from a wider range of breeds and feeding regimens seems warranted in order to further ameliorate the quality of the model.

^{− 1}). The grey shaded area depicts the spectra regions selected for generating the present partial least square models (968–1577, 1720–1808, and 2564–2965 cm^{− 1}).

Concordance correlation coefficient

_{4}

Methane

_{2}

Carbon dioxide

Fourier transform mid-infrared

Heat production

Metabolic body weight

Mean squared error of prediction

_{2}

Nitrogen

_{2}

Oxygen

Partial least square

Respiration chambers

^{2}CV

Coefficient of determination of cross validation

Root mean squared error of cross validation

Ratio of performance to deviation

The help of the staff at the “Tiertechnikum” and the cattle experimental unit at the Leibniz Institute for Farm Animal Biology (FBN), Dummerstorf (Germany) is greatly acknowledged in regards to animal care and handling as well as sample collection and preparation.

All experiments followed the instructions for the use of animals as experimental subjects of the State Government in Mecklenburg-Western Pomerania. The approval numbers are: 7221.3–1.1-053/13, 7221.3–1-052/17, 7221.3–1-016/18, and 7221.3–1.1-041/18.

SDM and BK conceived and designed the study, and wrote the manuscript. SDM was responsible for preparing the data, developing and validating the model. PH carried out the milk spectroscopy analysis. AE assisted with the prediction model development. MD collected and analyzed the data from respiration chamber experiments. All authors read and approved the manuscript.

One part of Experiment 1 (Supplementary Table 1) was executed within JPI FACCE program and another part in the optiKuh project, both financially supported by the German Federal Ministry of Food and Agriculture (BMBL) through the Federal Office for Agriculture and Food (BLE), grant number 2814ERA04A and 2817201313, respectively. Experiment 2 was performed within ERA-GAS program and financially supported by the BMBL through the BLE, grant number 2817ERA09C. Experiment 3 was financially supported by the BMBL through the Landwirtschaftliche Rentenbank (LR), grant number 28RZ3P077. Experiment 4 received funding from the core budget of the FBN. The authors acknowledge financial support for publication fom the Open Access Fond of the FBN and declare that the aforementioned funding parties had no role in the design of the study or in data collection, analysis, interpretation and writing of the manuscript.

All data generated or analyzed are available from the corresponding author on request.

Not applicable.

The authors declare that they have no competing interests.