A Machine Learning Framework Based on Extreme Gradient Boosting to Predict the Occurrence and Development of Infectious Diseases in Laying Hen Farms, Taking H9N2 as an Example

The H9N2 avian influenza virus has become one of the dominant subtypes of avian influenza virus in poultry and has been significantly harmful to chickens in China, with great economic losses in terms of reduced egg production or high mortality by co-infection with other pathogens. A prediction of H9N2 status based on easily available production data with high accuracy would be important and essential to prevent and control H9N2 outbreaks in advance. This study developed a machine learning framework based on the XGBoost classification algorithm using 3 months' laying rates and mortalities collected from three H9N2-infected laying hen houses with complete onset cycles. A framework was developed to automatically predict the H9N2 status of individual house for future 3 days (H9N2 status + 0, H9N2 status + 1, H9N2 status + 2) with five time frames (day + 0, day - 1, day - 2, day - 3, day - 4). It had been proven that a high accuracy rate > 90%, a recall rate > 90%, a precision rate of >80%, and an area under the curve of the receiver operator characteristic ≥ 0.85 could be achieved with the prediction models. Models with day + 0 and day - 1 were highly recommended to predict H9N2 status + 0 and H9N2 status + 1 for the direct or auxiliary monitoring of its occurrence and development. Such a framework could provide new insights into predicting H9N2 outbreaks, and other practical potential applications to assist in disease monitor were also considerable.

The Potentialities of Machine Learning for Cow-Specific Milking: Automatically Setting Variables in Milking Machines

In dairy farming, milking-related operations are time-consuming and expensive, but are also directly linked to the farm’s economic profit. Therefore, reducing the duration of milking operations without harming the cows is paramount. This study aimed to test the variation in different parameters of milking operations on non-automatic milking machines to evaluate their effect on a herd and finally reduce the milking time. Two trials were set up on a dairy farm in Northern Italy to explore the influence of the pulsation ratio (60:40 vs. 65:35 pulsation ratio) and that of the detachment flow rate (600 g/min vs. 800 g/min) on milking performance, somatic cell counts, clinical mastitis, and teats score. Moreover, the innovative aspect of this study relates to the development of an optimized least-squares support vector machine (LSSVM) classification model based on the sparrow search algorithm (SSA) to predict the proper pulsation ratio and detachment flow rate for individual cows within the first two minutes of milking. The accuracy and precision of this model were 92% and 97% for shortening milking time at different pulsation ratios, and 78% and 79% for different detachment rates. The implementation of this algorithm in non-automatic milking machines could make milking operations cow-specific.

Comparison and estimation of different linear and nonlinear lactation curve submodels in random regression analyses on dairy cattle

In the random regression model (RRM) for milk yield, by replacing empirical lactation curves with the five-order Legendre polynomial to fit fixed groups, the RRM can be transformed to a hierarchical model that consisted of a RRM in the first hierarchy with Legendre polynomials as individuals’ lactation curves resolved by restricted maximum likelihood (REML) software, and a multivariate animal model for phenotypic regression coefficients in the second hierarchy resolved by DMU software. Some empirical lactation functions can be embedded into the RRM at the first hierarchy to well fit phenotypic lactation curve of the average observations across all animals. The functional relationship between each parameter and time can be described by a Legendre polynomial or an empirical curve usually called submodel, and according to three commonly used criteria, the optimal submodels were picked from linear and nonlinear submodels except for polynomials. The so-called hierarchical estimation for the RRMs in dairy cattle indicated that more biologically meaningful models were available to fit the lactation curves; moreover, with the same number of parameters, the empirical lactation curves (MIL1, MIL5, and MK1 for 3, 4, and 5 parameters, respectively) performed higher goodness of fit than Legendre polynomial when modelling individuals’ phenotypic lactation curves.

Using quantile regression for fitting lactation curve in dairy cows

The main objective of this study was to compare the performance of different 'nonlinear quantile regression' models evaluated at the τth quantile (0·25, 0·50, and 0·75) of milk production traits and somatic cell score (SCS) in Iranian Holstein dairy cows. Data were collected by the Animal Breeding Center of Iran from 1991 to 2011, comprising 101 051 monthly milk production traits and SCS records of 13 977 cows in 183 herds. Incomplete gamma (Wood), exponential (Wilmink), Dijkstra and polynomial (Ali & Schaeffer) functions were implemented in the quantile regression. Residual mean square, Akaike information criterion and log-likelihood from different models and quantiles indicated that in the same quantile, the best models were Wilmink for milk yield, Dijkstra for fat percentage and Ali & Schaeffer for protein percentage. Over all models the best model fit occurred at quantile 0·50 for milk yield, fat and protein percentage, whereas, for SCS the 0·25th quantile was best. The best model to describe SCS was Dijkstra at quantiles 0·25 and 0·50, and Ali & Schaeffer at quantile 0·75. Wood function had the worst performance amongst all traits. Quantile regression is specifically appropriate for SCS which has a mixed multimodal distribution.

A new standard model for milk yield in dairy cows based on udder physiology at the milking-session level

Milk production in dairy cow udders is a complex and dynamic physiological process that has resisted explanatory modelling thus far. The current standard model, Wood’s model, is empirical in nature, represents yield in daily terms, and was published in 1967. Here, we have developed a dynamic and integrated explanatory model that describes milk yield at the scale of the milking session. Our approach allowed us to formally represent and mathematically relate biological features of known relevance while accounting for stochasticity and conditional elements in the form of explicit hypotheses, which could then be tested and validated using real-life data. Using an explanatory mathematical and biological model to explore a physiological process and pinpoint potential problems (i.e., “problem finding”), it is possible to filter out unimportant variables that can be ignored, retaining only those essential to generating the most realistic model possible. Such modelling efforts are multidisciplinary by necessity. It is also helpful downstream because model results can be compared with observed data, via parameter estimation using maximum likelihood and statistical testing using model residuals. The process in its entirety yields a coherent, robust, and thus repeatable, model.

Models to Estimate Lactation Curves of Milk Yield and Somatic Cell Count in Dairy Cows at the Herd Level for the Use in Simulations and Predictive Models

Typically, central milk recording data from dairy herds are recorded less than monthly. Over-fitting early in lactation periods is a challenge, which we explored in different ways by reducing the number of parameters needed to describe the milk yield and somatic cell count of individual cows. Furthermore, we investigated how the parameters of lactation models correlate between parities and from dam to offspring. The aim of the study was to provide simple and robust models for cow level milk yield and somatic cell count for fitting to sparse data to parameterize herd- and cow-specific simulation of dairy herds. Data from 610 Danish Holstein herds were used to determine parity traits in milk production regarding milk yield and somatic cell count of individual cows. Parity was stratified in first, second, and third and higher for milk, and first to sixth and higher for somatic cell count. Fitting of herd level parameters allowed for cow level lactation curves with three, two, or one parameters per lactation. Correlations of milk yield and somatic cell count were estimated between lactations and between dam and offspring. The shape of the lactation curves varied markedly between farms. The correlation between lactations for milk yield and somatic cell count was 0.2-0.6 and significant on more than 95% of farms. The variation in the daily milk yield was observed to be a source of variation to the somatic cell count, and the total somatic cell count was less correlated with the milk production than somatic cells per milliliter. A positive correlation was found between relative levels of the total somatic cell count and the milk yield. The variation of lactation and somatic cell count curves between farms highlights the importance of a herd level approach. The one-parameter per cow model using a herd level curve allows for estimating the cow production level from first the recording in the parity, while a two-parameter model requires more recordings for a credible estimate, but may more precisely predict persistence, and given the independence of parameters, these can be easily drawn for use in simulation models. We also conclude that using total somatic cell count may stabilize models, and therefore, the dilution factor is of importance in Danish Holstein.

On the analysis of Canadian Holstein dairy cow lactation curves using standard growth functions

Six classical growth functions (monomolecular, Schumacher, Gompertz, logistic, Richards, and Morgan) were fitted to individual and average (by parity) cumulative milk production curves of Canadian Holstein dairy cows. The data analyzed consisted of approximately 91,000 daily milk yield records corresponding to 122 first, 99 second, and 92 third parity individual lactation curves. The functions were fitted using nonlinear regression procedures, and their performance was assessed using goodness-of-fit statistics (coefficient of determination, residual mean squares, Akaike information criterion, and the correlation and concordance coefficients between observed and adjusted milk yields at several days in milk). Overall, all the growth functions evaluated showed an acceptable fit to the cumulative milk production curves, with the Richards equation ranking first (smallest Akaike information criterion) followed by the Morgan equation. Differences among the functions in their goodness-of-fit were enlarged when fitted to average curves by parity, where the sigmoidal functions with a variable point of inflection (Richards and Morgan) outperformed the other 4 equations. All the functions provided satisfactory predictions of milk yield (calculated from the first derivative of the functions) at different lactation stages, from early to late lactation. The Richards and Morgan equations provided the most accurate estimates of peak yield and total milk production per 305-d lactation, whereas the least accurate estimates were obtained with the logistic equation. In conclusion, classical growth functions (especially sigmoidal functions with a variable point of inflection) proved to be feasible alternatives to fit cumulative milk production curves of dairy cows, resulting in suitable statistical performance and accurate estimates of lactation traits.

Comparison of modelling techniques for milk-production forecasting

The objective of this study was to assess the suitability of 3 different modeling techniques for the prediction of total daily herd milk yield from a herd of 140 lactating pasture-based dairy cows over varying forecast horizons. A nonlinear auto-regressive model with exogenous input, a static artificial neural network, and a multiple linear regression model were developed using 3 yr of historical milk-production data. The models predicted the total daily herd milk yield over a full season using a 305-d forecast horizon and 50-, 30-, and 10-d moving piecewise horizons to test the accuracy of the models over long- and short-term periods. All 3 models predicted the daily production levels for a full lactation of 305 d with a percentage root mean square error (RMSE) of ≤ 12.03%. However, the nonlinear auto-regressive model with exogenous input was capable of increasing its prediction accuracy as the horizon was shortened from 305 to 50, 30, and 10 d [RMSE (%)=8.59, 8.1, 6.77, 5.84], whereas the static artificial neural network [RMSE (%)=12.03, 12.15, 11.74, 10.7] and the multiple linear regression model [RMSE (%)=10.62, 10.68, 10.62, 10.54] were not able to reduce their forecast error over the same horizons to the same extent. For this particular application the nonlinear auto-regressive model with exogenous input can be presented as a more accurate alternative to conventional regression modeling techniques, especially for short-term milk-yield predictions.

Lactation curves of dairy camels in an intensive system

Weekly milk records of 47 she-camels in a multibreed dairy camel herd were collected for over a period of 5 years. A total of 72 lactation curves were defined, and relationships with parity, calving season, lactation length, milk production level, following lactations, and dam weight were analyzed. Overall mean values were milk yield up to 12 months, 1,970 ± 790 l; lactation length, 12.5 months; persistency, 94.7 %; weekly peak yield, 50.7 l; monthly peak yield, 220 ± 90 l; and the number of weeks to reach peak yield, 28. The highest productivity was recorded in summer with a weekly mean of 48.2 ± 19.4 l, compared with 34.1 ± 16.3 l in winter. The highest average yield recorded was for camels at sixth parity, whereas the highest weekly peak was at eighth parity, and highest persistency at fifth parity. Camels that calved during the cold months (November to February) were most productives, with the highest persistency, peak yield, and longest lactation length. Four types of curves were identified corresponding to different parities and milk yield levels. Based on these data, specific models for camels are proposed.