Comparison of methods for predicting cow composite somatic cell counts

Abstract

One of the most common and reliable ways of monitoring udder health and milk quality in dairy herds is by monthly cow composite somatic cell counts (CMSCC). However, such sampling can be time consuming, and more automated sampling tools entail extra costs. Machine learning methods for prediction have been widely investigated in mastitis detection research, and CMSCC is normally used as a predictor or gold standard in such models. Predicted CMSCC between samplings could supply important information and be used as an input for udder health decision-support tools. To our knowledge, methods to predict CMSCC are lacking. Our aim was to find a method to predict CMSCC by using regularly recorded quarter milk data such as milk flow or conductivity. The milk data were collected at the quarter level for 8 wk when milking 372 Holstein-Friesian cows, resulting in a data set of 30,734 records with information on 87 variables. The cows were milked in an automatic milking rotary and sampled once weekly to obtain CMSCC values. The machine learning methods chosen for evaluation were the generalized additive model (GAM), random forest, and multilayer perceptron (MLP). For each method, 4 models with different predictor variable setups were evaluated: models based on 7-d lagged or 3-d lagged records before the CMSCC sampling and additionally for each setup but removing cow number as a predictor variable (which captures indirect information regarding cows' overall level of CMSCC based on previous samplings). The methods were evaluated by a 5-fold cross validation and predictions on future data using models with the 4 different variable setups. The results indicated that GAM was the superior model, although MLP was equally good when fewer data were used. Information regarding the cows' level of previous CMSCC was shown to be important for prediction, lowering prediction error in both GAM and MLP. We conclude that the use of GAM or MLP for CMSCC prediction is promising.

Comparison of machine learning methods to predict udder health status based on somatic cell counts in dairy cows

“…Predictions obtained using different ML methods on data collected at different time points were reported by Anglart et al 24 , who predicted cow composite SCC by using quarter and cow milk data regularly recorded in cows milked in an automatic milking system in a 8-week trial. The authors evaluated three ML methods (generalized additive model, RF, and multilayer perceptron), developing models with different variables setups, and found generalized additive model and multilayer perceptron to be promising for udder health prediction.…”

“…Anglart et al 24 also suggested to include information on cows’ previous composite SCC in model training to lower the prediction error. Although we recognize that predictions between monthly TD sampling would represent an asset for mastitis monitoring, such application is feasible only if automatic milking systems are used by the farmers, which is not the case of the current situation in Italy.…”

“…Nevertheless, in those studies both outcome (mastitic or healthy according to SCC value) and features (milk traits) were recorded at the same milking time. Predictions based on data collected at different time points were reported by Anglart et al 24 , who explored different ML methods for predicting cow composite SCC by using quarter and cow milk data regularly recorded in cows milked with automatic milking system. However, data spanned a 2-mo period and were collected in only one dairy farm.…”

Detecting and predicting changes in milk homogeneity using data from automatic milking systems

“…Alternative methods for prediction that include parameter estimation, for example, generalized additive mixed models, and thus the possibility of evaluating the contribution of the different predictor variables, were considered. However, because the aim of the study was prediction performance rather than inference, and based on previous experiences with both generalized additive mixed models and MLP (Anglart et al, 2020), the MLP was chosen for this study.…”

“…The multilayer perceptron (MLP), a classic feed-forward artificial neural network (ANN), was used. Artificial neural networks have potential to capture nonlinear relationships and interactions between predictors in a flexible manner and have previously been suggested for CM detection (e.g., Nielen et al, 1995;Sun et al, 2010;Ankinakatte et al, 2013), pathogen prediction (Heald et al, 2000;Hassan et al, 2009), and SCC prediction (Anglart et al, 2020). Similar to a milker, who uses faculties such as taste, smell, vision, and memory (Hillerton, 2000) to decide whether or not the milk should be discarded, ANN are designed to process information in a similar way, basing decisions on detected patterns and relationships in data and learning from them (e.g., Agatonovic-Kustrin and Beresford, 2000;Haykin, 2009).…”