Locomotion system issues are major issues in dairy herds, affecting both animal welfare and farm productivity. Early detection could improve the effectiveness of treatments and increase the chances to cure lame cows. Currently, locomo- tion issues detection requires direct observation of cows walk- ing (locomotion score). However, this is a time-consuming task and is not always an available option in large dairy farms. Aim of this preliminary study was to build a predictive model for locomotion system issues in Italian Holstein dairy cows using some novel phenotypes from automatic recording systems (milking parallel parlour and SCR Heatime and DataFlow2 system) as predictors. Data was recorded from a commercial farm located in the province of Mantua (Lombardy, Northern Italy) for a total of 413 animals, daily monitored for two years (Sept. 2014 - Dec. 2016). The response variable was binary (0/1: healthy and diseased, respectively). The selected variables were daily rumination time, parity, DIM, daily milk production, daily activity and month of recording. Summary statistics (mean and SD) were calculated: rumination time, 563.28 ± 88.48 min/day; parity, 1.93 ± 1.28; DIM, 171.34 ± 124.14 days; milk production, 24.14 ± 13.74 kg/day; activity, 614.00 ± 134.64 min/day. DIM were classified in four classes to assess the potential effect of the lactation stage: dry, early, mid, and late lactation (no lactation, <120 days, 120-240 days, and >240 days, respect- ively). Three different datasets were prepared, where rumin- ation, milk, and activity were averaged as means of 1, 3, and 5 days before the response variable record. On each dataset, two models were fitted: logistic regression and random forest. All the analyses were performed in R using the caret package. Data were divided into a training and a testing dataset (pro- portion 80/20). Training data was used to train two different algorithms which were used to predict the class variable. The two selected algorithms were: 1. a logistic regression 2. a ran- dom forest. For all the datasets, logistic regression was not able to predict diseased individuals, assigning all to the 'healthy' class. Random forest performed better, although with a high-class error. The 5-day window had the lowest OOB error rate (0.24%) and the lowest class error (0.71). Further tuning of the selected models will be necessary to build a valuable tool to predict locomotion system issues.
Predictive models for locomotion issues in Italian Holstein dairy cows
Chessa S;
2017
Abstract
Locomotion system issues are major issues in dairy herds, affecting both animal welfare and farm productivity. Early detection could improve the effectiveness of treatments and increase the chances to cure lame cows. Currently, locomo- tion issues detection requires direct observation of cows walk- ing (locomotion score). However, this is a time-consuming task and is not always an available option in large dairy farms. Aim of this preliminary study was to build a predictive model for locomotion system issues in Italian Holstein dairy cows using some novel phenotypes from automatic recording systems (milking parallel parlour and SCR Heatime and DataFlow2 system) as predictors. Data was recorded from a commercial farm located in the province of Mantua (Lombardy, Northern Italy) for a total of 413 animals, daily monitored for two years (Sept. 2014 - Dec. 2016). The response variable was binary (0/1: healthy and diseased, respectively). The selected variables were daily rumination time, parity, DIM, daily milk production, daily activity and month of recording. Summary statistics (mean and SD) were calculated: rumination time, 563.28 ± 88.48 min/day; parity, 1.93 ± 1.28; DIM, 171.34 ± 124.14 days; milk production, 24.14 ± 13.74 kg/day; activity, 614.00 ± 134.64 min/day. DIM were classified in four classes to assess the potential effect of the lactation stage: dry, early, mid, and late lactation (no lactation, <120 days, 120-240 days, and >240 days, respect- ively). Three different datasets were prepared, where rumin- ation, milk, and activity were averaged as means of 1, 3, and 5 days before the response variable record. On each dataset, two models were fitted: logistic regression and random forest. All the analyses were performed in R using the caret package. Data were divided into a training and a testing dataset (pro- portion 80/20). Training data was used to train two different algorithms which were used to predict the class variable. The two selected algorithms were: 1. a logistic regression 2. a ran- dom forest. For all the datasets, logistic regression was not able to predict diseased individuals, assigning all to the 'healthy' class. Random forest performed better, although with a high-class error. The 5-day window had the lowest OOB error rate (0.24%) and the lowest class error (0.71). Further tuning of the selected models will be necessary to build a valuable tool to predict locomotion system issues.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
