Presentation is loading. Please wait.

Presentation is loading. Please wait.

Classifiers based on 3D-head acceleration data

Similar presentations


Presentation on theme: "Classifiers based on 3D-head acceleration data"— Presentation transcript:

1 Classifiers based on 3D-head acceleration data
Kiel University Kiel University Faculty of Agricultural and Nutritional Science Faculty of Agricultural and Nutritional Science Institute of Animal Breeding and Husbandry Institute of Animal Breeding and Husbandry Classifiers based on 3D-head acceleration data used for lameness detection in dairy cows Introduction Lameness is a great health & welfare problem Changed behaviour & movement patterns Decrease in activity level Compensatory head movement Material & Methods Ear tag with a 3D-Acceleration sensor with data frequency of 1 Hz & 10 Hz Locomotion scoring 3-times a week 1 (normal) to 5 (severely lame) (SMARTBOW GmbH, Weibern, Austria) Aim: Determination of classifiers based on 3D-head acceleration data to detect lameness in dairy cows Classifier calculation for each cow Lame animals  Hz: 6 lameness of 5 cows 10 Hz: 17 lameness of 10 cows Day of detected lameness (locomotion score > 2) + 21 days prior Healthy animals Equal number of cows Matching the lame cows in lactation number & used time period Learning dataset: Included first 10 days of the dataset & Thresholds: Ranges of observed features were extended Divided by 250  Threshold-steps (0-250) Raw dataset Learning dataset & Thresholds Each day 100 sections  10 pre-processing methods median, mean, standard deviation, percentile 25 & 75, skewness, kurtosis, p-variation (p = 1, 2, 3)  100 values / day & cow Test dataset: Included last 12 days, disease blocks were defined as the day of detected lameness + 3 days prior & Classifiers: Application and evaluation of classifiers defined by a pair of pre-processing method & feature - Alerts were given if the threshold was exceeded - True positive/negative, false negative/positive events were defined by comparing alerts to disease blocks Each threshold (0 to 250) corresponds to one point on the ROC (receiver operator characteristics) curve Pre-processing methods Test dataset & Classifiers 100 sections  daily features 17 feature calculations i.e. sum, minimum, maximum, range, skewness  1 value / day & cow Features Results AUC (area under ROC curve), Sensitivity & Specificity for the three best classifiers ROC curve for the best classifiers Pre-processing Feature AUC Sensitivity Specificity 1 Hz Percentile 25 Variance 0.90 0.82 Standard deviation 0.83 Range 0.89 0.95 0.79 10 Hz Kurtosis Median 0.80 0.71 0.72 Percentile 75 Skewness 0.76 Interquartile range 0.73 1 Hz (AUC=0.90) Percentile 25 – Variance 10 Hz (AUC=0.80) Kurtosis – Median Random choice (AUC=0.50) Discussion The 10 Hz dataset had lower AUC values than the 1 Hz dataset The 1 Hz dataset included less animals, and all lameness were treated according to diagnose  clear lameness The 10 Hz dataset included cows with multiple detected lameness without further treatment For further analysis these cows should be excluded or treatment data should be used Conclusion Classifiers based on percentile 25, percentile 75, kurtosis & skewness had the best results Previous studies showed, that these methods indicated changes in activity level In lame cows acceleration data accumulated around the mean (e.g. higher kurtosis) Distribution of acceleration data for a lame and a non lame cow Yvonne Linka , Jennifer Salaua, Susanne Karstenb, Joachim Krietera aInstitute of Animal Breeding and Husbandry, Christian-Albrechts-University, Kiel, Germany bTiDa Tier und Daten GmbH, Bosser Str.4c, Westensee, Germany Funded by: Stiftung Schleswig-Holsteinische Landschaft


Download ppt "Classifiers based on 3D-head acceleration data"

Similar presentations


Ads by Google