Technical note: Random forests prediction of daily eating time of dairy cows from 3-dimensional accelerometer and radiofrequency identification

摘要

Feed intake and time spent eating at the feed bunkare important predictors of dairy cows’ productivityand animal welfare, and deviations from normal eatingbehavior may indicate subclinical or clinical disease.In the current study, we developed a random forestsalgorithm to predict dairy cows’ daily eating time (ofa total mixed ration from a common feed bunk) usingdata from a 3-dimensional accelerometer and a radiofrequencyidentification (RFID) prototype device (logger)mounted on a neck collar. Models were trained oncontinuous focal animal observations from a total of 24video recordings of 18 dairy cows at the Danish CattleResearch Centre (Foulum, Tjele, Denmark). Each sessionlasted from 21 to 48 h. The models included boththe present time signal and observations several secondsback in time (lag window). These time-lagged signalswere included with the purpose of capturing changesover time. Because of the high costs of installing anRFID antenna in the feed bunk, we also investigateda model based solely on 3-dimensional accelerometerdata. Furthermore, to address the trade-off betweenprediction accuracy and reduced model complexity andits implications for battery longevity, we investigatedthe importance of including observations back in timeusing lag window sizes between 8 and 128 s. Performancewas evaluated by internal leave-one-cow-outcross-validation. The results indicated that we obtainedaccurate predictions of daily eating time. For the mostcomplex model (a lag window size of 128 s), the medianof the balanced accuracy was 0.95 (interquartile interval:0.93 to 0.96), and the median daily eating timedeviation was 7 min 37 s (interquartile interval: −6 to15 min). The median of the average daily eating timeduring sessions was 3 h 41 min with an interquartileinterval of 2 h 56 min to 4 h 16 min. Exclusion ofRFID data resulted in a considerable decrease in predictionaccuracy, mainly due to a decreased sensitivityof locating the cow at the feed bunk (median balancedaccuracy of 0.87 at a lag window size of 128 s). Incontrast, prediction accuracy only slightly decreasedwith decreasing lag window size (median balanced accuracyof 0.94 at a lag window size of 8 s). We suggesta lag window size of 64 s for further development of theprototype logger. The methodology presented in thispaper may be relevant for future automatic recordingsof eating behavior in commercial dairy herds.