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

Feed intake and time spent eating at the feed bunk are important predictors of dairy cows' productivity and animal welfare, and deviations from normal eating behavior may indicate subclinical or clinical disease. In the current study, we developed a random forests algorithm to predict dairy cows' daily eating time (of a total mixed ration from a common feed bunk) using data from a 3-dimensional accelerometer and a radiofrequency identification (RFID) prototype device (logger) mounted on a neck collar. Models were trained on continuous focal animal observations from a total of 24 video recordings of 18 dairy cows at the Danish Cattle Research Centre (Foulum, Tjele, Denmark). Each session lasted from 21 to 48 h. The models included both the present time signal and observations several seconds back in time (lag window). These time-lagged signals were included with the purpose of capturing changes over time. Because of the high costs of installing an RFID antenna in the feed bunk, we also investigated a model based solely on 3-dimensional accelerometer data. Furthermore, to address the trade-off between prediction accuracy and reduced model complexity and its implications for battery longevity, we investigated the importance of including observations back in time using lag window sizes between 8 and 128 s. Performance was evaluated by internal leave-one-cow-out cross-validation. The results indicated that we obtained accurate predictions of daily eating time. For the most complex model (a lag window size of 128 s), the median of the balanced accuracy was 0.95 (interquartile interval: 0.93 to 0.96), and the median daily eating time deviation was 7 min 37 s (interquartile interval: -6 to 15 min). The median of the average daily eating time during sessions was 3 h 41 min with an interquartile interval of 2 h 56 min to 4 h 16 min. Exclusion of RFID data resulted in a considerable decrease in prediction accuracy, mainly due to a decreased sensitivity of locating the cow at the feed bunk (median balanced accuracy of 0.87 at a lag window size of 128 s). In contrast, prediction accuracy only slightly decreased with decreasing lag window size (median balanced accuracy of 0.94 at a lag window size of 8 s). We suggest a lag window size of 64 s for further development of the prototype logger. The methodology presented in this paper may be relevant for future automatic recordings of eating behavior in commercial dairy herds.