PSXI-16 Inclusion of automated sensor data as a predictor of feed intake increases the variance explained by a random forest model. (30th November 2020)

• Record Type: Journal Article
• Authors: Siberski, Cori J; Mayes, Mary S; Gorden, Patrick J; Copeland, Adam; Healey, Mary; Goetz, Brady M; Beiki, Hamid; Kramer, Luke M; Baumgard, Lance H; Dixon, Philip; Koltes, James E
• Is Part Of: Journal of animal science. Volume 98(2020)Supplement 4
• Page Start: 394
• Page End: 395
• Publication Date: 2020-11-30
• Subjects: Precision livestock technologies -- feed intake -- prediction; Livestock -- Periodicals; Livestock; Electronic journals; Periodicals; 636.005
• DOI: 10.1093/jas/skaa278.694
• Languages: English
• ISSNs: 0021-8812

Abstract

Abstract: Prediction of feed intake from indicators would benefit the dairy industry since on-farm feed intake data are rare. The objective of this study was to examine the ability of sensor data to improve predictions of feed intake. Dry matter intake (DMI), milk yield (MY) and components, metabolic body weight (MBW; body weight 0.75 ), and veterinary health records were collected from two cow groups (n1 =47, n2 =60). Automated sensors (ear tags, rumen bolus, environmental) captured measurements of cow activity, temperature, rumination and rumen pH, and barn temperature and humidity which were used to calculate THI. Random forest (RF) models were trained in R (Caret package) by 10-fold cross validation, with DMI as the response variable. Training data originated from the full study with the exception of the final day, for which DMI was then predicted. Predictive ability was evaluated against a base model excluding automated sensor data to determine changes in accuracy and the percent of variance explained (VAR). The base model included MY and components, MBW, THI, health status and parity. Base model mean square error (MSE) was 9.86, 13.25 and 12.50 kg of DMI and VAR 44.71, 42.9 and 44.85% (n = 92, 56 and 41, respectively). The correlation between actual and predicted final day DMI (CORR) was 0.05, 0.03 and 0.02 (n = 92, 56 and 41, respectively). Adding activity and temperature (first ear tag; n = 92) reduced MSE to 9.70 kg and VAR increased to 45.62% (CORR=0.20). Independently adding bolus activity, rumen temperature and pH (n = 56) to the base model also decreased MSE to 12.53 kg (VAR=46.24% and CORR=0.26). Lastly, adding activity and rumination from the second ear tag (n = 41) to the base model decreased MSE to 12.32 kg (VAR=45.63%, CORR=0.18). Automated sensors appear to explain additional variation in DMI that is not captured in the typical energy sink variables utilized when predicting intake.