22 Comparison of linear and machine learning models to predict dry matter intake in feedlot cattle.

Abstract Accelerometers can provide continuous behavioral data that may partially correlate with intake. The objective of this research was to determine if measurements collected from ear tag accelerometers (CowManager; Harmelen, Netherlands) could be used to predict dry matter intake (DMI) in feedlot cattle. Data were collected from cattle (initial BW = 832 ± 324 kg) at Hy-Plains Feedyard (Montezuma, KS; n = 9,062 observations) across three feeding trials involving 69, 119, and 114 animals, respectively, with trial durations ranging from 35 to 47 d. Animals were fed a standard finishing beef diet once daily in Vytelle SENSE feeders (Vytelle, LLC; Lenexa, KS). The dataset represented beef × dairy crossbred (65%) and Red Angus cattle (35%), and included bulls (35%), heifers (16%), and steers (49%). Body weights were measured daily using Vytelle SENSE In-Pen-Weighing nodes. Observations ± 3 SD from the mean were removed to mitigate the influence of extreme values. Body weights (BW) were selectively included in specific prediction models to test their significance in predicting intake, as regular weighing is uncommon on most feedlots. Six predictive models were evaluated: (1) mixed-effects regression with daily BW (MIX+), (2) mixed-effects regression without daily BW (MIX-), (3) random forest regression with daily BW (RF+), (4) random forest without daily BW (RF-), (5) linear regression with daily BW (L+), and (6) linear regression without daily BW (L-). Variables included in all models were ear temperature (°C), trial day number, rumination, eating, and high activity (minutes/d), breed, and sex. Initial BW (kg; as a covariate), and daily BW (kg) were used in MIX+, RF+ and LIN+. Data were randomly partitioned into training (80%) and testing (20%) sets to assess model generalizability. Model performance was assessed using root mean square error (RMSE) and R² statistics. The MIX+ model provided the greatest predictive accuracy (RMSE = 3.62, R² = 0.42). The RF+ model closely approached this performance (RMSE = 3.67, R² = 0.40). The MIX- model had slightly lower performance (RMSE = 3.79, R² = 0.36), while the LIN+ model performed moderately (RMSE = 3.81, R² = 0.35). The RF- (RMSE = 4.04, R² = 0.27) and LIN- (RMSE = 4.21, R² = 0.21) models showed notably reduced predictive capability without daily BW. Although MIX+ provided the greatest predictive accuracy, including random animal effects limits its applicability for predicting intake in new animals due to impracticality of obtaining historical data for individual cattle. Consequently, RF- offered greater practical value to predict intake without frequent weighing. Integrating activity-based intake prediction models has the potential to enable real-time intake monitoring on commercial feedlots; however, additional data is needed to validate their accuracy and applicability across diverse cattle populations and management systems.