Daily Calving-Oriented Triage in Grazing Cattle via Ear-Tag Accelerometers: A Weakly Supervised Approach

We investigated whether daily behavior summaries derived from a behavior classifier applied to ear-tag accelerometer data could support calving-oriented monitoring under field conditions. Data came from a seedstock operation in subtropical Australia with grazing and feedlot components. Because exact birth times were unavailable, we used weak supervision over two candidate days: the farm-recorded calf-observation date and the previous day. A neural network model using an exactly-one formulation over this candidate window generated an uncalibrated compatibility score from daily behavior durations and cow-specific baseline-deviation features only after each 06:00–06:00 interval was complete. Cow-grouped nested cross-validation used 134 calving cows and 4891 cow-days. At the selected operating point, the model achieved 47.0% candidate-window coverage, an F1 score of 0.277, an AUPRC of 0.231 with a candidate-day prevalence reference of 0.055, and a 2.94% pooled false-alert rate on non-candidate days. Random forest and an at-least-one neural network formulation achieved higher candidate-window coverage at their selected thresholds but greater false-alert burden outside the candidate window. We also characterized out-of-domain output burden in an independent cohort containing no calving events: 8 of 1547 scored cow-days were Alert and 7 of 31 cows received at least one Alert. These findings indicate useful but modest calving-associated information in daily behavior summaries and support prospective evaluation for management triage. However, the present study did not establish intervention lead time. More precise event timing and external calving cohorts are needed to establish advance-warning value and transferability.