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

Prioritizing cows for closer observation around calving is important for livestock management, but reliable same-day calving detection from behavior alone remains challenging. We examine whether daily 06:00–06:00 behavior summaries derived from ear-tag accelerometers can support strictly causal, calving-oriented triage in grazing cattle. Because birth annotations provide only daily-resolution supervision, we frame the task as weakly supervised day-level risk assessment rather than precise event timing. Therefore, we train a compact weakly supervised multilayer-perceptron daily-risk model using an exactly-one likelihood over a two-day candidate window comprising the day before birth and the recorded day of birth. The model uses current behavior summaries, deviations from cow-specific causal behavior baselines for the current and previous retained days, a history-gap variable, and missingness indicators. We evaluate the model using cow-grouped nested cross-validation on data from 134 calving cows and 4,891 cow-days. At the primary base operating point, pooled out-of-fold predictions show a measurable but modest signal, with F1 score 0.277, MCC 0.244, AUPRC 0.231, and candidate-window detection 47.0%. Comparator methods recover similar signals with different detection–burden trade-offs: a supervised positive-bag random forest improves F1 score and candidate-window detection, but increases non-candidate false-alert burden. A higher-threshold Alert state reduces the retained non-candidate false-alert rate to 0.0121, but lowers recall to 0.194. Applying the frozen deployment model to an independent non-calving cohort yields low daily Watch/Alert rates but non-negligible cow-level exposure. Overall, daily behavior summaries are better suited to lightweight herd-level risk stratification and human-in-the-loop monitoring than to reliable standalone calving detection.