Event- Level Decision Quality Under Temporal Behavior Annotation Perturbations: Simulation and Sow–Piglet Ethogram Evaluation

Errors in temporal behavior annotations can change decisions about when behavioral events occur. This study examined how six types of annotation errors affect downstream event decisions. A temporal event decision model (TEDM) was evaluated using 100 synthetic reference sequences and manually annotated sow–piglet ethograms. Its results were compared with those of three rule-based methods under controlled perturbations. In repeated five-fold cross-validation on synthetic data, TEDM and a smoothing-based method achieved micro-averaged Event F1 scores of 0.9293 and 0.9300, respectively, without a statistically significant difference (Holm-adjusted p = 0.4570). Category errors and omitted events caused the largest reductions in event detection, whereas small timing shifts mainly affected temporal alignment. In the held-out manual ethogram evaluation, TEDM and the smoothing-based method achieved Event F1 scores of 0.822 and 0.796, respectively. These results showed that the effects of annotation errors depended on the error type and the temporal evidence required by each event criterion. Further evaluation using larger datasets and outputs from trained behavior recognition models is needed.