Farrowing Prediction Dataset

Activity-based time-to-farrowing: preprocessed training data Summary Per-instance activity and crating features for sows approaching farrowing, derived from two independent sensing modalities: an ear-tag accelerometer and video-derived optical-flow activity, for training and evaluating time-to-farrowing prediction models. Companion data release for the code and analysis described in Mayrhuber et al (citation to be added once published; see this dataset's own listing for an updated reference). Contents farrowing_prediction_dataset.csv: 9478 instances across 180 sow(s) (180 with accelerometer data, 21 with video data, 21 with both) -> 8549 accelerometer instances and 929 video instances in total. Columns Column Type Description sow_id integer Anonymized sow identifier (1..N). A sow with both modalities has the same sow_id in each, so cross-modal pairs stay linkable; it carries no relation to any real trial, tag or animal number. modality string accelerometer or video: which sensor this instance's activity features were computed from. mean_day float Mean activity during a fixed daytime window (10:00-18:00) near the start of this sow's recording - a per-sow baseline, constant across that sow's instances. mean_night float Mean activity during a fixed nighttime window (22:00-04:00) near the start of this sow's recording - a per-sow baseline, constant across that sow's instances. mean_24 float Mean activity over the first 24h of this sow's recording (before nest-building activity is expected) - a per-sow baseline, constant across that sow's instances. mean_48 float Mean activity over the first 48h of this sow's recording - a per-sow baseline, constant across that sow's instances. mean_last_12 float Mean activity over the 12h immediately before this instance's timestamp. Varies instance to instance. mean_last_24 float Mean activity over the 24h immediately before this instance's timestamp. Varies instance to instance. trend_last_12 float Slope of a linear fit of activity against time over the last 12h (activity trend). Varies instance to instance. trend_last_24 float Slope of a linear fit of activity against time over the last 24h (activity trend). Varies instance to instance. currently_crated 0/1 Whether the sow was confined to the farrowing crate at this instance's timestamp. crated_since float Minutes elapsed since crating began, at this instance's timestamp; 0 if not currently crated. remaining_minutes float Target variable: minutes remaining until farrowing onset, measured from this instance's timestamp. How the data was derived Each instance is one hourly snapshot in the 48h window before a sow's known farrowing time, built only from data available up to thattimestamp (no look-ahead). The activity features use the same feature-extraction code as the companion accelerometer-only study, restricted to the subset that needs neither prepartum-examination data nor farm-management variables (e.g. pen type), so the released features depend only on information either sensor can supply on its own. A sow with a short or gappy recording can have a baseline feature (mean_day/mean_night/mean_24/mean_48) missing for some instances; these are forward/back-filled from that sow's own other instances rather than dropped, since dropping would lose an otherwise usable instance over one weak feature. License CC BY 4.0 (Creative Commons Attribution 4.0 International). You may copy, redistribute, and adapt this data for any purpose, including commercially, as long as you give appropriate credit. Full text. Citation If you use this dataset, please cite the companion paper (Mayrhuber et al., citation to be added once published) and, if relevant, the code release it accompanies. Contact Elisabeth Mayrhuber [ORCID]