Hidden phase-type Markov model for the prediction of onset of farrowing for loose-housed sows

• Algorithm for the prediction of onset of farrowing based on online sensor records.
• Directly calculates the mean and standard deviation of remaining time to farrowing.
• A framework to integrate different sensors; possible to rank the sensor combination.
• Satisfactory performance in the warning system; promising in Precision livestock farming.
• Low complexity of the algorithm means easy implementation on the herd level computer.

High piglet mortality is an issue in the pig production. Evidence indicates that if the time of farrowing can be predicted, the mortality can be reduced through planned supervision or improved climate regulation. The aim of the study was to improve the prediction of onset of farrowing by monitoring pre-parturient behaviour of sows using several sensors and by developing an automated system for the prediction of time to farrowing. The resulting prediction model, named as Hidden Phase-type Markov Model (HPMM), assumes that a sow passes through the behavioural states Before Nest-Building, Nest-Building and Resting before reaching the Farrowing state. Each state was further split into phases, to allow a more realistic distribution of sojourn times. As these phases and states are unobservable, HPMM was used to calculate the probability of a sow being in given phase using the automatic sensor measures. Thus time to farrowing could be predicted at each time point. The prediction algorithm was validated on a sensor data set for about 35 sows, each followed from day 105 (day-105) since mating until the farrowing. Sensors include sow activity measured by video recording as well as by a photo-cell grid, and water consumption. The algorithm was evaluated using heuristic warning strategies e.g. that a warning should be generated when the expected time to farrowing was less than 12 h (inspired by the regulation of floor heating systems). The performance of the sensors was evaluated. Different combinations of sensor types outperformed the use of a single sensor type. Using a combination of water and activity sensors the prediction algorithm gave a coherent warning period prior to farrowing (true warning) in 97% of the cases. The duration from start of the warning period to farrowing had a mean 11.5 (SD = 4.6) h. False warning periods ending before farrowing lasted on average only 0.7 h per sow. The use of HPMM thus allowed a direct prediction of the time to farrowing, handling more than one sensor and a compact representation of historical sensor information.