Introduction
Today, pig farmers typically asses the health of their pigs while walking through their herd as part of the daily routine. In modern pig production, however, a herd will consists of thousands of pigs in a few hundred pens. If a farmer were to spend just two minutes observing each pen to assess the overall health of the pigs, this alone would easily take 4-5 hours. As real farmers won't take this time, problems can easily be overlooked, leading to needless welfare problems and costs. Implementing existing sensor technology enables automatic monitoring 24/7, and detection algorithms can then identify specific pens in need of extra attention. Here we evaluate the value of monitoring live weight, feed usage, humidity, drinking behavior and pen temperature in relation to early warnings of diarrhea and pen fouling in slaughter pigs.

Materials and methods
We used data collected in 16 pens (8 double-pens) between November 2013 and December 2014 at a commercial Danish farm. During this time, three new batches were inserted. We monitored the mean live weight of the pigs per pen (weekly, only in 4 pens), feed usage per double-pen (daily), humidity per section (daily), temperature at two positions per pen (hourly), water flow per double-pen (liters/hour/pig) and drinking frequency per pen (activations/hour/pig). Staff registrations of diarrhea and pen fouling were the events of interest. The data were divided into a learning set (15 events) and a test set (18 events). The data were modeled on pen level with a multivariate dynamic linear model, which was built using the learning set. Alarms were raised if the unified forecast errors were above a control limit for a sufficient number of consecutive hours (0-25) during a 24 h day. An alarm up to 3 days before or 1 day after an event observation was considered a true positive. If no alarm was raised within this window, it was considered a false negative. True negatives and false positives were counted per day. The predictive values of the various variables were estimated based on how much their omission affected the performance, measured as the area under the receiver operating characteristics curve (AUC).

Results
We achieved an AUC of 0.83 when all variables were included. Omitting live weight and humidity had no effect. Omitting drinking behavior, temperature, and feed usage reduced the AUC to 0.70, 0.81, and 0.82 respectively.