The drinking behavior of healthy pigs is known to follow predictable diurnal patterns, and these patterns are further known to change in relation to undesired events such as diarrhea. We therefore expect that automatic monitoring of slaughter pig drinking behavior, combined with machine learning, can provide early and automatic detection of diarrhea. To determine the best approach to achieve this goal, we compared 36 different strategies for combining a multivariate dynamic linear model (DLM) with an artificial neural network (ANN).

We used data collected in 16 pens between November 2013 and December 2014 at a commercial Danish pig farm. The pen level water flow (liters/hour/pig) and drinking bouts frequency (bouts/hour/pig) were monitored. Staff registrations of diarrhea were the events of interest.

Mean water flow and drinking bouts frequency were each modeled using three harmonic waves in a multivariate DLM. The DLM was optimized using the pen-groups for which no events were observed (n=26). The forecast errors produced by the DLM were normalized by the forecast variance and used as inputs for the ANN. In addition, the forecast errors were categorized based on the direction (positive or negative) and simga-1, sigma-2, or sigma-3 cutoff thresholds. Furthermore, observations from between 0 and 48 hours before the day of the observation, with steps of 6 hours, were included in the ANN training window. Thus between 87 and 277 diarrhea-associated observations were included. The diarrhea-associated observations were paired with an equal number of observations from healthy groups, based on the observation date and the age of the pigs. The complete set of diarrhea positive and negative observations was divided into a training set (80 %) and a test set (20 %).

The ANN's consisted of three layers: an input layer corresponding to the number of forecast error categories, a hidden layer with 50 nodes, and an output layer with one node. The various ANN's were applied to all observations in the test set. The observation-level performance of the ANN predictions was evaluated by the error rate, the specificity (SP), and the sensitivity (SE).

The best performance was seen when using a training window with a total of 42 hours for the numerical forecast errors, which produced an error rate=0.16, a specificity=0.88, and a sensitivity=0.80. For the other tested strategies, the ranges of error rates and the corresponding specificities and sensitivities were 0.55-0.28, 0.43-0.71, and 0.50-0.74, respectively.