THE GENETIC PARAMETERS OF FEEDING BEHAVIOUR AND PRODUCTION TRAITS, SOCIAL INTERACTION AND DISEASE CONTROL IN FINNISH PIGS

A major proportion of the costs of pork production is related to feed; thus, improving feed efficiency is one of the most important breeding goals in pig breeding programs. Feed efficiency measurements are based on the performance of pigs in a controlled test station environment with an automatic feeding system. During the test period, a large amount of feeding-related information, including time entering the feeder, time leaving the feeder and amount of feed consumed, are gathered. These data can be used to form different feeding behaviour traits (FBT) such as daily feed intake (DFI) or feed intake per visit (FPV). During the test period, pigs are group housed. Therefore, social behaviour between pen mates may affect feeding behaviour, production traits (PT) and welfare of other pen mates. Furthermore, unusual or radical changes in pigs’ behaviour can be a sign of diseases or discomfort and can be used for the detection of diseases for timely intervention. The thesis has three objectives. The first objective was to estimate the heritability of FBT and their correlation with PT in the Finnish Yorkshire pig population (I). The second objective was to estimate genetic parameters of feeding behaviour and PT in Finnish pig breeds using a model with a social genetic effect (SGE; study II). The third objective was to develop machine learning methods for detection of pigs that may need medical treatment or extra management based on changes in their feeding behaviour patterns (III). The data consisted of purebred Finnish Yorkshire and Finnish Landrace pigs and their F1-crosses. Pigs were raised in a controlled test station environment located in Längelmäki, Finland (Figen Oy). Pigs arrived at the test station at an average age of 89 ± 10 days (mean ± standard deviation) and an average weight of 34.7 ± 6.4 kg. The slaughter age was 186 ± 10 days, and the slaughter weight was 121.2 ± 12.9 kg. The feeding, production and sickness data were collected from 10,275 pigs that had entered the test station between 2011 and 2016 (October). From obtained feeding observation, daily values were calculated as the number of visits per day (NVD), time spent in feeding per day (TPD), DFI, time spent feeding per visit (TPV), FPV and feed intake rate (FR). The final records of FBT were calculated as averages of the daily records separately for the five testing time periods: 0 to 20 days (period 1), 21 to 40 days (period 2), 41 to 60 days (period 3), 61 to 80 days (period 4) and 81 to 93 days (period 5). The PT were average daily gain (ADG), feed conversion rate (FCR) and backfat thickness (BF) as well as residual feed intake (RFI), which was calculated as the difference between the observed and predicted DFIs. For the last objective, daily observations of the health status of pigs (794,509 daily observations) were used. The data included 13,018 observations of symptoms (cough, limp, loss of appetite, skin damage, bitten tail). Pigs with symptoms were classified as “sick” while the pigs with no symptoms were classified as “healthy”. For the first objective (I), a univariate animal model was applied for the estimation of heritability of FBT and PT and a bi-variate model for the estimation of genetic correlations between the traits. For the second objective, single-trait social genetic models were applied with two different group size parametrisations: a fixed group size with a random sampling of eight pigs for each pen and a variable group size model where the number of pen mates varied. The restricted maximum likelihood (REML) method and the DMU software were used for the variance component estimations (I and II). For the last objective, the machine learning algorithm (Xgboost) was designed with different window lengths (one-day, three-day and seven-day window lengths) using features calculated from the feeding data to predict the individual health status (sick or healthy). Heritability estimates for FBT and PT were moderate (around 0.3). Only the DFI had strong genetic correlations with PT (e.g., over 0.8 with RFI). Interestingly, the FR and FPV were associated with BF in early testing time (I). The SGE was significant for FBT and the FCR but not for the ADG and BF (II). For those traits, the total heritable variation was considerably higher than the pure classical heritability. The fixed and variable group size models produced very similar estimates (II). Using the Xgboost machine learning method, the best performance for predicting pigs’ daily health status was obtained with a seven-day window length (80% area under the Receiver Operator Characteristics (ROC) curve) [AUC], 7% F1-score, 67% sensitivity and 73% specificity). However, the precision was very low (0.04), possibly due to an imbalanced dataset (III). In conclusion, results indicate that the FBT are moderately inherited but their correlation with PT are weak; thus, FBT are not important in breeding programs. The social genetic component was important in FBT but also in the FCR; thus, applying SGEs in breeding value estimation would increase the accuracy of selection for the FCR. Predictions based on changes in feeding behaviour can help in the detection of sick pigs, leading to better pig welfare and increased sustainability of pork production