Analysis and prediction of backfat thickness in gestating sows using machine learning algorithms

Abstract

Sow backfat (BF) thickness is a key indicator for predicting the nutrient requirements and influencing on reproductive performance of gestating sows. The purpose of this study was to determine feature importance for healthy piglets, define the optimal BF at farrowing and change trend of BF during gestation, as well as to establish the prediction models of BF in gestating sows using 10 machine learning (ML) models. A database with 64,298 observations including 3 categorical and 18 numerical features was used for data analysis and modeling. Compared to other features, BF at farrowing was the most important feature for healthy piglets. The optimal BF at farrowing of parity 1, 2, and ≥3 was 18 mm, 16 mm, and 16 mm, respectively, and the early to middle stage of gestation was the best period for body condition restoration. The eXtreme gradient boosting regression (XGBR) and gradient boosting regression (GBR) exhibited best prediction performance with lowest RMSE (30d, 60d, 90d and farrowing of gestation were 1.17, 1.09, 1.01 and 0.81 mm, respectively) and MAPE (30d, 60d, 90d and farrowing of gestation were 5.57 %, 4.93 %, 4.44 % and 3.54 %, respectively), showing best accuracy and stability among the 10 ML models. The analysis and prediction of BF during gestation based on ML methods provide technical support for accurately predicting nutrient requirements and formulating precise feeding strategy of gestating sows.