Prediction of fat-free lean percentage in pigs from four genetic lines using in vivo real-time ultrasonography applying different statistical approaches

Fat-free lean percentage (FFL%) in pigs represents an important market and economic quality indicator for producers, calculated by taking several measurements throughout a pig's life, entailing high economic and management costs for modeling using predictive approaches. However, existing approaches are either outdated or not adapted to the specific production conditions in Colombia. Accordingly, the aim of this study is to estimate the FFL% in pigs from four genetic lines through a prediction equation using in vivo real-time ultrasonography and applying different statistical approaches. A total of 412 pigs were evaluated between July 2024 and January 2025 under typical management conditions in Antioquia, Colombia. All animals were monitored biweekly from nine weeks of age until slaughter. Body weights and in vivo ultrasound measurements of fat thickness and muscle depth were registered at the 10th and last ribs using a linear transducer following standardized longitudinal scanning protocols. Stepwise multiple linear, multivariate, and PCA-based multivariate regressions, generalized additive models (GAM), Legendre polynomials, and Cisneros' traditional equation were applied to predict FFL%. Models were evaluated using the coefficient of determination ( R ²), root mean square error (RMSE), and Akaike and Bayesian information criteria. Correlations between predicted and observed values and the coefficient of variation for predicted values were used to assess predictive performance. GAM showed the best predictive performance, indicating moderate accuracy and stability. The proposed equation (model) integrated key biometric and carcass indicators, allowing a reliable FFL% estimation in live pigs. Among the evaluated models, the generalized additive model (GAM) showed the best predictive performance for fat-free lean percentage, achieving a moderate coefficient of determination (R² = 0.59), the lowest RMSE (2.64), and the best AIC. The GAM also exhibited the strongest agreement with slaughterhouse data (r = 0.82) and the lowest coefficient of variation (6.25%), indicating superior stability and robustness. In contrast, the traditional Cisneros equation showed poor agreement with current carcass data, high dispersion, and a negative correlation, highlighting its inadequacy for modern genetic lines. Resampling validation confirmed the robustness of the GAM, with nearly identical RMSE values between the full dataset and jackknife analysis, supporting its applicability for in vivo lean estimation under Colombian production conditions. Further validation across diverse genotypes and production systems is recommended to ensure broader applicability.