Comparação de algoritmos e aplicação em dados de produtividade e condenações em frangos de corte em função dos indicadores bioclimáticos

Abstract

Although the poultry industry plays a crucial role in food security, meeting the growing demand for healthy, flavorful, and high-quality meat, it still faces a significant challenge: partial and total carcass condemnations. These condemnations result in considerable economic losses, reduce market confidence, and raise concerns about animal welfare and product quality. In light of this, this study investigates the application of machine learning, considering environmental, productive, and carcass variables, to build and validate a predictive model for carcass condemnations in broiler chickens. Data from post-mortem inspections conducted in the southwestern region of Paraná, Brazil, were used, collected by an integrator company operating in the area. The information includes animal performance indicators and bioclimatic indices, aggregated with the Thermal Environmental Productivity Index for Broiler Chickens (IAPfc), developed by Medeiros (2005). After thorough data preprocessing, a set of machine learning algorithms was employed to train and test predictive models for carcass condemnations. The climatic environment was categorized based on the IAPfc, ranging from comfortable (21 to 24 points) to dangerous (>35 points). This study applied the IAPfc and found that 98.1% of the birds were in comfortable thermal conditions, indicating that the current environment was suitable for bird welfare. Artificial neural networks (ANNs) showed superior performance, with an AUC (Area Under the ROC Curve) of 99.1%, demonstrating their effectiveness in capturing complex data and classifying the IAPfc. For carcass condemnation rates based on IAPfc, the Random Forest model excelled, correctly classifying 98.49% of instances with a Kappa coefficient of 0.9258, indicating high accuracy in categorizing carcasses across different thermal comfort levels. The relevance of the IAPfc and carcass condemnation indices in this study lies in their ability to predict and improve broiler welfare conditions, directly influencing product quality and production efficiency. The high efficiency of the ANN and Random Forest models underscores the importance of implementing advanced machine learning techniques. Furthermore, genetic and environmental factors were identified as influential in carcass condemnation rates. The integration of these factors can help mitigate issues such as ascites syndrome and cellulitis. Thus, this study demonstrates the feasibility and potential of using machine learning as an effective tool for predicting carcass condemnations in broiler chickens, which can be applied to prevent such losses. The application of this model can assist producers in making more precise decisions, optimizing production, and ensuring the delivery of highquality meat to the market to satisfy consumers.