Inteligência artificial para identificação de mastite bovina em sistemas de produção compost barn

Abstract

The numerous factors that affect subclinical mastitis contamination in cows within controlled environments, such as Compost Barns, make the study of its detection a complex challenge. While there is research in the literature that focuses on subclinical mastitis detection through machine learning using various models and variables, the application of such a technique for subclinical mastitis detection remains relatively unexplored, especially in the context of automated machine learning. Therefore, using automated machine learning models for subclinical mastitis detection is a promising research topic. Inspired by this gap, this study aims to apply advanced automated machine-learning techniques to enhance early subclinical mastitis detection in dairy cattle housed in Compost Barns, focusing on animal welfare and early disease identification. Additionally, it aims to assist producers in reducing the significant financial impacts caused by this disease. Machine learning is crucial in mastitis detection, enabling efficient analysis of subtle patterns in dairy cattle data. This approach facilitates early disease identification, resulting in prompt interventions that preserve animal welfare and reduce costs associated with late treatments. The ability to process large volumes of information contributes to a healthier and more productive livestock industry. To address this issue, the study proposed applying four automated machine learning models: AutoGluon, H2O, TPOT, and a combination. These models were tested on three distinct datasets and demonstrated promising results in all analyses. It is worth highlighting the effectiveness of the H2O approach, which achieved a performance of 0.80% in the Area under the ROC Curve metric using only 12 variables. On the other hand, TPOT also showed great potential due to its genetic programming. The best percentage result, regardless of the number of variables, was achieved with the combination of the three models, reaching 0.81% in the Area under the ROC Curve metric when applied to a dataset composed of 24 variables.