8 Prediction of Culling and Mortality Risks in Group-Housed Broilers Using Machine Learning Methods Trained with Time-Series Data of Feeding Behavior Traits. (21st September 2022)

• Record Type: Journal Article
• Authors: Alves, Anderson A Carvalho; Fernandes, Arthur F A; Lopes, Fernando B; Breen, Vivian; Hawken, Rachel; Rosa, Guilherme J J M
• Is Part Of: Journal of animal science. Volume 100(2022)Supplement 3
• Page Start: 2
• Page End: 2
• Publication Date: 2022-09-21
• Subjects: animal welfare -- classification algorithms -- radio-frequency identification; Livestock -- Periodicals; Livestock; Electronic journals; Periodicals; 636.005
• DOI: 10.1093/jas/skac247.002
• Languages: English
• ISSNs: 0021-8812

Abstract

Abstract: In this study, we investigated the performance of different machine learning (ML) methods for predicting withdrawing events (culled or dead animals) according to feeding behaviors (FB) time series. The raw data comprised a total of 1, 492, 482 daily observations for six FB traits from 55, 400 birds allocated into 88 trials. After data editing, the overall class distribution was 17:1 (44, 352 healthy animals and 2, 689 culled or dead). The event classification (0 or 1) was performed one day in advance, by treating healthy birds randomly as control groups for each day. The FB daily trends were used for extracting 21 time-series features per trait, generating a structured feature dataset (day before the event + 126 time-series features). The trained ML algorithms were the gradient boosting machine (GBM), multilayer perceptron neural network (MLP), naïve Bayes (NB), random forest (RF), and support vector machine (SVM). The models were compared based on the area under the ROC (AUC) and precision-recall (AUPRC) curves, computed with 20-fold cross-validation. The performance metrics ranged from 0.76 to 0.85 for AUC and from 0.32 to 0.48 for AUPRC, so that models achieved considerable superior performance than that expected for a random classifier (0.50 for AUC and 0.06 for AUCPR). The better classifier in terms of AUC was the RF (0.85±0.02), although not statistically different from the average performance obtained with the GBM and SVM (0.84±0.02 in both cases). The MLP achieved the greatest AUPRC (0.48±0.05), while the NB performed poorly considering this criterion (0.32±0.02), indicating that a low precision-recall gain is expected for this model. Broiler barns presenting a critical number of animals with high-risk scores could be indicative of a disease outbreak or management failures. Hence, the proposed approach offers a potential tool for real-time monitoring of health status and welfare in broilers.