Monitoring mortality events in floor-raised broilers using machine learning algorithms trained with feeding behavior time-series data

•Continuous observation of individual health in poultry production is laborious.•Abnormal feeding behavior could indicate the onset of diseases in broilers.•Electronic feeders allow the monitoring of feeding behavior in real-time.•Disease-related bird removal events were predicted based on feeding behavior.•Machine learning methods achieved promising performance one day in advance. In this study, we explored the integration of machine learning (ML) techniques with feeding behavior (FB) time series data to predict mortality events (animals culled or found dead) in floor-raised broilers. Our dataset included 2,667,617 daily observations for eight FB traits from 95,711 birds across 146 feeding trials. After data cleaning, the class distribution was 93.7 % healthy birds and 6.3 % withdrawn birds (culled or found dead), coded as 0 and 1 respectively. Mortality predictions were made one or three days before the observed events. Time series data for different FB traits were utilized to extract 22 time series features per trait, creating a structured feature dataset (days in the feeding trial + 128 time series features). We compared different ML algorithms: gradient boosting machine (GBM), multilayer perceptron neural network (MLP), logistic regression (LR), random forest (RF), and support vector machine (SVM). Due to the imbalanced nature of the data, we evaluated two sampling strategies: a random under-sampling technique (RUS) and a combined strategy (RUS + SMOTE). Models were assessed using 20-fold cross-validation and an independent test set. Statistical tests indicated consistent differences in most FB traits between control and withdrawn birds at least 7 days before the event. Features derived from traits like daily feed intake, number of visited feeders, visiting activity interval, and number of meals presented high predictive importance for mortality monitoring in broilers. In the cross-validation, classifiers achieved an average (standard deviation) of up to 0.87 (0.02) for the area under the ROC curve (AUC) and 0.55 (0.03) for the area under the precision-recall curve (AUPRC). This demonstrated a significant increase in classification performance compared to a no-skill classifier. However, performance dropped notably when extending the prediction window from one to three days in advance. The performance observed in the independent set was similar to that observed during cross-validation, indicating the robustness of our approach. The RUS + SMOTE strategy