Predicting and explaining high dead-on-arrival outcomes in meat-type ducks using deep learning: A path to improved welfare management

Abstract

Dead-on-arrival (DOA) rates are a critical welfare and economic concern in poultry production, reflecting the cumulative impact of handling, transport, and lairage conditions on bird mortality. Compared to broilers and layers, meat-type ducks have received less attention in DOA research, despite their distinct physiological responses to preslaughter stressors and increasing relevance in commercial poultry production. Although machine learning models have been widely applied for DOA prediction, their limited transparency can hinder practical application in real-world settings. This study analyzed 8220 truckload entries of meat-type ducks recorded between 2022 and 2023, with the objective of developing an explainable deep learning model to predict high DOA outcomes using preslaughter management and environmental data. Deep learning models, owing to their complex architecture, offer superior predictive capacity and can capture nonlinear interactions in high-dimensional datasets. To enhance model interpretability and support practical application, SHapley Additive exPlanations (SHAP) was applied to identify the most influential predictors of DOA classification. The final model demonstrated strong classification performance, with an accuracy of 80.29 %, precision of 79.25 %, recall of 80.29 %, F1-score of 79.66 %, and an AUC-ROC of 76.03 %. Key predictors of high DOA included duck head count, lairage temperature, duck age, and transport duration. Notably, a higher number of ducks per truckload was strongly associated with elevated DOA risk (i.e., truckloads classified in the high DOA group), along with lairage temperatures and duck ages below the respective medians. Additionally, shorter transport durations were linked to increased mortality, highlighting the complex interplay of preslaughter stressors. By leveraging SHAP analysis, this study provided both global and local interpretability, ensuring that model outputs were not only accurate but also explainable. These findings support precision-driven preslaughter interventions, enabling industry stakeholders to optimize handling, transport, and lairage practices to reduce mortality rates and enhance duck welfare.

Keywords: Dead on arrival; Deep learning; Duck welfare; Meat-type duck; SHAP analysis.

Fig. 1

Deep learning workflow for DOA prediction. The process includes data preprocessing, data splitting, and random oversampling for class balancing. The model undergoes 10-fold cross-validation, development, and hyperparameter optimization before final evaluation on the test set. SHAP analysis is then applied to interpret key predictors of high DOA cases.

Fig. 2

Distribution of key variables in the dataset.

Fig. 3

Correlation matrix of predictor variables.

Fig. 4

SHAP summary plot of variable importance. The SHAP summary plot illustrates the contribution of each predictor variable to the model’s classification of high DOA cases. The x-axis represents the SHAP value, indicating the impact of a given variable on the model’s output, while the y-axis lists the predictor variables in order of importance. Each dot represents a single prediction, with color gradients (red to blue) denoting high to low variable values. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.).

Fig. 5

SHAP dependence plots illustrating the relationship between duck head count and its impact on DOA classification. (A) Interaction with duck age, where red represents older ducks and blue represents younger ducks. (B) Interaction with transport duration, where red indicates longer transport times and blue indicates shorter transport times. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.).

Fig. 6

SHAP dependence plots illustrating the effects of (A) lairage temperature and (B) transport duration on DOA classification, with interactions by winter season.

Fig. 7

SHAP waterfall plot illustrating local interpretability for an individual truckload.

Fig. 8

SHAP decision plot illustrating the contribution of each variable to the model’s high DOA predictions across multiple truckloads. Each line represents an individual case, showing how variables cumulatively influence the final model output.