Interpretable AI for Animal Health: Bayesian Networks with Uncertainty-Aware Decision Making

Timely and trustworthy animal disease detection is critical for animal welfare, food security, and farm profitability, yet existing AI systems struggle with noisy sensors, class imbalance, and limited interpretability. We present U-DeepBN, an Uncertainty-Aware Bayesian Network with Deep Learning, that integrates deep feature learning, uncertainty quantification, and Bayesian reasoning for robust and explainable disease prediction. By modeling feature-level uncertainty and explicit dependencies, U-DeepBN enables intrinsic interpretability and transparent inference under uncertainty and adversarial perturbations, supported by theoretical analysis of consistency and robustness properties. Experiments on three real-world animal disease datasets and a cross-domain dataset show that U-DeepBN improves accuracy and robustness, achieves higher explanation fidelity and stability, and substantially reduces runtime compared to strong baselines. Ablation studies reveal that optimal uncertainty types and feature representations are dataset-dependent, with adaptive uncertainty modeling and hybrid features yielding the most reliable performance. These results position U-DeepBN as a practical, deployable decision-support tool for precision livestock monitoring.