Development of an AI-Enabled Livestock Biosecurity and Disease Risk Intelligence Framework for U.S. Poultry and Livestock Supply Chain

Abstract

This quantitative study developed and validated an AI-enabled livestock biosecurity and disease risk intelligence framework for integrating epidemiological, biosecurity, production, environmental, geographic, movement, and supply-chain network information across U.S. poultry and livestock systems. A longitudinal observational design was applied to 18,742 premises-period observations representing 3,864 unique premises and supply-chain facilities across 32 U.S. states. The overall disease positivity rate was 14.1%, with substantial variation among production systems, including 18.9% for turkey, 17.1% for broiler, 16.8% for swine breeding, 14.7% for layer, 14.4% for swine finishing, 11.6% for dairy, and 8.7% for beef operations. Disease-positive observations had a substantially lower mean biosecurity score than disease-negative observations (57.9 vs. 74.8) and exhibited greater movement intensity, livestock density, proximity to infected premises, and network connectivity. Logistic regression and multiple machine-learning algorithms were evaluated using training, validation, cross-validation, temporal-holdout, geographic-holdout, and independent-test procedures. XGBoost achieved the strongest independent-test performance, with 93.6% accuracy, 91.8% sensitivity, 94.1% specificity, an F1-score of 85.6%, ROC-AUC of 0.968, PR-AUC of 0.902, and balanced accuracy of 92.9%. Its false-negative rate was 8.2%, compared with 23.6% for logistic regression. Multivariable analysis identified proximity within 10 km of an infected premises as the strongest risk factor (adjusted OR = 3.41, 95% CI [3.01, 3.87]), followed by weak biosecurity (OR = 2.96) and high network connectivity (OR = 2.48). Explainable-AI analysis ranked total biosecurity score, proximity to infected premises, node strength, incoming shipments, and degree centrality among the strongest predictors. Validation demonstrated limited performance deterioration, with ROC-AUC decreasing from 0.982 during training to 0.968 during independent testing, while a Brier score of 0.061 and calibration slope of 0.97 indicated strong calibration. The findings established that integrated multidomain AI modeling provided accurate, interpretable, robust, and generalizable quantitative disease-risk classification across heterogeneous U.S. poultry and livestock supply-chain environments.