Leveraging predictive modeling and explainable AI to understand veterinary safety profiles and health outcomes using OpenFDA data

Abstract

The safe use of pharmaceuticals in food-producing animals is essential for protecting animal welfare and human food safety. Adverse events (AEs) in veterinary medicine may reflect unexpected pharmacokinetic or toxicokinetic effects that cause violative residues in the food chain. This study proposes a predictive modeling framework to classify animal health outcomes as either “Death” or “Recovered”, using the U.S. Food and Drug Administration’s OpenFDA Center for Veterinary Medicine database, containing approximately 1.25 million reports from 1987 to 2024 Q3. The preprocessing pipeline linked relational tables and mapped AEs and drugs to standardized ontologies, including the Veterinary Dictionary for Drug Regulatory Activities (VeDDRA) and the World Health Organization’s Anatomical Therapeutic Chemical (ATC) classification system for veterinary use. It further standardized units, handled missing values, and simplified high-cardinality categorical variables. Multiple supervised learning models were evaluated, including Random Forest, CatBoost, XGBoost, and the transformer-based ExcelFormer. Class imbalance was addressed using random undersampling and the Synthetic Minority Oversampling Technique (SMOTE), with a particular focus on improving recall for fatal outcomes. A semi-supervised learning approach with Average Uncertainty Margin (AUM)-based pseudo-labeling was applied to utilize reports with uncertain outcomes, incorporating high-confidence predictions into retraining. XGBoost with undersampling achieved the best overall performance (accuracy 91.60%, F1-score 92.24%, death recall 91.56%), while adding AUM-filtered pseudo-labeled data improved death recall to 93.34% with minimal loss in overall accuracy. Model explainability was achieved through Gain-based feature importance and SHapley Additive exPlanations (SHAP) analysis identified biologically plausible predictors, including specific organ system disorders (e.g., bronchial and lung disorders, heart disorders) that were strongly linked to Death outcomes, as well as animal demographic information and drug characteristics, which contributed to distinguishing between “Death” and “Recovered” outcomes. This work demonstrates that combining rigorous data engineering, advanced machine learning, and explainable AI enables accurate, interpretable prediction of veterinary safety profiles and health outcomes, supporting earlier detection of high-risk profiles and informing evidence-based regulatory and clinical decision-making.