TLDR: A new AI framework uses machine learning and explainable AI to predict animal health outcomes (Death or Recovered) from veterinary drug adverse events. It analyzes 1.28 million FDA reports, integrating drug physicochemical properties and standardized medical ontologies. Ensemble models achieved 0.95 F1-scores, and SHAP identified key predictors like organ system disorders, animal demographics, and drug chemical features, enhancing early detection of high-risk drug profiles for animal welfare and human food safety.
Ensuring the safe use of pharmaceuticals in animals, especially those destined for human consumption, is a critical aspect of public health and animal welfare. This involves carefully monitoring for adverse events (AEs) and preventing harmful drug residues from entering the food chain. A recent study introduces a groundbreaking predictive modeling framework designed to classify animal health outcomes, specifically distinguishing between “Death” and “Recovered” cases, using a vast dataset from the U.S. FDA’s OpenFDA Center for Veterinary Medicine.
The research, conducted by a team including Hossein Sholehrasa, Xuan Xu, Doina Caragea, Jim E. Riviere, and Majid Jaberi-Douraki, delves into approximately 1.28 million reports spanning from 1987 to the first quarter of 2025. This extensive dataset provides a rich foundation for understanding how veterinary drugs impact animal health. The core of their approach lies in a comprehensive data preprocessing pipeline. This pipeline meticulously links various relational tables and maps adverse events and drugs to standardized ontologies like the Veterinary Dictionary for Drug Regulatory Activities (VeDDRA) and the WHO’s Anatomical Therapeutic Chemical (ATC) system for veterinary use. This standardization is crucial for ensuring consistency and comparability across the diverse reports.
Beyond clinical data, the study also integrated physicochemical properties of drugs. These properties, such as molecular weight and partition coefficients, help capture the intricate relationships between a drug’s chemical characteristics, how it distributes within an animal’s body, and its potential to leave residues. By combining these molecular-level features with demographic and clinical information, the researchers aimed to build a more holistic predictive model.
The team evaluated a range of supervised machine learning models, including popular ensemble methods like Random Forest, CatBoost, and XGBoost, as well as more advanced transformer-based models like ExcelFormer, and even in-context learning with large language models (LLMs) such as Gemma 3-27B and Phi 3-12B. A significant challenge in this type of data is class imbalance, where “Recovered” cases far outnumber “Death” cases. To address this, various techniques like undersampling, oversampling, and SMOTE+ENN were applied, with a particular focus on improving the recall for fatal outcomes, which is paramount in veterinary safety.
The results were compelling. Ensemble methods consistently delivered the strongest performance, with CatBoost, Voting, and Stacking classifiers achieving impressive precision, recall, and F1-scores of 0.95. This indicates their robust ability to accurately predict outcomes. Furthermore, the study explored semi-supervised learning by incorporating Average Uncertainty Margin (AUM)-based pseudo-labeling for uncertain cases. This technique helped to expand the training data with confidently predicted unlabeled cases, leading to further improvements, especially in recall for minority classes in models like ExcelFormer and XGBoost.
A key strength of this framework is its emphasis on explainable AI (XAI), specifically using SHapley Additive exPlanations (SHAP) values. SHAP helps to interpret model predictions by identifying which features are most influential in determining an outcome. The analysis revealed biologically plausible predictors for fatal outcomes, including specific organ system disorders (e.g., bronchial, lung, and heart conditions), animal demographics (like age and weight), and certain drug physicochemical properties. For instance, older companion animals and heavier livestock were more prone to fatal outcomes, while drugs with greater atom stereocenter counts often correlated with increased risk.
The SHAP analysis also provided insights into specific adverse events and active ingredients. Protective features associated with recovery included stomach disorders and epidermal reactions, while severe conditions like bronchial and lung disorders or abnormal pathology reports were strong indicators of death. Certain active ingredients, such as Grapiprant and Maropitant Citrate, were linked to recovery in companion animals, whereas others like Tilmicosin and Tildipirosin were strongly associated with fatal outcomes in livestock.
Also Read:
- Adaptive AI Optimizes Molecular Design in Data-Scarce Environments
- Streamlining Pharmacokinetic Data Extraction with AutoPK’s AI-Powered Framework
This research significantly advances veterinary pharmacovigilance and human food safety by providing an accurate and interpretable framework for predicting veterinary safety outcomes. The framework aligns with the mission of the Food Animal Residue Avoidance Databank (FARAD), supporting the early detection of high-risk profiles and informing crucial regulatory and clinical decisions. The code and hyperparameters for this study will be made available on GitHub, fostering transparency and further research. You can read the full research paper here: Predictive Modeling and Explainable AI for Veterinary Safety Profiles, Residue Assessment, and Health Outcomes Using Real-World Data and Physicochemical Properties.
