A modern AI framework integrating deep imputation, synthetic data balancing, and explainable modeling for survival prediction in horse colic.

Abstract

BACKGROUND: Artificial intelligence (AI) has emerged as one of the most transformative tools for developing clinical decision-support systems in veterinary medicine. Despite its growing use, its full potential remains underutilized in equine medicine, an area of both high economic and clinical importance. Accurate survival prediction in horses with colic is crucial for timely intervention and improved clinical outcomes. METHODS: This study aimed to predict survival outcomes in horse colic cases by developing models that combine traditional machine-learning algorithms (XGBoost, Light Gradient Boosting Machine [LightGBM], and Categorical Boosting [CatBoost]) with advanced deep-learning architectures (TabNet, Feature Tokenizer Transformer [FT_Transformer], and Neural Oblivious Decision Ensemble [NODE]). Missing clinical data were imputed using deep-learning-based approaches-Generative Adversarial Imputation Networks (GAIN-OneHot, GAIN-Emb) and Missing Data Imputation via Denoising Autoencoder (MIDAS). Class imbalance was addressed through Conditional Tabular Generative Adversarial Network (CTGAN) and Tabular Variational Autoencoder (TVAE). Model interpretability was assessed using the SHapley Additive exPlanations (SHAP)-based Explainable Artificial Intelligence (XAI) framework to identify the most influential features contributing to survival prediction. RESULTS: Among the tested combinations, the TVAE-GAIN-OneHot-LightGBM pipeline achieved the highest classification performance, with an area under the curve (AUC) value of 0.928, outperforming conventional statistical and machine-learning baselines. SHAP analysis revealed that total_protein, abdomo_appearance, mucous_membrane, packed_cell_volume, and temp_of_extremities were the most decisive clinical variables influencing the model's predictions. CONCLUSIONS: The findings demonstrate that ensuring data integrity, optimizing model complexity, and integrating XAI-based interpretability substantially enhance the reliability and clinical applicability of AI-driven models in veterinary medicine. The proposed framework provides a pioneering and explainable approach for developing accurate prognostic systems in equine colic, paving the way for broader AI adoption in clinical veterinary practice.