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

Overview

• This study developed an advanced artificial intelligence framework that integrates deep imputation techniques, synthetic data balancing, and explainable modeling to accurately predict survival outcomes in horses suffering from colic.
• The proposed AI models demonstrated superior performance compared to traditional methods and revealed key clinical factors important for survival prediction, improving decision-support tools in equine veterinary medicine.

Background and Importance

• Artificial intelligence (AI) is increasingly used to develop clinical decision-support systems in veterinary medicine, although its potential in equine medicine remains underexploited despite the economic and clinical significance of horse health.
• Colic, a common and potentially life-threatening condition in horses, requires timely intervention. Accurate prediction of survival outcomes can guide clinical decisions and improve prognosis.

Research Methods

• Goal: To predict survival outcomes for horses with colic by integrating advanced AI models with data preprocessing and interpretability techniques.
• Machine Learning Models: Combined traditional algorithms and deep-learning architectures including:
  • XGBoost, Light Gradient Boosting Machine (LightGBM), Categorical Boosting (CatBoost) – traditional machine-learning algorithms.
  • TabNet, Feature Tokenizer Transformer (FT_Transformer), Neural Oblivious Decision Ensemble (NODE) – advanced deep-learning models specialized for tabular data.
• Handling Missing Data: Missing clinical information was addressed with deep-learning imputation techniques:
  • Generative Adversarial Imputation Networks (GAIN) with both OneHot and embedding-based input encoding.
  • Missing Data Imputation via Denoising Autoencoder (MIDAS).
• Addressing Class Imbalance: Since survival data may be imbalanced, synthetic data generation methods were applied:
  • Conditional Tabular Generative Adversarial Network (CTGAN).
  • Tabular Variational Autoencoder (TVAE).
• Model Interpretability: SHapley Additive exPlanations (SHAP) framework was used as an Explainable AI (XAI) tool to:
  • Identify important clinical features driving the survival predictions.
  • Enhance trustworthiness and clinical interpretability of AI results.

Key Results

• The best performing pipeline combined:
  • TVAE for synthetic data balancing.
  • GAIN with OneHot encoding for data imputation.
  • LightGBM as the predictive model.
• This combination achieved an area under the curve (AUC) of 0.928, indicating very high classification accuracy in survival prediction.
• The model outperformed both traditional statistical approaches and baseline machine-learning algorithms, verifying the benefit of deep learning-based imputation and synthetic data generation in the pipeline.
• SHAP analysis spotlighted key clinical variables influencing model decisions:
  • Total protein levels.
  • Abdominal appearance.
  • Mucous membrane condition.
  • Packed cell volume.
  • Temperature of the extremities.

Conclusions and Implications

• Ensuring high data quality through deep imputation and addressing class imbalance with synthetic data generation are critical for developing robust AI models in veterinary medicine.
• Integration of explainable AI methods such as SHAP enables clinicians to understand the rationale behind predictions, which is vital for clinical acceptance and decision-making.
• The study’s proposed AI framework offers a novel, accurate, and interpretable approach for survival prognosis in equine colic, with potential application across other clinical scenarios in veterinary care.
• By demonstrating improved predictive performance and transparency, this work paves the way for wider adoption of AI systems in clinical veterinary practice, potentially leading to better health outcomes for horses.