Machine-learning and deep-learning approaches have shown promising applications in medical diagnosis, imaging, and laboratory testing; however, the use of these technologies in veterinary medicine remains relatively limited (Cihan et al., 2017). Nevertheless, studies conducted in this field have demonstrated that artificial intelligence models yield encouraging results in clinical diagnosis (Cihan et al., 2021, Xiao et al., 2025), disease classification (Cihan et al., 2024, Guitian et al., 2023), production efficiency prediction (Topuz, 2021), and image analysis (Cihan et al., 2025, Saygılı et al., 2024).

Equine health represents one of the most significant areas for the application of artificial intelligence technologies. Horses possess high economic value, particularly due to their use in racing and breeding (Tinker et al., 1997). Animal models, particularly equine models, have been widely used in medical and surgical research due to their anatomical and physiological relevance, and have been shown to provide valuable insight into clinical decision-making processes in veterinary and translational medicine (Choudhary, 2025, Choudhary and Sarkar, 2025). Acute abdomen (colic) is the most common and serious health condition in horses, which can result in death if not treated promptly. Colic causes substantial economic losses for both animal welfare and owners. Although the majority of cases respond to medical treatment, some require surgical intervention, and the mortality rate is considerably higher in this group. Therefore, early diagnosis, the selection of appropriate treatment, and accurate survival prediction are important in veterinary clinical practice.

In this process, clinicians have long relied on traditional prognostic approaches. These approaches typically consider basic clinical parameters such as heart rate, mucous membrane color, hematocrit, and pain severity; however, the subjective interpretation of these variables and their inconsistent results across different populations limit their predictive power. In addition, pathophysiological factors such as systemic inflammatory response syndrome (SIRS) and coagulation disorders play a critical role in the progression of colic cases. Particularly, the early detection of subclinical coagulopathy findings may improve the accuracy of prognostic predictions. Viscoelastic coagulation tests (VCT) provide a comprehensive method for monitoring such changes; however, their use has remained limited due to high cost, limited accessibility, and analytical complexity (Macleod et al., 2025).

Recently, there has been a growing interest in the application of artificial intelligence in veterinary anatomy, education, and clinical decision support systems. AI-based systems have demonstrated significant potential in enhancing anatomical assessment and veterinary medical training, highlighting the increasing role of AI in the clinical decision-making process (Choudhary et al., 2025, Choudhary et al., 2023).

In this context, machine-learning algorithms demonstrate superiority over traditional regression methods due to their ability to simultaneously process large volumes of clinical variables and model complex biological relationships. Studies aiming to predict survival outcomes in colic cases based on fundamental clinical data have shown that machine-learning-based models achieve higher sensitivity and specificity compared to conventional statistical approaches (Bishop et al., 2022, Macleod et al., 2025).

Cihan, 2024 aimed to predict the need for surgical intervention and the likelihood of survival in horses presenting with acute abdomen (colic) using artificial intelligence based on clinical data. A total of 15 different models were tested on raw, imputed, and balanced datasets, with the highest accuracy achieved by the Random Forest model. The best performance was obtained from the imputed and balanced dataset, yielding 85.83 % accuracy (AUC = 0.906) for surgery prediction and 80.75 % accuracy (AUC = 0.888) for survival prediction. Reducing the number of features further improved model performance. The findings indicate that appropriate handling of missing data and feature selection significantly enhance the performance of AI models and demonstrate the broad application potential of this technology in veterinary medicine.

Fraiwan and Abutarbush, 2020 predicted the need for surgical intervention (surgery) and survival status (survivability) in horses presenting with acute abdomen (colic). These two parameters were evaluated using Decision Trees, Multilayer Perceptron, Bayesian Network, and Naive Bayes algorithms. As a result, the Random Forest algorithm achieved the highest average accuracy, with the models predicting surgical necessity at 76 % accuracy and survival at 85 % accuracy, respectively. The researchers concluded that artificial intelligence technology could serve as a valuable tool in various clinical areas of veterinary medicine and merits further investigation.

Macleod et al., 2025 investigated the combined use of viscoelastic coagulation tests and clinical data to predict survival outcomes in horses with acute abdomen (colic). In the study, prediction models developed using machine-learning algorithms—particularly Random Forest and other classification models—were compared with traditional statistical methods. According to the results, the most successful model, Random Forest, achieved approximately 91 % sensitivity and 83 % specificity in predicting survival. These rates were significantly higher than those obtained with conventional regression models. The findings demonstrate that AI-based approaches hold substantial potential for prognostic assessment in horses experiencing abdominal pain due to colic.

Bishop et al., 2022 aimed to evaluate the performance of survival prediction models in horses undergoing surgical intervention for colic. Using data from a large clinical population in the United States, the study tested the validity of previously proposed statistical models, including univariate and multivariate regression approaches. The results indicated that the overall performance of existing models was limited, with the best-performing model achieving an accuracy of 74 %, sensitivity of 80 %, and specificity of 60 %. The findings suggest that models developed using conventional statistical methods may have low generalizability across different populations and highlight the need for developing new artificial intelligence–based models to improve predictive accuracy.

In conclusion, artificial intelligence technologies offer strong potential for the early diagnosis and prognostic assessment of colic-related acute abdominal cases in horses. The integration of clinical and laboratory data with machine-learning models can enable more accurate prediction of surgical requirements and survival probabilities, thereby substantially supporting veterinarians in their clinical decision-making processes.

This study was structured around four main objectives: (i) to complete the missing values in the dataset using deep-learning-based imputation approaches (GAIN-OneHot, GAIN-Emb, MIDAS) and to compare their performance; (ii) to balance the imbalanced dataset using advanced synthetic data generation techniques (CTGAN, TVAE) and to evaluate their effectiveness; (iii) to identify the most influential variables in the model’s decision-making process through SHAP analysis and to assess survival prediction performance using reduced feature subsets; and (iv) to compare the survival prediction performance of machine-learning models (XGBoost, LightGBM, CatBoost) and deep-learning models (TabNet, FT_Transformer, NODE) to determine the most effective model.

Missing data were completed using modern deep-learning-based imputation approaches (GAIN-OneHot, GAIN-Emb, MIDAS).

The class imbalance problem was addressed through advanced synthetic data generation techniques (CTGAN, TVAE), and the model performances revealed that TVAE achieved superior results.

Using the SHAP explainable artificial intelligence method, the most influential variables contributing to the model’s decision-making process were identified, and it was observed that survival prediction could be successfully performed with a reduced number of features after feature selection.

The survival prediction performances of machine-learning and deep-learning models were compared, and the LightGBM model demonstrated the best performance.

Consequently, the TVAE–GAIN-OneHot–LightGBM model was proposed for survival prediction in horses, achieving the highest classification performance with an AUC value of 0.928.