Comparative performance of bagging and boosting ensemble models for predicting lumpy skin disease with...

Area of Science:

• Veterinary Medicine
• Machine Learning
• Epidemiology

Background:

• Lumpy Skin Disease (LSD) significantly threatens livestock health and causes economic losses.
• Ensemble machine learning (ML) offers advanced decision support for disease prediction in veterinary contexts.
• Accurate prediction models are crucial for effective risk management in livestock populations.

Purpose of the Study:

• To develop and evaluate ensemble ML models for predicting Lumpy Skin Disease (LSD) in Egyptian livestock.
• To address the challenge of multiclass imbalance in veterinary disease datasets.
• To identify key predictors for LSD to inform targeted intervention strategies.

Main Methods:

• Collected 1,041 records from six Egyptian governorates (June 2020-October 2022).
• Applied data balancing techniques: SMOTE, Random Oversampling (ROS), and Random Undersampling (RUS).
• Evaluated five ensemble models (DT, RF, AdaBoost, GBoost, XGBoost) using hyperparameter tuning and 10-fold cross-validation.

Main Results:

• The Random Forest model combined with ROS (RF-ROS) achieved the highest accuracy (82%) and AUC (0.93).
• Balanced XGBoost also showed strong performance (81.25% accuracy, AUC=0.93).
• SHAP analysis identified vaccination status as the most significant predictor of LSD outcomes.

Conclusions:

• Combining resampling techniques with hyperparameter tuning significantly improves ML model performance on imbalanced veterinary datasets.
• Ensemble ML models, particularly Random Forest with ROS, are effective tools for LSD prediction.
• Vaccination status is a critical factor for LSD prevention and control, guiding targeted interventions.