Table 6 Key contribution and novelty of our approach against the used methods.

Baseline (Default)

Ensemble models (RF, XGB, AdaBoost, etc.) without resampling or tuning.

Poor performance for minority classes; moderate overall due to high bias toward the majority class.

DT performed poorly; RF and XGBoost showed slightly better robustness.

Establishes a baseline for assessing improvements.

Improved Tuning Only

Ensemble models with hyperparameter tuning only.

Moderate improvement, especially in DT and GBoost.

Tuning improved the performance of weak learners (DT, Ada, GBoost).

Highlights the role of tuning in mitigating class imbalance effects.

Resampling Only

Resampling using ROS, RUS, SMOTE with default model parameters.

ROS > SMOTE > RUS in improving minority class recall.

Oversampling improved recall; RUS reduced overall accuracy.

Validates ROS as a simple and effective resampling method.

Combined (Resampling + Tuning)

Ensemble models + resampling + tuning.

Achieved the highest overall performance across all metrics.

RF-ROS with tuning gives the best results.

First study to evaluate and compare both resampling and tuning for multiclass LSD prediction.

Model Rankings

Across all combinations

RF > XGB > Ada > GBoost > DT.

RF dominant in precision and F1, even with imbalance.

Provides empirical evidence of model robustness.

SHAP Interpretation

Applied to the best-performing model ((RF + ROS + tuning)).

Identified vaccination status, grazing system, and season as top predictors.

Neethling vaccine linked to healthy cases.

Enhances model transparency and ML interpretability using SHAP.