Abstract

Early and correct identification of dengue is vital in the reduction of mortality and in better clinical decision-making, especially in healthcare systems that lack resources. The proposed research finds a strong and decipherable machine learning model in predicting dengue based on clinical hematological samples. The mathematical models evaluated were five state of the art models: Random Forest, Extra Trees, XGBoost, Support Vector Classifier, and CatBoost with the optimization of hyperparameters and rigorous preprocessing. The experimental findings indicate that CatBoost has the best overall performance with an accuracy of 98.80%, an AUC of 0.9999, and a high Matthews Correlation Coefficient (MCC) of 0.9722, which is reflective of an excellent and balanced predictive capacity. Additionally, 5-fold crossvalidation was used to verify the generalization capacity of the model, achieving the best mean accuracy of 0.9990 with low variance across folds. CatBoost was also computationally efficient with a fast inference time (0.0036 s) meaning it can be used in real-time clinical applications, but the associated training cost is moderate. Explainable artificial intelligence (XAI) with SHAP was incorporated to increase transparency, and it showed the main contributions of key features, which include WBC Count and Platelet Count. Altogether, the suggested framework offers a predictable, effective, and interpretable dengue prediction solution that solves the shortcomings of the current methodology and offers a practical healthcare management solution.