Dengue fever represents a significant public health challenge in Bangladesh and is increasingly becoming a global concern. This research presents a machine learning system for the early diagnosis of dengue utilizing Bangladeshi hematological clinical data. To determine the primary risk factors, we conducted comprehensive statistical evaluations, including tests for normality, variance analysis, t-tests, and Z-tests. The SMOTEEN balancing strategy significantly improved performance, with the KNN model achieving the best individual accuracy of 96.83%. Additional ensemble and optimized models exhibited robust performance, exceeding 92% accuracy. Integrating all models using a stacking classifier improved performance, resulting in an optimal total accuracy of 98.1%. Additionally, methods of XAI were employed to identify the most crucial elements in dengue diagnosis. The SHAP study confirmed the diagnostic relevance of clinically relevant measures such as monocyte and neutrophil percentages. The results highlight the efficacy of combining sophisticated data preparation with ensemble learning to provide precise and reliable dengue identification. This approach not only increases prediction precision but also ensures its practical utility in real clinical settings. In addition, a real-time diagnostic interface was implemented using Streamlit, serving as a clinical decision support system to facilitate disease identification and assist both healthcare professionals and patients.