An interpretable machine learning model for dengue detection with clinical hematological data

Abstract

Dengue fever remains a major global health concern that demands rapid and accurate diagnosis to prevent severe complications and support timely patient care. Traditional approaches relying on environmental variables often lack patient-level precision, limiting their clinical applicability. This study focuses on hematological parameters as more reliable indicators for early dengue detection. A novel machine learning framework, DengueStackX-19, was developed using 1,523 clinically verified patient records from Jamalpur 250-Bedded General Hospital, Jamalpur, Bangladesh. The dataset underwent rigorous preprocessing, normalization, and imbalance handling using various resampling techniques. Comparative evaluation across five balancing methods demonstrated that DengueStackX-19 consistently achieved the highest accuracy and robustness, performing effectively both before and after outlier removal. The model achieved 93.65 % accuracy and 89.63 % F1 during 10-fold cross-validation under SMOTEENN, and further attained 96.38 % accuracy and 94.20 % F1 in dengue classification, demonstrating robust generalization and consistent high performance across evaluation phases. Sensitivity analysis further verified its stability under feature perturbations. To ensure interpretability, SHAP and LIME were applied to identify the hematological factors most influential to the model's predictions, and the resulting patterns aligned with established clinical understanding. The model was deployed as an accessible web-based diagnostic tool, allowing healthcare professionals to perform real-time dengue detection without specialized laboratory infrastructure. This study demonstrates that hematology-driven AI models can significantly enhance diagnostic accuracy, reduce decision-making time, and improve patient outcomes, particularly in resource-limited settings.