A Stacking Ensemble Approach for Robust Dengue Patient Detection from Complete Blood Count Data

Dengue, a virus transmitted by mosquitoes, is a growing global health concern given its escalating incidence, significant death rates, and intense clinical signs. Despite the severity of the disease, there are substantial difficulties in accurately classifying dengue patients, correct patient classification remains problematic, which is crucial for timely intervention and patient management. In this study, we propose a comprehensive method for classifying 248 dengue-positive cases and 252 dengue-negative patients using tabular complete blood count (CBC) data from two different hospitals. There are missing variables in the dataset which is handled using Multivariate Imputation by Chained Equations (MICE) algorithm. This involves rigorous data preprocessing like data cleansing, statistical analysis, and missing data imputation. Among the different CBC parameters and demographic variables, by employing feature ranking and selection techniques, we are able to identify key characteristics. Thirteen classical machine learning (ML) models were trained for 5-fold cross-validation and finally, a Stacking-Based Meta-Classifier was trained using three top-performing model for Dengue patient identification along with a Nomogram-Based Scoring System. Extra Tree, Adaboost, and CatBoost Meta classifiers excel in their in-performance metrics. XGBoost Meta classifier achieves the highest F1-score of 97.8%. The Area Under Receiver Operating Characteristic Curve (ROC-AUC) scores are 0.976 for AdaBoost and 0.972 for Extra Tree and CatBoost, while the XGBoost meta-classifier attains an AUC score of 0.978. Shapley values shed light on feature contribution characteristics. Our proposed approach offers a robust framework for reliable dengue detection, facilitating timely medical response and easing the burden on health systems.