Automated Differentiation of Hantavirus Pulmonary Syndrome (HPS) and Hemorrhagic Fever with Renal Syndrome (HFRS) Using Stacking Ensemble Based on Clinical Manifestation Profiles

Abstract

The clinical differentiation of HPS and HFRS is a critical but unresolved diagnostic challenge, particularly in resource-limited endemic settings where confirmatory laboratory diagnostics are unavailable. Despite the rising global burden of hantavirus infections and their high case fatality rates up to 40% for HPS and 12% for HFRS no prior study has specifically developed an automated machine learning system for HPS/HFRS binary differentiation using solely clinical manifestation profiles. This study proposes a stacking ensemble classification framework combining XGBoost and LightGBM as base learners with Logistic Regression as the meta-learner to automatically differentiate HPS from HFRS based on clinical symptom data. A synthetic clinical dataset of 8,000 patient records encoding 22 unique symptom features was used. Preprocessing included binary encoding, SMOTE-based class balancing, and 5-fold stratified cross-validation. SHAP analysis provided feature-level explainability. The stacking ensemble achieved accuracy of 94.87%, precision of 95.12%, recall of 94.61%, F1-score of 94.86%, and AUC-ROC of 0.9821, outperforming individual base learners by 3–13 percentage points and improving upon the best prior hantavirus classification result (88.5%) by 6.4 percentage points. SHAP analysis confirmed discriminative alignment with established pathophysiology: Cough, Tachycardia, and Pulmonary Edema dominated HPS, while Proteinuria, Facial Flushing, and Conjunctival Injection dominated HFRS. Ensemble learning on clinical symptom profiles provides a clinically interpretable, laboratory free tool for automated HPS/HFRS differentiation, with direct applicability for clinical decision support in endemic and resource-constrained settings.