Differentiating Hantavirus Pulmonary Syndrome (HPS) andHemorrhagic Fever with Renal Syndrome (HFRS) Using a Stacking Ensemble

Abstract

Distinguishing Hantavirus Pulmonary Syndrome (HPS) from Hemorrhagic Fever with Renal Syndrome (HFRS) becomes difficult once organ-specific manifestations emerge, while delayed laboratory confirmation can hinder timely clinical decisions. A major limitation is the absence of a publicly available structured dataset describing HPS/HFRS symptom profiles. As a methodological first step, this study compiled a synthetic dataset reflecting documented clinical patterns from the literature to evaluate a stacking ensemble using XGBoost and LightGBM as base learners with Logistic Regression as the meta-learner. Using 8,000 synthetic records and 22 symptom features, preprocessing included binary encoding, SMOTE applied only within cross-validation folds, and 5-fold stratified cross-validation with grid-search hyperparameter tuning. The proposed model achieved 94.87% accuracy, 95.12% precision, 94.61% recall, 94.86% F1-score, 95.52% specificity, an MCC of 0.891, and an AUC-ROC of 0.9821, outperforming both individual base learners. SHAP analysis identified cough, tachycardia, and pulmonary edema as the strongest HPS indicators, whereas proteinuria, facial flushing, and conjunctival injection were the strongest HFRS indicators, consistent with reported organ-specific manifestations. Although limited to synthetic data, the proposed framework demonstrates methodological potential for HPS/HFRS differentiation and provides a foundation for future validation using prospective clinical datasets.