HEXAID: A Task-adaptive Multimodal Ensemble Framework for Disease Identification and Severity Stratification in Emergency Departments

Clinical decision-making in emergency departments (EDs) requires rapid integration of heterogeneous patient information, including structured vital signs, physical examination findings, and free-text clinical narratives, to support disease identification and patient severity stratification.However, a single modeling strategy may not be equally effective across clinical prediction tasks with different decision-boundary characteristics.This study proposes HEXAID, a task-adaptive multimodal ensemble framework that integrates structured clinical data with n-gram TF-IDF representations of free-text clinical narratives for unified ED decision support.The study used 5,000 electronic medical records from an Indonesian ED, of which 4,374 valid records remained after a multi-stage preprocessing pipeline, resulting in a 719-dimensional feature matrix.A systematic ablation study comprising ten experimental configurations evaluated baseline models, hybrid RF-XGBoost soft voting, Chi-Square feature selection, Bayesian hyperparameter optimization using Optuna TPE, and class-imbalance handling strategies.For disease identification, the Optuna-optimized full-feature ensemble achieved an F1-Macro of 0.9255 and AUC of 0.9922, outperforming the baseline models.For severity stratification, the selected imbalance-handling strategy achieved an F1-Macro of 0.8579 and AUC of 0.9567, with a Severe-class recall of 0.75.The ablation results demonstrate that optimal ensemble configurations are task-dependent: disease identification favored XGBoost-dominant weighting, whereas severity stratification benefited from a more balanced RF contribution.SHAP analysis further showed that temperature was the dominant feature for disease identification, while systolic blood pressure was prominent for severity stratification, providing clinically plausible model interpretations.These findings demonstrate that task-adaptive optimization is important for multimodal clinical decision-support models, particularly when prediction tasks differ in decisionboundary structure and class distribution.