A Global and Local SHAP-Driven Interpretable Framework for Early Hospital Admission Prediction With Machine Learning at Emergency Department Triage

Background: Emergency department (ED) crowding and constrained hospital resources drive the need for early risk stratification to efficiently identify patients requiring inpatient admission or escalation of care. Machine learning (ML) models are promising as helpful tools for admission prediction, yet there remains limited literature exploring the integration of explainability techniques to enhance the transparency of these models for clinicians and resource allocation staff end-users. Explainability, in this context, refers to techniques that show which patient features contributed most to each individual prediction and to the model's overall behavior. Objectives: To develop ML models predicting hospital admission at ED triage and to demonstrate population-level and individual-level explainability as a step toward building trust and fair evaluation of AI-assisted triage models. Methods: We conducted a retrospective study of ED visits (2020-2024) across three tertiary-care EDs. We extracted structured triage data, including patient demographics, vital signs, Emergency Severity Index (ESI), arrival mode, and comorbidities from the electronic medical record. Six models were developed and evaluated on a temporally distinct holdout test set. SHapley Additive exPlanations (SHAP) were applied to the two top-performing models (CatBoost and XGBoost) on a population and individual prediction level. Results: The training cohort included 620,729 visits (32.4% admitted); the test cohort included 183,789 visits (32.3% admitted). CatBoost achieved an area under the receiver operating characteristic curve (AUC) of 0.870 (F1 0.701; sensitivity 79.0%; specificity 78.0%). XGBoost achieved an AUC of 0.848. ESI level, vital signs, means of arrival, and presenting complaints were among the most influential predictors. Post-hoc Platt scaling reduced the Brier score from 0.146 to 0.134 and corrected a systematic over-prediction of admission probability (calibration intercept moved from -0.69 to 0.02). Conclusions: ML models using triage data predict hospital admission with good discrimination. Population-level and individual-level SHAP provide transparency that may facilitate clinician trust and fair model evaluation.