Related Experiment Videos
A Global and Local SHAP-Driven Interpretable Framework for Early Hospital Admission Prediction With Machine Learning
Adam E Brown1,2, Chance W Marostica1,2, Nicole R Hodgson3
1Emergency Medicine, Mayo Clinic Alix School of Medicine, Phoenix, USA.
Cureus
|July 20, 2026
Summary
Machine learning models accurately predict hospital admission using emergency department triage data. Explainability techniques like SHAP enhance transparency, fostering clinician trust and fair AI evaluation in healthcare.
Area of Science:
- Artificial Intelligence in Medicine
- Clinical Decision Support Systems
- Health Informatics
Background:
- Emergency department (ED) crowding necessitates efficient risk stratification for patient admission.
- Machine learning (ML) models show promise for predicting hospital admissions.
- Limited research exists on integrating explainability into ML models for clinical transparency.
Purpose of the Study:
- Develop ML models to predict hospital admission from ED triage data.
- Demonstrate population- and individual-level explainability for AI-assisted triage models.
- Enhance clinician trust and enable fair evaluation of AI models.
Main Methods:
- Retrospective study of over 800,000 ED visits (2020-2024) across three tertiary care centers.
- Extracted structured triage data including demographics, vital signs, ESI, and comorbidities.
- Developed and evaluated six ML models, applying SHapley Additive exPlanations (SHAP) to top performers (CatBoost, XGBoost).
Main Results:
- CatBoost model achieved an AUC of 0.870; XGBoost achieved an AUC of 0.848.
- Key predictors included Emergency Severity Index (ESI) level, vital signs, arrival mode, and presenting complaints.
- Platt scaling improved model calibration, reducing systematic over-prediction of admission probability.
Conclusions:
- ML models utilizing ED triage data demonstrate strong predictive performance for hospital admission.
- Population- and individual-level SHAP explainability enhance model transparency.
- Explainable AI can facilitate clinician trust and support fair evaluation of AI-driven triage tools.