ORCID

Abstract

Emergency departments face constant pressure from overcrowding, making early prediction of patient admission a valuable support for clinicians. In this study, we used the MIMIC-IV-ED v2.2 dataset, containing about 296,000 visits and 38 triage-level features, to develop and compare multiple machine learning models for admission prediction. Across five approaches - Logistic Regression, Decision Tree, Random Forest, XGBoost, and a Deep Neural Network - performance ranged from moderate to strong, achieving AUROC values up to 0.84 and balanced accuracy around 77%. Despite these results, recall for admitted patients remained around 60%, indicating that many potential admissions were not detected. Explainable AI methods (SHAP and LIME) identified triage acuity, patient age, arrival transport, and medication counts as key drivers of model decisions. Fairness analysis revealed demographic disparities, with younger patients predicted more accurately than older adults, and elderly women particularly disadvantaged. Compression experiments further showed that quantisation and pruning reduced model size and latency with minimal performance loss. The study highlights the potential of predictive triage systems while underscoring the importance of fairness monitoring, calibration, and regulatory compliance before deployment.

Keywords

Admission Prediction, ED, EU AI Act, XAI Fairness

Publication Date

2026-02-16

Event

3rd International Conference on Human-Centred AI - Education and Practice, HCAI-ep 2026

Publication Title

HCAI-ep 2026 - Proceedings of the 2026 Conference on Human Centered Artificial Intelligence - Education and Practice

Publisher

Association for Computing Machinery (ACM)

ISBN

9798400721533

First Page

119

Last Page

119

Deposit Date

2026-11-04

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Download


Share

COinS