International Journal of Medical Discoveries

Background: The efficient and equitable allocation of hospital resources, specifically critical care beds, represents a major operational and ethical challenge in healthcare. Traditional predictive models for this task often suffer from a “black box” nature, hindering clinician trust and adoption.

Aim: The study aimed to develop, validate, and evaluate an Explainable AI Decision Support System (XAI-DSS) framework for optimising hospital bed allocation by integrating high-accuracy predictive modelling with stakeholder-centric XAI.

Methodology: The study employed a sequential mixed-methods approach, beginning with a retrospective cohort analysis of the MIMIC-IV database to train and validate three Machine Learning models (XGBoost, Random Forest, Neural Network). A randomized controlled user study and a simulated operational trial then evaluated the XAI utility and operational impact.

Results: The Random Forest model achieved near-perfect predictive accuracy, recording an AUC-ROC of 0.999998 and, critically, a 1.00 Recall with zero False Negatives. Global SHAP identified Creatinine and Respiratory Rate as key drivers, confirming clinical objectivity. The XAI-DSS achieved a ‘Good’ System Usability Scale score (Target >70) and demonstrated a 15% reduction in mismatched bed allocations in the simulated trial.

Conclusion: The XAI-DSS provides a safe, trustworthy, and operationally efficient framework for critical resource management, validating the necessity of multi-modal interpretability for clinical adoption.

Future Recommendation: Future studies should focus on a prospective, real-world shadowed deployment to externally validate the 15% efficiency gain.