Explainable Machine Learning for Hospital Overcrowding Prediction: A Comparative Study of Predictive Models

Abstract

Hospital overcrowding requires capacity-based or flow-based measurements, such as occupied beds relative to staffed capacity, boarding time, or emergency-department census. The public NHS Hospital Admitted Patient Care Activity 2023-2024 data used in this study do not contain such an endpoint. We therefore reframe the analysis as a descriptive classification audit of an emergency admission intensity label, defined as emergency admissions divided by total admissions exceeding 85%. K-Nearest Neighbors (KNN), Random Forest (RF), and Gradient Boosting (GBM) were compared using accuracy, precision, recall, and F1-score. In the reported held-out row split, GBM obtained accuracy 0.967, precision 0.9255, recall 0.9062, and F1-score 0.9158. However, because the label is deterministically constructed from contemporaneous admission counts and emergency admissions are among the model inputs, these scores quantify reproduction of the constructed rule, not prospective forecasting of overcrowding. LIME examples are consequently interpreted only as local descriptions of model behavior. The study demonstrates both the apparent performance and the methodological limits of applying explainable classifiers to annual aggregate admission data, and specifies the hospital-by-time, independently measured outcomes required for a valid overcrowding forecasting study.