AI-powered models for overcrowding prediction at TUMS hospitals.

When the emergency department (ED) of a hospital reaches its saturation capacity, adverse events such as increased waiting times, increased mortality, clinical mistakes, and a financial burden escalate. Accurate overcrowding forecasting renders a powerful alarming tool for increasing the operational efficiency of hospitals. This study provides a novel contribution by leveraging a time series dataset for 5 hospitals of Tehran University of Medical Sciences (TUMS) to predict daily ED-overcrowding. Firstly, we identified 121 possible features per ED. Then, in the attribute selection, we used a new time series features selection method using minimum redundancy and maximum relevance (mRMR) with mutual information selector, and elastic net, both implemented via time series cross validation with growing window (TSCVG). In time series modeling, we implement several base learners, namely linear regression, elastic net, neural network, long short-term memory (LSTM), support vector machine (SVM), random forest, least square boosting, and a convolutional neural network with Bayesian tuning upon TSCVG folds. These base learners are combined by using ensemble methods, namely weighted and quantile averaging, and variance-penalized Bayesian Model Averaging (VP-BMA). Results indicated that reducing the initial feature set using the feature selection method significantly improves the modeling performance for all EDs. It is observed that the binary calendar feature and the delays of the target are frequently selected in all EDs. According to the VP-BMA model component analysis, the SVM and LSTM showed the most reliable performance across all EDs. As a result, ED A indicated a stable performance across most overcrowding thresholds (ROC ≈ 0.95-0.98), EDs C and E a reasonable at moderate cutoffs, while EDs B and D showed poor performance in all overcrowding thresholds.