Real-time prediction of emergency department admissions using an AI model and its integration into hospital bed-planning

OBJECTIVE: To design, validate, and implement a tool based on a machine-learning model capable of predicting emergency patient admissions in real time, and to develop an application that integrates the model together with traditional indicators of health care pressure. METHODS: We analyzed anonymized data from a total of 328,696 emergency episodes. Fifteen variables with the highest predictive capacity were identified through Pearson correlation analysis. The RandomForest model was selected. Python was used for model construction, and "Jupyter Notebook" for design and validation. A total of 209,530 episodes (64%) were used for training, 52,383 (16%) for validation, and 66,783 (20%) for testing. The model was integrated into a Control Panel using Microsoft Power BI. RESULTS: The model showed a high predictive ability to identify patients who would require admission, with a receiver operating characteristic (ROC) area under the curve (AUC) of 0.97. For each metric, a decrease in training score and a concurrent increase in validation score were observed, ruling out overfitting behaviors. The saturation point in the learning curves was reached at 130,000 episodes. Model calibration yielded a Brier score of 0.0414. CONCLUSIONS: The model demonstrates excellent performance and significant potential for predicting hospital admissions among emergency department patients. Combining machine-learning algorithms with interactive visualization systems represents a major step toward more efficient management and planning, helping avoid bottlenecks and overcrowding.