AI-Based Decision Support and Data-Driven Optimisation in Emergency Departments

Sammanfattning

Emergency Departments uncertainty introduces key concerns into the scheduling problem. Overcrowding depletes medical personnel, and increases morbidity within EDs. This thesis investigates whether Large Language Models (LLMs), like llamaa 3.3 70b-versatile, can be used for enhancing patients acuity prediction and to see whether introducing acuity levels to the model affect the hybrid robust-stochastic scheduling algorithm. Such modelling were carried out using discrete event simulation (DES). The assessment used fundamental evaluation metrics (accuracy, precision, recall, F1-score), Cohen’s Kappa Statistic, discrepancy breakdown, interview results received by a specialist, Door-to-Provider (DTP) times, Length of Stay (LOS) and Door Violations. The results show that LLMs can decrease human-introduced bias within acuities, provide professional reasoning, and can serve as decision-support system. Yet the LLMs performance deteriorates substantially when class imbalance is present. The hybrid robust-stochastic optimisation algorithm combined with patients acuities improved system performance, positively affecting DTP, LOS and Door Violation. These findings suggest that the proposed pipeline, combining acuities and optimisation algorithm, is suitable for a simulation-only environment. Yet, future evaluation with more developed LLM models (like GPT-4-Turbo), and in nurse-observed simulation settings should be examined before conclusion can be stated related to real-world deployment.