Short-Horizon Queue-Length Forecasting for Outpatient Merge Stations

Accurate short–horizon forecasts of the queue length at downstream merge stations—where multiple upstream desks converge—are vital for dynamic triage, load balancing and anticipatory staffing in outpatient clinics. We model patient arrivals to each upstream desk as a non-homogeneous Poisson process and encode real-time queue states into a compact feature vector. Within this supervised-learning framework we benchmark five algorithms—k-nearest neighbors, random forest, Bayesian ridge regression, single-layer perceptron and multi-layer perceptron (MLP)—for one-step-ahead prediction of the merge-station queue length. The results indicate that a lightweight MLP consistently delivers the best accuracy across all load segments and markedly reduces under-forecasting during peak periods, while deeper or wider networks offer diminishing returns. The proposed predictor operates with modest computational overhead and can be embedded in real-time routing and staffing policies to shorten queues and enhance service consistency during congestion.