In large hospitals, outpatient phlebotomy units (OPUs) often face overcrowding during peak hours due to limited resources, leading to longer patient queues and compromised care quality. Factors such as service capacity, patient age, and wheelchair usage can significantly impact queue length. Accurately predicting queue lengths and identifying key influencing factors is essential for effective resource management. Machine learning (ML) methods are commonly used in clinical settings for their ability to handle complex feature interactions. However, ML methods provide limited explanations of how a key factors influencing the predictive outcome. These detailed insights could help managers to be more informed when planning resources allocations. Moreover, sufficient time is required to respond to and mobilize resources when managing a real-case scenario. Proactively forecasting future queue lengths can provide information to support managers in reacting effectively during real-time service operations. It is a challenging and complex task for managers to manage OPUs as wide variety of aspects are needed to be considered. To address the challenges, this study develops a feature selection scheme that incorporates SHapley Additive exPlanations (SHAP) to gain more detailed feature insights and direct strategy of multi-step ahead forecasting to predict future queue length at different time steps in OPUs. Using OPU data from a Taiwanese medical center (2017–2019) and three well-known ML methods of random forest (RF), least absolute shrinkage and selection operator regression (Lasso) and extreme gradient boosting (XGB) under the proposed scheme, RF emerged as the most accurate model across all horizons. Wheelchair usage was consistently the most influential feature, while elder patients became critical in three-step ahead forecasting. SHAP provided detailed insights into how these features affect queue length, supporting better resource planning and operational decision-making.
