Application of healthcare data mining techniques to planning for nursing length of stay in surgical departments

Effective allocation of nurse resources in surgical departments is essential for improving patient care and controlling operating costs in a health society. Length of stay (LOS) is the metric that connects clinical workload to
staffing decisions, yet ward-level forecasting and its translation into daily nursing schedules remain limited. This study presents a hybrid, data-driven decision-support system that combines machine-learning LOS prediction with Reinforcement Learning (RL) for the surgical ward. A dataset of 137,145 records is used to evaluate Random Forest, Gradient Boosting, Decision Tree, and a Multi-layer Perceptron. Random Forest achieved the most accurate and stable performance (R² = 0.84; RMSE = 1.63), and its predicted LOS states drive an RL agent that adjusts staffing and triggers early-discharge reviews. The novelty lies in focusing on the understudied surgical ward, converting predicted LOS into a daily scheduling policy, and integrating forecasting with RL-based scheduling. The hybrid model reduced average LOS from 6.12 to 4.82 days, lowered weekly nurse overtime by approximately 47%, and improved staff utilization.