Evaluation of Predictive Models of Inpatient Nursing Workloads

The efficient allocation of medical personnel to clinical care requires adapting to fluctuating requirements. Nursing workloads at hospital wards depend not just on patient numbers (occupancy), but are also associated with individual features that are increasingly available in Electronic Health Records (EHRs). These EHRs thus promise to support human resource management with automated predictions of patient-specific nursing needs. To evaluate EHR-based workload estimates, we trained Random Forest (RF) and Gradient Boosting (GB) models in an anonymous dataset of 27,178 inpatient cases and 53 attributes from an acute-care hospital. The calculated statistical metrics demonstrated that GB models outperformed RF models, which appeared to suffer from overfitting. Reducing input data to 25 relevant attributes had little effect on predictive performance. Future investigations might explore how well such methods would work in different settings, or study other predictive algorithms such as neural networks.