Construction of a Model to Estimate Nurse Workload in Real-Time and Proactively under Fluctuating Conditions

Abstract

This study aimed to construct a model that proactively predicts each nurse’s workload five minutes ahead in real time by utilizing information obtainable within the ICU, through simulation experiments simulating nursing tasks in the ICU. An experiment was conducted where 30 teams of five members each performed simulated ICU nursing tasks for two hours, collecting NASA-TLX data every five minutes along with various data concerning the nurses （experimental participants） and patients （simulators）. We applied five machine learning methods to predict five workload levels based on the 5-minute NASA-TLX WWL score, using 130 features derived from aggregated nurse and patient data, and compared their accuracy. While the prediction accuracy for the five levels did not exceed 40% for any model, it was found that the GBDT （gradient-boosting decision tree） model could achieve nearly 80% accuracy if a ±1 difference was tolerated. Analysis of feature importance revealed that incorporating data not only from the nurse being predicted for, but also from other nurses and patients not under their care, was effective for predicting each nurse’s workload level.