Identifying Nurses at Risk of Nursing Interruptions During Medication Administration Using Machine Learning: A...

Nursing interruptions during medication administration (NIMA) critically contribute to medication administration errors. However, validated prediction tools for NIMA risk assessment are unavailable. This study aimed to develop and internally validate three machine learning-based prediction models (logistic regression [LR], decision tree [DT], and Naive Bayes [NB]) for identifying nurses' individualized risk of NIMA.

A total of 4758 Chinese nurses were recruited from 12 tertiary hospitals between November 2023 and January 2024. The outcome was defined as the occurrence of ≥ 1 NIMA during the latest work shift. We used univariate analysis and LR to identify predictors. Participants were randomly allocated to training (n = 3806; 80%) and internal validation (n = 952; 20%) sets. Three machine learning models were implemented in Python, with performance evaluated via 1000-iteration stratified bootstrapping. Key performance metrics included AUC, accuracy, recall, specificity, precision, F1-score, and G-mean. Models comparisons used the DeLong test. The best model was further assessed with the Hosmer-Lemeshow test and calibration curves.

Overall, 52.1% of nurses experienced at least one NIMA event. Predictive features of NIMA included 18 factors, such as department type, marital status, shift type, general self-efficacy level, and the needs of multiple people. In the test set, AUCs ranged from 0.679 to 0.748. The LR model performed best, achieving the highest AUC of 0.748 (95% CI: 0.717-0.779), an accuracy of 0.694 (95% CI: 0.664-0.724), and other performance metrics, including precision, recall, specificity, F1-score, and G-mean. The calibration of the LR model was supported by the Hosmer-Lemeshow test (X² = 7.062, p = 0.530).

This study systematically evaluated multiple influencing factors of NIMA and developed three internally validated risk prediction models of NIMA. The LR-based nomogram and the web-based calculators showed the most consistent performance and may support risk stratification and targeted nursing management, pending external validation and feasibility assessment.