Machine-learning model shows moderate accuracy for predicting specialist nurse transfer intentions

A machine-learning model developed from 471 samples of specialist nurses produced moderate discrimination between respondents who intended to transfer jobs and those who intended to stay, according to the study abstract. The final model achieved an area under the receiver operating characteristic curve (AUC) of 0.766 after cross-validation and hyperparameter optimisation using the Tree-structured Parzen Estimator (TPE).

The model-development study compared several classification algorithms, including a multilayer perceptron classifier, random forest, XGBoost, TabNet, AdaBoost, LightGBM and CatBoost. The dataset was highly imbalanced: 367 participants, or 77.9%, intended to transfer, while 104, or 22.1%, intended to stay. To address this imbalance, the researchers used the synthetic minority oversampling technique (SMOTE), applied within cross-validation to mitigate class imbalance. SMOTE creates synthetic observations for the minority class so that nurses intending to stay reach the same sample size as those intending to transfer. The researchers separately cautioned that feature selection, scaling and other preprocessing require careful management to avoid leakage between training and testing data.

Shapley additive explanations (SHAP) were used to examine feature contributions. The analysis identified variables associated with stated transfer intention: whether satisfaction with performance distribution was high, monthly take-home income, the number of specialist training or continuing-education sessions attended, whether the nurse worked in a specialist nursing outpatient clinic, and the number of children younger than 12. These are predictive associations and do not show that changing income, training attendance, clinic duties or family circumstances causes a nurse to transfer. The study reported an unexpected positive association between satisfaction with performance distribution and the intention to transfer; the authors said this anomaly may reflect class imbalance or data biases rather than a robust underlying relationship.

The researchers described the model’s performance as moderate and recommended collecting more representative data and refining feature selection in future work. They identified additional limitations: XGBoost and CatBoost can be difficult to interpret, one-hot encoding can produce high-dimensional data when categorical variables have many values, and hyperparameter tuning across multiple models can require substantial computation. SMOTE may also generate synthetic observations that do not reflect real-world resignation patterns, introducing noise into training. The supplied evidence does not establish that the model is ready for operational use or that its predictions would improve retention outcomes, but it identifies variables that may help guide further evaluation of specialist-nurse workforce retention tools.

Subject of Research: Machine-learning prediction of specialist nurse job-transfer intentions

Article Title: Construction of a machine learning-based model for predicting specialist nurse transfer tendencies

Article References: Wang, G., Luo, Z., Wu, H., Wang, J., Ma, W., & Luo, L. (2026). Construction of a machine learning-based model for predicting specialist nurse transfer tendencies. BMC Health Services Research. https://doi.org/10.1186/s12913-026-15528-5

Image Credits: AI Generated

DOI: 10.1186/s12913-026-15528-5

Keywords: specialist nurses, job transfer intentions, machine learning, model development, SHAP explanations, class imbalance, SMOTE, workforce retention, Construction, machine, learning-based, model