Construction of a machine learning-based model for predicting specialist nurse transfer tendencies
Year: 2026
Type: article
Abstract: The shortage of specialist nurses poses a significant challenge to healthcare systems worldwide, impacting both patient care and hospital operations. This study explores the application of machine learning to predict the job transfer tendencies of specialist nurses, as well as determining the major causes of the transfer tendencies, aiming to provide hospitals with actionable insights for retention strategies. We applied multiple classification algorithms, such as Multilayer Perceptron (MLP) Classifier, Random Forest Classifier, XGBoost and TabNet Classifier. The final model achieved an area under the receiver operating characteristic curve (AUC) score of 0.766 after cross-validation and hyperparameter optimisation using the Tree-structured Parzen Estimator (TPE). To better understand feature contributions, we utilised Shapley additive explanations (SHAP) values to perform feature importance analyses. Our findings identified that key predictors of nurse transfer include whether satisfaction with performance distribution is high, monthly take-home income, number of specialist trainings/continuing education attended, whether working in a specialist nursing outpatient clinic and number of children under 12. However, certain results, such as the unexpected positive correlation between satisfaction with performance distribution and job transfer, revealed anomalies that are likely attributable to class imbalances or data biases. We also discuss the tradeoffs between different data preprocessing techniques, particularly regarding oversampling within cross-validation to mitigate class imbalances. While the performance of the model was moderate, the insights from the SHAP analysis provided valuable directions for future research. Hence, we recommend collecting more representative data and refining feature selection to improve model accuracy. Therefore, this research contributes to the development of predictive tools for hospital management to address the specialist nurse shortage by identifying nurses who are at risk of job transfer. Not applicable. Strengths: • Multiple model comparison: This study explores a wide range of algorithmic approaches to optimise predictive performance by implementing various machine learning models (TabNet Classifier, Random Forest, XGBoost, MLP Classifier, AdaBoost, LGBM Classifier, and CatBoost Classifier). • Cross-validation: The use of cross-validation ensures robust model selection and reduces the risk of overfitting, thus enhancing the generalisability of the results. • Hyperparameter tuning: The application of Optuna’s TPE-based optimisation allows for efficient hyperparameter selection, which leads to improved model performance. • Handling imbalanced data: Our dataset is highly imbalanced. Among the 471 samples collected, 77.9% (367) of the participants intended to transfer while only 22.1% (104) of the participants intended to stay. This study addresses class imbalance by incorporating the synthetic minority over-sampling technique (SMOTE). SMOTE will oversample the minority class (nurses who do not intend to transfer) until it has the same number as the majority (nurses who intend to transfer), thus ensuring the better prediction of minority cases. • Feature importance analysis: The use of SHAP to interpret feature importance adds transparency and interpretability, which facilitates understanding the decision-making processes of the models. Weaknesses: • Limited interpretability in some models: While powerful, models such as XGBoost and CatBoost can be complex, which makes it harder to interpret the underlying relationships compared with simpler models, such as logistic regression. • One-hot encoding scalability: Using one-hot encoding in preprocessing can lead to high-dimensional data, especially when categorical features have many unique values, which potentially impacts computational efficiency. • Potential data leakage risk: The extensive preprocessing steps, including feature selection and scaling, require careful management to avoid leakage between the training and testing data. • Hyperparameter tuning computational cost: Although Optuna optimises the process, the computational cost of hyperparameter tuning across multiple models can be high, particularly when combined with cross-validation. • SMOTE limitations: While SMOTE helps balance the dataset, it may generate synthetic data that do not perfectly represent real-world nurse resignation patterns, which could introduce noise into the training process.
Source: BMC Health Services Research
Institutions: Fudan University, Shanghai Jiao Tong University
Language: English
Topic: AI and HR Technologies
Domain: Social Sciences
Open Access status: gold
Estimated APC paid: $3,290
OpenAlex ID: W7222354302
FWCI:
Cites:
Cited by:
Related to: