Identifying the Most Critical Predictors of Workplace Violence Experienced by Junior Nurses: An Interpretable Machine Learning Perspective

Abstract

Background: Workplace violence, defined as any disruptive behavior or threat to employees, seriously threatens junior nurses. Compared with senior nurses, junior nurses are more vulnerable to workplace violence due to inexperience, low professional recognition, and limited mental resilience. However, there is an absence of research discussing the workplace violence risk of junior nurses, in particular, the lack of analysis of critical factors within the multiple influences and the lack of targeted risk prediction models. Objective: Considering the multiple influencing factors faced by junior nurses, this study aims to predict the risk of workplace violence using interpretable machine learning models and identify the critical influencing factors and their nonlinear effects. Design: An observational, cross-sectional study design. Participants: A total of 5663 junior registered nurses in 90 tertiary hospitals in Sichuan Province, China. Methods: Data are all obtained through a questionnaire survey. An interpretable machine learning framework, including the Light Gradient Boosting Machine (LightGBM) model and two post hoc interpretable methods, Accumulate Local Effect and SHapely Additive exPlanations (SHAP), are conjoined. Results: The LightGBM model is more accurate than other machine learning methods, achieving an area under the receiver operating characteristic curve of 0.761 and a Brier score of 0.198 on the workplace violence prediction task. Among the dozens of potential influences input into the predictive model, seeing medical complaints, psychological demands, professional identity, etc., are the most critical predictors of workplace violence. Conclusions: The proposed LightGBM-SHAP-ALE approach dynamically and effectively identifies junior nurses at high risk of workplace violence, providing a foundation for timely detection and intervention.

Keywords: critical predictors identifying; interpretable machine learning; junior nurses; risk prediction; workplace violence.

Figure 1

The predictor filtering framework in this study.

Figure 2

Hierarchical cluster analysis of predictors in this study.

Figure 3

Overall predictive analysis framework. (a) Baseline predictive analysis and (b) robustness analysis.

Figure 4

ROC curves of all utilized machine learning prediction models.

Figure 5

Global-level most important predictors under LightGBM-based SHAP analysis (top 20). (a) Feature importance is shown in the bar plot. (b) Feature importance is shown in the Beeswarm plot.

Figure 6

SHAP-based instance-level result of a randomly selected true positive sample.

Figure 7

SHAP-based instance-level result of a randomly selected true negative sample.

Figure 8

ALE-based analysis of the effects of top-seven important features on prediction.

Figure 9

Logistic regression-based nomogram. Top 14 predictors with the largest coefficient absolute values are shown.