Development and External Validation of a Machine Learning Model for Risk Stratification of Presenteeism in Clinical...

Presenteeism impairs clinical nurses' work performance and patient safety, yet externally validated risk estimation tools capturing nonlinear effects are lacking.

To develop and externally validate a machine learning-based risk estimation model for presenteeism among clinical nurses, and to explore predictor roles across risk levels.

This multicenter cross-sectional survey recruited 19,729 clinical nurses from 153 hospitals across 9 cities in Guizhou, China (July-November 2025), using the Stanford Presenteeism Scale as the outcome. Predictors were selected via Least Absolute Shrinkage and Selection Operator regression and multiple linear regression. Nine machine learning models were developed (derivation: n = 13,323) and externally validated (n = 6406). Nurses were stratified into low-, medium-, and high-risk groups by tertiles of predicted scores. The optimal model was interpreted using SHapley Additive exPlanations (SHAP) and partial dependence plots. Regularized partial correlation networks were estimated within each risk group.

Five predictors were retained: burnout (Maslach Burnout Inventory-General Survey, MBI-GS), perceived social support (PSSS), organizational climate, workplace violence, and night shift involvement. The generalized additive model (GAM) performed optimally. SHAP showed burnout had the highest main effect, most pronounced in low- and high-risk strata. Partial dependence plots revealed a linear increasing effect of MBI-GS, a gently descending trend of PSSS, and an inverted U-shaped relationship for organizational climate. Network analysis showed denser networks with more negative edges as risk escalated; department leadership/communication consistently served as a core node, while burnout node centrality increased markedly in the high-risk group.

This study developed and cross-regionally validated a GAM-based risk estimation model for clinical nurse presenteeism. The advantage of GAM over linear regression was its ability to capture nonlinear dose-response relationships, notably the inverted U-shaped organizational climate association. Burnout was the strongest predictor across all risk levels. These findings provide preliminary evidence and visualization approaches that may inform risk-stratified management of presenteeism, pending prospective validation.

A stepped, risk-stratified approach to presenteeism may be considered for future evaluation: preventive resource building for low-risk nurses, threshold monitoring for medium-risk nurses, and enhanced organizational support for high-risk nurses. Strengthening department leadership and communication warrants particular attention in intervention design.