Public Health. 2026 Jul 1;258:106393. doi: 10.1016/j.puhe.2026.106393. Online ahead of print.
ABSTRACT
OBJECTIVES: To develop and interpret a predictive model that detects key factors associated with nurses' intention to rectify online health misinformation, and to offer an evidence foundation for targeted interventions that strengthen frontline responses to misinformation.
STUDY DESIGN: A nationwide cross-sectional survey that incorporated machine-learning modelling and explainable artificial intelligence.
METHODS: Convenience sampling was used to conduct this cross-sectional online survey among registered nurses employed in Chinese healthcare facilities between May and June 2025. A dichotomous item was used to measure the intention to rectify online health misinformation. Thirty-nine predictors were collected, including sociodemographic traits, psychological constructs, health information attitudes, eHealth literacy, and work environment factors. Stratified data splitting, the Synthetic Minority Over-sampling Technique, and grid-search-based hyperparameter tuning were used to train seven machine-learning algorithms (i.e., logistic regression, random forest, AdaBoost, CatBoost, ExtraTrees, XGBoost, and LightGBM). Accuracy, recall, precision, F1 score, and the area under the receiver operating characteristic curve (AUC) were used to assess the model's discrimination ability. The best-performing model was interpreted with SHapley Additive exPlanations (SHAP). Structural equation modelling (SEM) was subsequently conducted to test theoretically hypothesized relationships among the key predictors identified by SHAP, providing theory-driven validation of the findings.
RESULTS: Data analyses from 1120 nurses suggested that 80.2% of them intended to actively rectify online health misinformation. On the test set, the ExtraTrees model performed the best (AUC = 0.919; accuracy = 0.827). The model demonstrated a balance between sensitivity (recall = 0.827) and precision (0.791), indicating reasonable performance despite class imbalance. SHAP analysis identified perception of information-related harm, working in general wards, education level, perceived information trustworthiness, years of clinical experience, and eHealth literacy skills as the most influential predictors. These variables clustered into three overarching domains, namely risk perception, professional competence, and clinical practice context. Higher values of these features were consistently associated with positive SHAP values, indicating an increased likelihood of correction intention. Notably, perception of information harm exhibited a clear monotonic positive relationship with the outcome. Visual inspection of SHAP dependence plots suggested that interaction effects among key predictors were limited, with most variables demonstrating consistent monotonic relationships with the outcome. Structural equation modelling further supported these findings, showing that capability factors (eHealth literacy) exerted a strong positive effect on psychological factors (risk perception and information attitudes) (β = 0.761, p < 0.001), a direct effect on intention (β = 0.336, p < 0.001), and an indirect effect mediated by psychological factors (β = 0.124, 95% CI [0.065, 0.188]), accounting for 27.0% of the total effect. The model explained 22.4% of the variance in intention (R2 = 0.224) with acceptable fit (SRMR = 0.056).
CONCLUSIONS: This study provides novel machine-learning-based and explainable evidence on the determinants of nurses' intention to correct online health misinformation, offering a comprehensive framework that integrates predictive modelling and theory-driven analysis. The ability of nurses to recognise harm caused by misinformation, critically appraise data, and mobilise digital health competencies within supportive practice settings influences their intention to rectify online health information. Enhancing eHealth literacy and risk communication skills may increase frontline capacity to reduce the negative effects of misinformation on public health, particularly in high-interaction ward settings.
PMID:42385290 | DOI:10.1016/j.puhe.2026.106393