Machine learning and SHAP interpretation for predicting high-level compassion fatigue among operating room nurses: A multicenter cross-sectional study

Abstract Background Operating room nurses (ORNs) are at high risk for compassion fatigue (CF), which significantly impairs individuals’ well-being, undermines the stability of the nursing workforce, and jeopardizes patients’ safety. The study aimed to analyze the prevalence and symptom characteristics of CF among ORNs, construct and compare predictive models using machine learning, and determine the relative importance of distinct features in CF prediction. Methods This is a multi-center cross-sectional study. The questionnaires used in the study included a sociodemographic questionnaire, the Professional Quality of Life Scale (ProQoL), the Patient Health Questionnaire-9 (PHQ-9), the Generalized Anxiety Disorder 7-item Scale (GAD-7), and the Pittsburgh Sleep Quality Index (PSQI). Five machine learning-based predictive models were constructed and compared: decision tree (DT), logistic regression (LR), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost). SHapley Additive exPlanation (SHAP) were drawn to show the contribution of each feature to the models. SPSS version 26.0 and R software version 4.4.0 were used for statistical analyses. Results In this study, a total of 1024 ORNs from 20 cities across China were recruited. According to ProQoL, 1000 (97.7%) ORNs reported CF (moderate or severe: 637 cases; mild: 363 cases), with only 24 (2.3%) unaffected. Among the three dimensions, the incidence of secondary traumatic stress was the highest (95.4%), followed by low compassion satisfaction (61.3%) and burnout (35.0%). In five machine learning-based predictive models, the RF model stood out with the highest AUC at 0.956 (95%CI: 0.954–0.958) in training set and 0.851 (95%CI: 0.795–0.907) in testing set. Following closely, the XGBoost model showed favorable efficacy with the AUC at 0.865 (95%CI: 0.837–0.893) in training set and 0.824 (95%CI: 0.769–0.879) in testing set, outperforming the remaining algorithms. The results of the two SHAP plots (RF and XGBoost) were consistent: depression, anxiety, self-mental health training, sleep quality, and length of service emerged as the five most significant contributors to the models. Conclusion This study identified an extremely high incidence of CF among ORNs, and the most serious symptom was secondary traumatic stress. The model of RF exhibited the best performance for predicting CF of ORNs and SHAP improved the interpretability of the model. All of those enabled medical managers and researchers to better understand the causes of CF, to give early identification and timely interventions. Clinical trial number Not applicable.