A convenience sample of registered nurses was recruited from nine tertiary hospitals, and data were collected from March to December 2023. The eligibility criteria for nurses included being registered nurses with full-time employment and providing informed consent to participate in this study. Trainee nurses were excluded. The data were collected via questionnaires through a web-based system. The researchers obtained approval from the principal nurses to allow participants to complete the questionnaires after they explained the purpose of the study to the directors of the nursing departments. Before the questionnaires were distributed, the researchers explained the purpose of the study to prospective participants. The questionnaires were then distributed and collected by the researchers. To adhere to the principle of having at least 10 subjects per variable in the prediction model, a minimum sample size of 250 subjects was required (22). A total of 3,808 nurses were enrolled in this study.

The incidence of nurses’ psychological disturbance was the main outcome of our study. The Symptom Checklist 90 (SCL-90) was used to collect data related to psychological health. We employed the Chinese version of the SCL-90 translated by Wang (23), which consists of 90 items covering 10 dimensions: somatization, compulsion, interpersonal sensitivity, depression, anxiety, hostility, terror, paranoia, psychosis, and others. Each item is rated on a 5-point Likert scale ranging from 1 (not at all) to 5 (extremely), reflecting the severity of symptoms. The total score, calculated as the sum of all 90 items, ranges from 90 to 450, with higher scores indicating poorer psychological health. Scores for each dimension were computed as the mean of all items within that dimension. A total score exceeding 160, more than 43 positive items (scores ≥ 2), or any dimension score above 2 was considered indicative of a potential psychological disturbance. In this study, the scale demonstrated excellent internal consistency, with a Cronbach’s α coefficient of 0.99.

The selection of potential predictors for the psychological disturbance prediction model was guided by existing literature and framed within the Job Demands-Resources (JD-R) model (24), which highlights the complex and multifactorial nature of nurses’ psychological health. According to the JD-R model and previous studies, psychological outcomes may be influenced by the balance between job demands and available resources. Specifically, job demands-such as workload and shift patterns-may increase stress and contribute to psychological distress (25), while resources like social support and organizational career management probably serve as protective factors that buffer against these adverse effects (26–29). Based on this framework, the predictors included in this study cover three main domains (a total of 26 variables): living conditions (including 9 variables: educational background, average monthly income, marital status, whether to raise children, regular degree of three meals, weekly leisure frequency, weekly leisure time, weekly family companionship frequency, weekly family time), working situation (including 14 variables: hospital list, working years, employment type, working type, daily/night shifts per month, patients in day/night shift care, day/night shift work hours, overtime work hours per week, articles published in recent 5 years, experience of scientific research project, scientific research training experience), and social psychological indicators (including 3 variables: perceived social support, perceived organizational career management, and bullying at work), which were presented in Table 1.

Statistical analysis was performed using R software version 4.3.2. Continuous variables were expressed as mean (x̄) ± standard deviation (s), and categorical variables were expressed as frequencies (n) and percentages (%). We selected the most statistically significant predictors using stepwise regression based on the Akaike Information Criterion (AIC) criterion. The model used logistic regression, and the probability of psychological disturbances in nurses was predicted using a binomial distribution. We assessed multicollinearity by calculating the Variance Inflation Factor (VIF) for each variable. Generally, a VIF value less than 5 indicates a low likelihood of multicollinearity, suggesting minimal correlation between the variables and a more stable model. Based on the stepwise regression results, the nomogram model was developed, which was used to predict the probability of nurses’ psychological disturbance and score the proportion of each factor in the model. A receiver operating characteristic (ROC) curve was used to evaluate the discriminatory ability of the prediction model. The greater the area under the ROC curve (AUC), the greater the discrimination of the prediction model. An AUC greater than 0.7 indicates a reasonable estimate (19). To assess the clinical benefit of the prediction model, we used Decision Curve Analysis (DCA). DCA measures the clinical benefit of the model by calculating the standardized net benefit at different thresholds. The net benefit takes into account the costs of false positives and false negatives and visually demonstrates the model’s performance at various risk thresholds.

To further validate the stability of the model, we performed 1,000 bootstrap resamples. In each resample, we randomly selected samples with replacement from the training dataset and evaluated the AUC of the model generated from each resample. AUC was used as the performance metric, and we reported the average AUC along with its 95% confidence interval (CI). To assess the generalizability and stability of the model, we applied 10-fold cross-validation. In each fold, the dataset was randomly divided into 10 subsets, with 9 subsets used for training the model and the remaining subset used for testing. The evaluation metrics for cross-validation included AUC, sensitivity, and specificity, and we adjusted the model’s threshold to optimize the balance between sensitivity and specificity. To evaluate whether hospital-level heterogeneity affects the predictive ability of the model, we used a generalized linear mixed-effects model (GLMM), considering hospital as a random effect. There were no missing values in the predictive variables in this study. Two-tailed p < 0.05 was considered to indicate statistical significance.

A total of 3,808 nurses were included in our study, 883 (23.2%) of whom were considered to have a potential psychological disturbance. 2,471 (64.9%) nurses were married, and the average working years of nurses was 11.80 ± 8.67. There are 2,906 undergraduate nurses, accounting for 76.3%. Nurses’ income levels were mainly concentrated in 4,000–8,000 RMB (34.2%) and 8,001–12,000 RMB (37.0%). Additional information was shown in Table 1. In addition, the results showed that the factor scores of obsessive compulsive (1.53 ± 0.56), somatization (1.41 ± 0.47), and depression (1.39 ± 0.52) ranked in the top three (Table 2).

The model was constructed using the entire population sample. The initial model included 26 variables, and 10 significant predictors were ultimately retained through stepwise regression. As shown in Table 3, the independent protective factors included perceived social support, organizational career management, weekly leisure time, meal regularity, and the number of published articles. The independent risk factors included negative acts, working years, raising children, the number of patients in day shift care, and night shift work hours. The VIF for all variables were less than 2, indicating minimal correlation between them. The residual deviance of the model was 3280.5, and the AIC value was 3326.5, suggesting that the model demonstrated a good fit. Based on psychological disturbance as the outcome variable and these 10 independent predictors, a risk prediction model for nurses’ psychological disturbance was developed (Figure 1). The corresponding score for each independent predictor on the score scale at the top of the nomogram model was determined, and the scores of all the predictive factors were summed to calculate the total score. The point where the total scores corresponded to the risk line at the bottom of the nomogram model was the predicted value of the nurses’ psychological disturbance.

The AUC value for the final model was 0.803 (95% CI: 0.786–0.819), which showed that the nomogram had good discrimination and prediction abilities (Figure 2). Furthermore, the average AUC value calculated using the bootstrap approach was 0.810 (95% CI: 0.785–0.817), with a standard error of 0.0078, indicating that the model demonstrated good stability across different sample sampling conditions. Through 10-fold cross-validation, we assessed the model’s performance. Across all folds, the model’s AUC ranged from 0.749 to 0.841, with an average AUC of 0.794, suggesting that the model exhibited strong classification performance. After adjusting the threshold (from 0.5 to 0.8), the model’s sensitivity was 0.794, and its specificity was 0.784, successfully balancing the impact of false positives and false negatives. This adjustment made the model suitable for classification tasks in practical applications. The DCA results indicated good clinical benefit of the nomogram (Figure 3). The area between the model curve and the black and light gray lines indicated the net benefit of this model in predicting nurses’ psychological disturbance. Compared to the simple strategies of classifying “all nurses at risk” or “no nurses at risk,” the model showed higher clinical benefits across different thresholds, particularly within the range of 0.1 to 0.8, where the net benefit was maximized. It is recommended to select this range for clinical decision-making. Additionally, the GLMM results indicated that the variance of random effects between hospitals was minimal (variance = 0.001, standard deviation = 0.037), indicating that heterogeneity across hospitals had minimal impact on the final model, supporting the decision to pool the data from all nine hospitals.

This study developed a prediction model to evaluate the risk of psychological disturbance among nurses for the first time. The results have great implications for the ability of nursing managers to provide effective psychological support to nurses. Nursing managers can utilize this visualized nomogram prediction model to predict nurses’ risk of psychological disturbance and identify individualized risk factors. The findings can help nursing managers maintain nurses’ psychological health, and they can provide a reference for implementing preventive measures to reduce the occurrence of psychological disturbance among nurses. Nursing managers can implement effective communication, emotional support, spiritual guidance, tangible rewards, and other forms of support to help nurses manage stress and maintain a positive psychological state. By formulating relevant positive policies, nursing managers can reduce occupational stress and workplace bullying, create a professional and supportive working environment, and ultimately improve the quality of nursing work.

Limitations exist in this study. First, data from a single follow-up cannot establish a strict causal relationship between variables. To clarify causality, longitudinal data or panel data are needed for further research. Second, the data were based on self-reports, which can lead to reporting bias, as respondents might exaggerate or conceal psychological health symptoms. Future studies should consider cross-referencing self-reports with clinical records and health and social services records.