In this multicenter cross-sectional study, we used data from the first survey of Chinese nurses conducted between December 2020 to June 2021 from the Nurses' Health Cohort Study of Shandong. The cohort study mentioned above is an ongoing prospective cohort study of Chinese nurses conducted in Shandong province, China (registration number: ChiCTR2100043202). The study adopted a multistage sampling method for a more extensive sample coverage. The steps are as follows: (a) Shandong was categorized into eastern, western, southern, northern and central regions; (b) hospitals were selected by convenience sampling (at least one secondary and one tertiary hospital were selected from each region); and (c) participants were selected through cluster sampling.

Registered nurses with nurse qualification certificates who worked in shifts, and agreed to participate in this study were included in the present study. Retired nurses, refresher nurses, and student nurses; nurses who had been working for a duration of <6 months; and nurses who were on leave during the investigation were excluded from the present study.

The questionnaire consisted of the following four parts: demographic and lifestyle variables; SWSD assessment; shift work characteristics; standardized questionnaires.

Baseline information was collected through the Wenjuanxing network platform between December 2020 to June 2021. The survey included eight electronic questionnaires, each of which would take about 10–15 min to complete. Nurses were sent these questionnaires via WeChat (a widely used instant messaging app used in China) in batches to avoid impatience. Subjects who did not participate in the survey within 1 week of sending the questionnaires received a WeChat reminder to complete the survey once per week until the questionnaires were completed. Liaisons were set in every hospital involved in this study, functioning as regulators for questionnaire completion and communicators with the researchers when facing unsolved problems. Only professionals who signed nondisclosure agreements had direct access to the data. To ensure data quality, we adopted the required questions and data logic control design for the electronic questionnaires.

This study was carried out in accordance with the International Guidelines of Good Epidemiology Practice and the Declaration of Helsinki principles. The nurses' participation was voluntary, and written informed consents were obtained from all the respondents. The study was approved by the Ethics Committee of Scientific Research of Hospital with the registration number (KYLL-202011-085).

All the variables in this study, including one dependent variable and 43 independent variables, were categorical variables. Descriptive statistics included frequency and percentage for them. The distribution of characteristics between shift nurses with and without SWSD was assessed using the chi-squared test or Fisher's exact test for categorical variables. Bonferroni correction was applied for correcting the multiple test and p < 0.0012 was considered statistically significant for the univariate analysis. Statistically significant variables in the univariate analyses were all considered for inclusion in a forward and backward stepwise multivariate logistic regression analysis to identify the main factors associated with SWSD and to estimate the adjusted odds ratios (ORs) and 95% confidence intervals (CIs). The final logistic regression model with Akaike Information Criterion (AIC) minimization was used, and variables in the final model with p < 0.05 denoted statistically significant differences. Logistic regression was used because this classic classifier has been widely used in disease prediction and it is easy to interpret the results.

In order to more accurately explore the importance of factors, we adopted a machine learning method, random forest. The random forest is a tree-based ensemble classification algorithm, which has been widely used to build disease prediction models with small variable sets and high generalization ability. We use grid search for hyperparameter adjustment when training the model. The data was divided into 7:3 training data and test data. The training set was used to train and obtain the final model, and the test set was used to draw receiver-operating characteristic (ROC) curves to determine the prediction performance of the model. The random forest was implemented using Python 3.7.4 (Python Software Foundation, USA) with Scikit-learn (Packt Publishing, UK). Other statistical analyses were carried out using R software version 4.0.5 (R Development Core Team, Vienna, Austria).

The participants (n = 4,655) of the present study were recruited from nurses participating in shift work in 14 hospitals across six cities in Shandong, China. Of these, nurses who did not complete the questionnaire or those whose questionnaires contained logical inconsistencies were eliminated, leading to the exclusion of 8.17% of respondents, leaving 4,275 participants. The study participants ranged in age from 21 to 55 years and 93.2% (3,985) were women. The average age of the respondents was 31.19 ± 5.48 years, and the average working duration was 8.62 ± 5.16 years. In this study, the prevalence of SWSD among participating nurses was as high as 48.5%. Table 1 shows all the respondents' features and univariate analysis of SWSD in detail.

Table 2 shows the final forward and backward stepwise multivariate logistic regression model with adjusted OR and 95% CI of the main variables. In terms of physical and psychological factors, compared to nurses who reported a low level of fatigue, those who reported a moderate or high level of fatigue were significantly more likely to experience SWSD (OR = 1.35, 95% CI: 1.08–1.67; and OR = 2.50, 95% CI: 2.02–3.10, respectively). A high perceived stress level (OR = 4.63, 95% CI: 2.60–8.63) had a statistically significant association with SWSD. Additionally, shift nurses who experienced physical discomfort during their night shift (OR = 2.05, 95% CI: 1.74–2.43) or experienced psychological stress before or after the night shift (OR = 1.51, 95% CI: 1.14–2.01; and OR = 1.52, 95% CI: 1.30–1.77, respectively) were more likely to develop SWSD than shift nurses who did not have such experiences.

In terms of environmental and occupational factors, nurses who worked shifts of 7–9 or 10–12 months per year were significantly more likely to experience SWSD than those who worked shifts of ≤ 3 months per year (OR = 1.55, 95% CI: 1.03–2.35; and OR = 1.79, 95% CI: 1.26–2.58, respectively). Furthermore, 5–9 and >10 night shifts per month showed a detrimental effect to the sleep of shift nurses, with ORs of 1.34 (95% CI: 1.10–1.64) and 1.91 (95% CI: 1.41–2.58), respectively. Shift nurses who worked more than 40 h per week were also more likely to develop SWSD than those who worked less (>40 h per week for more than 4 weeks/month: OR = 2.36, 95% CI: 1.88–2.95). Compared to nurses who rested for 1 day after each night shift, those who took 1.5 days and ≥2 days off after each night shift were significantly less likely to experience SWSD. (OR = 0.76, 95% CI: 0.61–0.95; and OR = 0.77, 95% CI: 0.65–0.91, respectively). Nurses with ≥8-day intervals between shifts were less likely to develop SWSD than those with ≤ 4-day intervals (OR = 0.69, 95% CI: 0.52–0.92). Busyness during night shift was also associated with SWSD (OR = 1.30, 95% CI: 1.09–1.55), and shift nurses are less likely to experience SWSD when two clinical nurses were on the night shift and took turns to rest (OR = 0.69, 95% CI: 0.52–0.92).

In terms of social and behavioral factors, nurses who had a good work-family balance (OR = 0.41, 95% CI: 0.29–0.56) or good social support (OR = 0.74, 95% CI: 0.60–0.91) were less likely to experience SWSD. Nurses who responded that they could get >8 h of sleep before night shift were statistically significantly more likely to have SWSD (OR = 1.95, 95% CI: 1.08–3.55) than those who got <3 h of sleep before night shift. In addition, napping 2–3 times per week was a protective factor for SWSD (OR = 0.77, 95% CI: 0.63–0.94). Shift nurses who often used sleep medication before or after night shifts were more likely to experience SWSD than those never used sleep medication (OR = 2.40, 95% CI: 1.39–4.20; and OR = 3.25, 95% CI: 1.86–5.89, respectively). Excessive eating and irregular eating were also main factors for SWSD. Finally, high-intensity physical activity was associated with increased odds of SWSD (OR = 1.39, 95% CI: 1.17–1.64).

Before analyzing the predictors, we evaluated the performance of the two models. The results showed that the predictive efficiency of logistic regression model (sensitivity = 0.73, specificity = 0.71, accuracy = 0.72, area under the receiver operating characteristic curve [AUROC] = 0.81) was similar to that of random forest model (sensitivity = 0.72, specificity = 0.71, accuracy = 0.72, AUROC = 0.80) (Figure 1). Therefore, the model prediction performance were guaranteed in our study.

To illustrate the contribution of each factor to predictive capacity, we calculated the feature importance based on random forest. The feature importance of the top 50% of factors was visualized in Figure 2, including fatigue, physical discomfort, working overtime per week, using sleep medication, mental stress, diet, length of rest days, work-family balance, and departments. We found that the results of the two statistical methods were consistent, which can indirectly prove the robustness of the results of the logistic regression model.

The incidence of SWSD among participating nurses was 48.5%, which is higher than the rates of 37.6% (11) and 35.2% (4) reported in two studies in Norwegian nurses, and the rate of 24.4% reported by The Nurses' Sleep Health Project in Japan (10). Previous studies on the prevalence of SWSD in different populations have reported a prevalence of 23.3% in Norwegian oil rig shift workers (36), 19.9% in Chinese textile shift workers (37), 12.7% in Australian shift workers (38), and 8.1% in American shift workers from the community (39).

This discrepancy in the prevalence among the different studies could be attributable to the different shift schedules of the target populations, different national conditions or different times. Although China and Japan have similar nurse shift schedules, we collected data from 14 hospitals of different levels in areas of different economic levels, so as to more comprehensively reflect the prevalence of SWSD in nurses. Moreover, the incidence of SWSD may be higher since the time of our survey was the COVID-19 pandemic. Our result was somewhat supported by a systematic review stating that 36.36% of healthcare workers experienced work-related sleep problems during the COVID-19 pandemic (40). In addition, differences in disease definitions or questionnaire content might also reflect this divergence. In 2009, some scholars developed a self-administered questionnaire based on the ICSD-2 (36). In 2014, ICSD-3 updated the criteria for SWSD, and Waage et al. updated the questionnaire content according to the updated criteria (4). In ICSD-2, SWSD is defined as insomnia and/or excessive sleepiness that overlaps with shift work hours and cannot be explained by other diseases (7). In ICSD-3, the criteria for SWSD included a reduction in total sleep time as a symptom and increased the minimum duration of symptoms associated with shift work scheduling from 1 to 3 months (8). ICSD-3 was more rigorous in terms of duration and total sleep time was added as an indicator to make the diagnosis more sensitive, leading to significant differences between the two criteria in estimating the incidence of SWSD. Unlike most previous studies based on ICSD-2 (10, 11, 36), we used the latest questionnaire based on the ICSD-3 criteria to obtain a more rigorous and sensitive diagnosis of SWSD.

Similar to the results of previous studies, characteristics of physical and psychological factors showed a statistically significant relationship with the prevalence of SWSD among nurses; however, there were some differences in specific variables. For instance, SWSD is essentially a circadian rhythm disorder that has been associated with polymorphisms in various molecular clock genes (41). Languidity and flexibility are two characteristics of the circadian rhythm, and several previous studies have shown that fatigue may influence SWSD, while flexibility is a protective factor (19, 42, 43). Contrary to our hypothesis, we found no relationship between SWSD and circadian rhythm characteristics. One possible reason for this unexpected finding could be that approximately 70% of the circadian rhythms of shift nurses measured by the questionnaire in this study were at a moderate level, suggesting that a more objective and sensitive measurement method could help explore the relationship between SWSD and circadian rhythm characteristics in greater detail. Furthermore, in our study, nurses who reported experiencing fatigue and physical discomfort (e.g., headache, heart discomfort, gastrointestinal symptoms) during their night shifts were more likely to develop SWSD than those who did not, supporting the results reported previously (15, 23, 42, 44). Interestingly, high work stress during shift work and difficulty in falling asleep were previously found to be the typical symptoms of SWSD (45), consistent with our findings. In addition, the sensitivity of sleep to stress is a known predictor of SWSD (19, 46, 47), suggesting that reducing the mental burden of nursing work has potential benefits in reducing the incidence of SWSD. In addition, Afshari et al. reported that nurses have low levels of psychological resilience during the COVID-19 pandemic (48). Therefore, it is necessary to improve psychological resilience to relieve stress levels and indirectly reduce the impact of stress on SWSD.

Haile et al. proposed that three shifts are statistically significantly related to SWSD (49), which is different from our finding. We found that nurses on two shifts had a higher prevalence of SWSD than nurses on three shifts, although the results of the stepwise multivariate logistic regression analysis showed no statistically significant correlation between SWSD and rotation schedule, which supported the results of Vanttola et al. (19). Thus, we will continue to investigate the effect of shift schedule on SWSD with prospective follow-up data to verify the current results. An interesting finding of our study was that the number of shift years had no effect on SWSD, but more than 6 months of shift scheduling per year was associated with a higher likelihood of having SWSD. In addition, SWSD was correlated with the average number of nights per month, and more than four night shifts were detrimental to nurses' sleep, which is similar to that reported in other studies (21, 49). This suggests that in nursing management, the head nurse should schedule no more than four night shifts per month to prevent SWSD among nursing staff. Short rest periods (<11 h) between work shifts have been shown to increase the likelihood of SWSD in nurses (44, 50). This supports our finding that more than 1 day of rest (>12 h) after a night shift was a protective factor for SWSD in our study. Moreover, our study reported a novel finding that a busy night shift with one nurse can easily lead to SWSD, but a night shift where two nurses take turns to rest reduces the odds of SWSD. An interval of 8 days or more between night shifts helps nurses adapt to shifts and reduces the odds of SWSD. Working overtime (>40 h/week) for more than a week per month affected sleep. Di Milia et al. also found a statistically significant association between SWSD and weekly work hours (43).

In our study, the marital status of nurses and whether or not they had children had no association with SWSD, but a higher work-family balance or social support (i.e., family support and friend support) was associated with a lower likelihood of SWSD. Among nurses, the level of social support was statistically significantly and negatively associated with insomnia and daytime sleepiness (51), which supported our conclusions. Our results differed from those of previous studies (49), which only took into account marital status and children, possibly ignoring the more important variable of work-family balance. Afshari et al. reported that nurses have increased demands for social support and work-family balance in the context of the COVID-19 pandemic (48). Therefore, future research can focus more on the influence of psychological factors related to family and society on SWSD.

Numerous studies have repeatedly identified the use of multiple hypnotic agents (such as exogenous melatonin and injectable sleep medications) as predictors of SWSD (21, 49, 52), which is consistent with the findings of our research. We also investigated the role of sleep time before and after the night shift, rest taken during the night shift, and weekly naps. Those with SWSD had greater sleep debt on shift days and slept less compensatory sleep on days off than those without SWSD (19). Hence, our results suggest that effective sleep supplementation should include sleeping for <8 h before the night shift, with 2–3 compensatory naps per week. It is vital to raise awareness and education on sleep hygiene practices among nurses (53). We innovatively explored the effects of diet on SWSD during shifts. In our study, we found that irregular eating during the shift was a predictor of SWSD, but the results should be interpreted with caution and studies should be carried out in the future to verify our findings. Our results show that high-intensity physical activity was positively correlated with SWSD. However, a previous study contradicted our findings (54). One possible reason for the contradictory results is that we used a self-administered questionnaire to assess physical activity whereas Hajo et al. used objective ActiGraph GT3X accelerometers for data collection (54, 55). Therefore, future studies can combine self-reporting and objective monitoring to further explore the effect of physical activity on SWSD.

One limitation is the cross-sectional nature of this analysis. Further studies should collect longitudinal data to better understand the causal relationship between factors and SWSD. In addition, this study only included hospitals in Shandong Province in China; therefore, caution should be exercised when generalizing the study's findings to other regions. Our study lacks objective indicators; thus, future studies with objective measurements will help to verify the accuracy of the results of the current study.