Five computer databases (PubMed, Embase, the Cochrane Database of Systematic Reviews, CNKI, and Wanfang) were searched. The search time was from the completion of the database to June 2021. No restrictions were imposed on language and publishing status. We searched databases with search formulas and keywords included in these formulas. This search strategy aimed to collect researches on sleep problems among Chinese medical students.

The search formula was as follows:

#1: ‘Sleep problems' OR ‘Poor sleep quality' OR ‘Low sleep quality' OR ‘Sleep disorders' OR ‘Sleep disturbance' OR ‘Insomnia' OR ‘Agrypnia' OR ‘Hypersomnia' OR ‘Hypersomnolence' OR ‘Parasomnia' OR ‘Dyssomnia' OR ‘Sleep deprivation' OR ‘Somnipathy' OR ‘Obstructive sleep apnea' OR ‘OSA' OR ‘Restless legs syndrome' OR ‘RLS' OR ‘behaviorally induced insufficient sleep syndrome' OR ‘BIISS' OR ‘circadian rhythm disorder' OR ‘nightmare' OR ‘Day time sleep'

#2: ‘Chinese medical students' OR ‘medical students' AND ‘China'

#3: #1 AND #2

This review investigated studies include researches of prevalence of sleep problems among Chinese medical students. Eligible studies should meet the inclusion criteria: (1) Cross-sectional studies, (2) Date of publication before 18 June, 2021, (3) Chinese medical students in universities are assessed as the target population, (4) Reported the prevalence of sleep problems of Chinese medical students or provided sufficient data for calculating the prevalence, (5) Validated measurements were utilized as assessing tool for sleep problems, such as the Pittsburgh Sleep Quality Index (PSQI) for sleep quality assessment, eight-item Epworth Sleepiness Scale (ESS) for excessive daytime sleepiness, Athens Insomnia Scale (AIS) and Diagnostic and Statistical Manual of Mental Disorders (DSM) for insomnia and short sleep. The exclusion criteria are: (1) Intervention studies, case reports and cases comments, (2) Lack of data to calculate the prevalence of sleep problems, (3) Abstracts with no full text, (4) Studies that reported results include Chinese medical students, but without a subgroup analysis of that, (5) Studies investigating data under special circumstance, such as observation during offshore internship.

This systematic review was conducted following the registration on the prospective register of systematic reviews (PROSPERO, CRD42021237303). The filtering of studies was not performed before registration.

Two independent reviewers performed the full-text reviewing and data extraction process, and a third researcher checked the consistency. The disputes in the process were resolved through discussion with methodology experts (Figure 1). After the study was included, two reviewers independently extracted the data, including (1) the basic information of the study: author, measurements, publication year, study design, sample size, and proportion of participants; (2) information of participants: region, average age, grade, gender, education background, location of universities, and so forth; (3) outcome data: total prevalence and prevalence of different subgroups; (4) data of the latitude, gross domestic product (GDP) and GDP per capita of the cities where the universities are. The latitude and GDP data were collected from city data website (https://www.swguancha.com/).

The Agency for Healthcare Research and Quality (AHRQ) with an 11-item checklist was applied to access the methodological quality and risk of bias of the included studies. An item was scored “0” for the answer is “NO” or “UNCLEAR” and the item was scored “1” when answered “YES.” Then The study quality was evaluated with three standards: low quality = 0–3, moderate-quality = 4–7, and high quality = 8–11 (26).

Two independent researchers collected and input the data into Stata 12.0 software (www.stata.com) for analysis, including meta-analysis, subgroup analysis, and meta-regression. Sleep problems was used as the main outcome. The extracted data were used to calculate the prevalence rate and standard errors in prevalence. Data of prevalence met the condition: n × p >5 and n × (1 – p) >5. The formula for calculating the standard error (SE) was as follows:

where p is the prevalence, and n denotes the number of patients. This formula was applied to both total prevalence and prevalence between subgroups.

The prevalence of included study was collected or calculated. The combined effect was obtained by the inverse-variance analysis, and the weighted mean was calculated using the SE. The P-value and 95% confidence interval (16) were statistically analyzed. The Cochran Q test and the Higgins I² test were performed to evaluate the heterogeneity across studies (27). The fixed-effect models were used for low between-study heterogeneity (P > 0.05 and I² <50%). The random-effects models were used for significant heterogeneity (P <0.05 and I² > 50%). Jackknife analysis will be performed as sensitivity analysis to assess the robustness of the synthesized results. The following variables were used in subgroup analysis: sex, education background, grade, University type, region. The following variables were used in meta-regression: sex, education background, grade, University type, region, publication year, measurements, average age, sample size, latitude, GDP and GDP per capita. The results were assessed with forest plots. Publication bias was conducted with Egger's and Begg's tests, and the results were evaluated with bubble plots.

The statistical significance of prevalence and other continuous variables were analyzed with the Spearman correlation coefficient (CC). Statistical Product and Service Solutions (SPSS) version 26.0 was used to do the correlation analysis.

Medical students involved in the analysis were all from China. Among the included studies, five studies were conducted in South China and five studies in North China. Two concealed the geographical area of the study object due to confidentiality agreement. Two studies were published in English, and 10 studies were published in Chinese. All included studies used the lists of validated self-report questionnaires as a measurement to evaluate students. PSQI, eight-item Epworth Sleepiness Scale (28), and Diagnostic and Statistical Manual of Mental Disorders (DSM) were used for assessing sleep problems (Table 1).

The methodology and quality of all included studies were evaluated through AHRQ. The quality scores are shown in the last column of Table 1. The meta-analysis included 11 moderate-quality studies and 3 high-quality studies (Supplementary Table 1).

Fourteen studies were included (19, 29–41). Random effect was used as between-study heterogeneity is significant (tau² = 0.01, Chi² = 686.42, I² = 98%, P <0.00001). As shown in Figure 2, the prevalence of sleep problems in Chinese medical students is 27.38% (95% CI: 0.23–0.32, P <0.00001), with a range from 12.6 to 59.83%. Sensitivity analysis revealed that no study influenced the synthesized results by more than 2% (Figure 3).

The subgroup analysis on the prevalence of sleep problems in Chinese medical students showed that the differences in the prevalence of sleep problems in male (23.36%, 95% CI: 0.19–0.28, P <0.00001) and female (23.71%, 95% CI: 0.18–0.30, P <0.00001) medical students were not statistically significant (P = 0.93, I² = 0%). The prevalence in medical students in north China (33.96%, 95% CI: 0.28–0.40, P <0.00001) was higher than that in south China (24.89%, 95% CI: 0.20–0.29, P <0.00001); the difference between subgroups was significant (P = 0.01, I² = 84.0%). The prevalence in undergraduate students (28.61%, 95% CI: 0.23–0.34, P <0.00001) was higher than that in graduate students (12.83%, 95% CI: 0.10–0.16, P <0.00001); the difference between subgroups was significant (P <0.00001, I² = 95.5%). Among all undergraduate students, the prevalence was slightly higher in Grade 1 (30.88%, 95% CI: 0.17–0.45) and increased with the grade since Grade 2 (Grade 2, 28.35%, 95% CI: 0.11–0.46; Grade 3, 30.84%, 95% CI: 0.14–0.48; Grade 40.87%, 95% CI: 0.15–0.67; Grade 5, 60%, 95% CI: 0.46–0.74); the difference between subgroups was significant (P = 0.01, I² = 69%). When stratified by the type of University into Western medicine and TCM, the medical students in Western medicine universities (28.02%, 95% CI: 0.23–0.33, P <0.00001) had a higher prevalence of sleep problems compared with students in TCM universities (19.35%, 95% CI: 0.17–0.22, P <0.00001); the difference between subgroups was significant (P = 0.001, I² = 90.40%) (Table 2).

The meta-regression was performed for sensitivity analysis. The results showed that region (adj R² = 51.15%, P = 0.020), latitude (adj R² = 73.49%, P = 0.003) and GDP (adj R² = 56.07%, P = 0.037) per capita were main sources of heterogeneity of sleep problems, but not gender (P = 0.992), average age (P = 0.952), education background (P = 0.314), grade (P = 0.272), University type (P = 0.532), publication year (P = 0.053), measurements (P = 0.461), sample size (P = 0.822) and GDP (P = 0.253) (Table 3).

We analyzed the correlations between prevalence, and various continuous variables (Supplementary Table 3). The prevalence showed positive correlations with publication year (r = 0.630, P = 0.016), latitude (r = 0.929, P <0.001), and negative correlations with GDP per capita (r = −0.919, P = 0.003). Negative correlations were found between latitude and GDP per capita (r = −0.891, P = 0.007).

The publication bias of studies was access by Begg's and Egger's tests. The results indicated no obvious publication bias in studies on sleep problems (P_Begg =0.228, P_Egger = 0.724) (Table 4, Figure 4).

(1) The between-study heterogeneity in this review were significant. Except the variables already included, there may be other factors leading to heterogeneity. Several subgroups included only one study each on graduate students, fifth-grade undergraduates, and TCM students, leading to potential bias. (2) All included studies are cross-sectional designed, and therefore the longtime prevalence for a period cannot be observed. Also, a control group of healthy people was lacking in the design of each study. (3) The prevalence were assessed by self-report measurements and represent specific aspects of sleep problems. Objective and standard assessment is needed for future researches. (4) Further, the cut-off lines of PSQI in studies were not the same. There are several common criteria for judging low sleep quality with PSQI, score > 5 (54), > 6 (55), and > 7 (56, 57). The PSQI cutoff value of studies included for assessment in this review are all over 7 or 8, and that may cause underestimated prevalence compared to studies using PSQI with a cutoff of 5 or 6. (5) The evidence that meets the inclusion criteria during COVID-19 were lacking. (6) We chose GDP data in 2019 to reflects the economic level of cities, but these levels were at the same year, which may cause bias in range of years.

Hence, caution was required in interpreting the results. Besides, the AHRQ evaluation results showed that the proportion of high-quality studies included in this analysis was not high, only three high-quality studies. Hence, more attention should be paid to these issues in future research, and more hierarchical researches with subgroup analysis of region difference and education background should be conducted to investigate sleep problems more clearly and comprehensively.