This study employed a cross-sectional study design. The field survey was carried out in two university-affiliated, governmental, tertiary hospitals in Xi’an City, Northwestern China, from July to August 2020.

The target population was the pregnant women who visited the obstetrics clinic of the two research sites for routine prenatal health assessments. The inclusion criteria for the research participants included: 1) aged ≥ 18 years; 2) cognitively independent; 3) with sufficient communication ability; 4) consent to participation. Individuals were excluded if they were diagnosed with any mental disorder or were participating in other studies. The research participants were recruited on a convenient basis, with all eligible pregnant women approached for participation invitation during the study period.

We did not perform formal sample size calculation because there are no generally acknowledged approaches to sample size estimation for studies on the construction and validation of prediction models (26). Because this is a secondary analysis based on the available dataset of a prevalence cross-sectional study, it makes sense to use the entire dataset, per guidelines from Moons (26).

The variables measured in this study included the outcome variable, namely prenatal depression, and the input variables, referring to the potential associated factors of the condition. The input variables were prespecified based on a comprehensive review of evidence regarding the associated factors of prenatal depression, which include a variety of sociodemographic, health-related, and pregnancy-related variables and several psychosocial variables. Exceptionally, anxiety was not included in the analysis of this study even though the condition was reported to be associated with prenatal depression (4). This decision was made under the consideration that anxiety and depression have shared underlying pathophysiological mechanisms, risk factors, and somatic symptoms, making it difficult to discriminate between the two conditions (27).

Upon the identification of a potentially eligible participant, the principal investigator elaborated the objectives, procedure, and significance of the study, and invited the pregnant woman to participate. Interested participants were required to provide written informed consent and, subsequently, administered with the research questionnaires. In a face-to-face manner, the principal investigator asked the questions one by one following the protocolized procedure and recorded the participants’ responses in the printed survey questionnaires. All data collection was conducted by the principal investigator, and thus, consistency across the collected data was maximized.

This study obtained ethical approval from the Ethics Committee of the Health Science Center, Xi’an Jiaotong University (reference identifier: 2020-1373) and permissions from the participating hospitals. Written informed consent was obtained from the participants before data collection. The participants’ rights and safety were protected by adhering to national and local laws/regulations, the Declaration of Helsinki, and institutional policies. Participants who scored ≥ 10 points in the PHQ-8 were referred to the psychological clinics of a local tertiary hospital for professional assessments and assistance on a willing basis.

The IBM SPSS version 22.0 and R version 4.4.0 were used for data analysis. Continuous data with normal distribution were presented as mean and standard deviation, while categorical data as count and percentage. Using the classical split-sample validation approach, the total samples were randomly assigned to the training set and the validation set at a ratio of 7:3 (26), serving as risk model (nomogram) construction and validation, respectively. The homogeneity in the sociodemographic, health-related, and pregnancy-related characteristics between sets was tested using the independent t-test, Chi-square test, or Fisher’s exact test where appropriate.

Univariate logistic regression models were employed to screen the candidate factors associated with prenatal depression, with a statistical significance level of α=0.1 (39). Subsequently, the candidate factors were entered into the multivariate binomial logistic regression models to identify the independent associated factors of prenatal depression, with a statistical significance level of α=0.05. Based on the results of multivariate binomial logistic regression models, two nomograms for the risk prediction of prenatal depression were constructed using the R package “rms”. The associated factors were ordered based on their importance using the mean decrease Gini method. Multiple parameters were employed to examine the performance of the novel prediction models. The receiver operating characteristic (ROC) curve with area under the curve (AUC) and the net classification improvement (NRI) were utilized to assess the discriminative efficiency of the prediction models. The calibration curves, which demonstrate the differences between the predicted and observed probability, the Hosmer-Lemeshow test with a 1,000-sample bootstrap, and the integrated discrimination improvement (IDI) were employed to examine the predictive efficiency of the prediction models. The decision curve analysis (DCA), which quantifies the net benefits at different threshold probabilities, was used to evaluate the clinical utility of the prediction models.

A total of 1,465 pregnant women were invited to participate, of which 1,020 (69.62%) agreed and 999 (68.19%) provided complete data. The age of the participants ranged from 21 to 44 years, with an average of 30.12 (standard deviation: 3.46). Around one-fifth and 50% of the participants had a history of gynecologic diseases and dysmenorrhea, respectively. More than half of the participants were in the third trimester and two-fifths had an unplanned pregnancy.

The participants averaged a total score of 45.88 (standard deviation: 11.97) and 60.61 (standard deviation: 13.37) in the PSRS and MSPSS, respectively. For the SCSQ, the participants had a mean score of 24.58 (standard deviation: 5.72) in the active coping dimension and 9.91 (standard deviation: 4.35) in the passive coping dimension.

Participants in the training and validation sets were homogeneous in all sociodemographic, health-related, pregnancy-related, and psychosocial characteristics. Detailed characteristics of the research participants are presented in Table 1 .

The participants had an average of 5.65 (standard deviation: 3.66) in the PHQ-8 total score. According to a cut-off point of ≥ 10, the prevalence of prenatal depression in the current study was 9.5% ( Table 1 ).

Multivariate binomial logistic regression model based on sociodemographic, health-related, and pregnancy-related variables (model I) suggested that unemployment, poor relationship with partners, the antecedent history of gynecologic diseases, unplanned pregnancy, an earlier stage of pregnancy, and more severe vomiting symptoms were associated with increased risk of prenatal depression ( Table 2 ). In the regression model that further included psychosocial indicators (model II), unemployment, the antecedent history of gynecologic diseases, unplanned pregnancy, an earlier stage of pregnancy, and a higher total score in the PSRS were found to be associated with prenatal depression ( Table 2 ).

Based on the multivariate binominal logistic regression models, two individualized nomogram models for the risk prediction of prenatal depression were constructed ( Figure 1 ), with the input variables ranked in a descending manner based on their importance ( Supplementary Figure 1 ). The nomograms assigned a specific score to each input variable, and the total score was calculated as the sum of these individual scores. The predicted risk of prenatal depression was thereafter determined based on the corresponding probability associated with the total score.

In the training set, the AUC values for model I and model II were 0.732 and 0.814, respectively. Meanwhile, the values were 0.781 and 0.787 in the validation set, respectively ( Figure 2 ). Such results indicated that both models had good discriminative efficiency, but model II was superior compared to model I. Furthermore, model II was found to have higher reclassification efficiency (NRI: 0.645 [95% CI, 0.314−0.981]; P<0.001) compared to model I, which supported the superior discriminative performance of the former.

Concerning the predictive performance, the calibration curves demonstrated a moderate level of agreement or consistency between the predicted risk and actual observation ( Figure 3 ). Besides, the coefficients of the Homser-Lemeshow test for model I and model II were 11.473 (P=0.1781) and 11.979 (P=0.1521), respectively, which supported the good predictive efficiency of both models. However, the IDI coefficient (0.120, 95% CI: 0.068−0.173; P<0.001) suggested that model II tended to have improved predictive efficiency compared to model I.

As for the clinical utility, the results of DCA suggested that both models had significant net benefits in most of the probability thresholds ( Figure 4 ), indicating good performance in guiding clinical decisions. By contrast, model II exhibited higher net benefits in the probability threshold of 0.2−0.95, compared to model I.

The prevalence of prenatal depression in the current study was 9.5%, which is significantly lower than the reported figures of approximately 20% in recent systematic reviews with meta-analyses (3, 4). The major reason for the significantly lower prevalence of prenatal depression in the current study could be its strict operational definition of depression. A cutoff point of ≥ 10 was adopted, which represents major depression or clinically significant depression (29, 40). Such a hypothesis is partly supported by a large-scale investigation that reported a prevalence of 6.1% for prenatal depression using the same criterion (40). Another explanation for the low prevalence of prenatal depression in our study might be the fact that a dominant proportion of the samples had a college degree or above, as a higher educational level has been viewed as a protective factor of the condition (4).

The result of this study showed that employed women were less likely to experience prenatal depression compared to unemployed women or housewives, with an odds ratio (OR) of 0.55 in model I and 0.54 in model II. Existing evidence tended to be consistent with the current study in supporting an association between unemployment and increased risk of prenatal depression (3, 41). Some researchers believe that unemployed women have a higher risk of prenatal anxiety and stress, which are known contributors to prenatal depression (42). Besides, unemployment might also lead to heavier economic burdens, more family conflicts, and less social interaction, which could further aggravate prenatal depression (39, 43). Participants rated a close relationship with their partner had a lower risk of prenatal depression compared to those rated a general relationship (OR: 0.25 in model I), which is consistent with the findings of previous studies (14, 44). The husband is the most important source of affective value and social support for pregnant women, and thus, a poor spousal relationship would inevitably increase the risk of prenatal depression.

It was found in this study that pregnant women who had an antecedent history of gynecologic diseases were more vulnerable to prenatal depression (OR: 1.79 in model I and 1.88 in model II). Even though relevant evidence is lacking, pregnant women may probably worry that the antecedent gynecologic diseases could result in adverse pregnancy outcomes, which could further lead to psychological burdens. Anyhow, considering the lack of direct evidence on the association between the antecedent history of gynecologic diseases and prenatal depression, a conclusion cannot be drawn and further investigations are valuable.

Accordant with existing evidence, unplanned pregnancy was found to increase the risk of prenatal depression in this study (OR: 2.21 in model I and 2.38 in model II) (3, 45, 46). Human beings find it more difficult to cope with unexpected and undesired events (47). Similarly, women could be more likely to experience psychological problems in the face of an unplanned pregnancy. Compared to those in the first trimester, participants in the second and third trimester were less likely to have depression in this study. Even though subgroup analysis of a systematic review suggested that the pooled prevalence of prenatal depression was the highest in studies that recruited pregnant women in the third trimester (3), updated longitudinal and large-scale multicenter cross-sectional evidence has supported the findings of the current study (16, 17). To further clarify the relationship between the stage of pregnancy and prenatal depression, more empirical studies are encouraged. The decreasing tendency in the risk of prenatal depression during pregnancy could be interpreted as the result of a gradual adaption to the stressful event of pregnancy. The results of multivariate binomial logistic regression for model I indicated that pregnant women who had moderate to severe vomiting symptoms were 2.26 times more likely to suffer from prenatal depression compared to those who had no or minimal vomiting symptoms. Considering the bi-directional mechanisms between gastrointestinal and psychological symptoms, it is reasonable to accept such an association (19).

This study suggested that every increased point in the PSRS would increase the risk of prenatal depression by 5.48 times, defining prenatal stress as a predominant contributor to the condition. The close association between stress and depression has been well-documented, not only in pregnant women but also in the general population (48, 49). Stress could have profound effects on the development of depression either in a direct pathway via epigenetic changes or in an indirect pathway mediated by multiple variables, for example, anxiety symptoms (48–50).

A variety of methods and parameters were used to assess the validity and performance of the novelly constructed nomograms for the risk prediction of prenatal depression. Several parameters suggested model II is superior considering its better performance in discriminative efficiency, predictive efficiency, and clinical utility. Moreover, it could be used for the precise risk estimation of individuals with extremely high risk of prenatal depression (0.6−0.95, as is shown in Figure 1 ). However, it does not necessarily indicate that model I is not valid or meaningful. The model I reached a satisfactory level in all the tested parameters. Besides, it could be advantageous when used as a quick screening tool as healthcare professionals will not spend time on stress assessment, which could be time-consuming in the busy clinical setting.

After a comprehensive search of relevant evidence, only one recently proposed nomogram for the risk prediction of prenatal depression was identified (23). The referenced model included nine input variables, namely employment, planned pregnancy, pregnancy number, conception methods, gestational diabetes mellitus, twin pregnancy, placenta previa, umbilical cord encirclement, and educational attainment. Despite its large sample size and equivalent discriminative efficiency compared to the current study, it only included pregnant women in the third trimester, which could limit its generalizability. Moreover, some other important candidate variables were neglected in the risk prediction model construction, such as relationships with family members and the history of gynecologic diseases.

This study has several remarkable strengths. Foremostly, it is among the very limited studies that developed monograms for the precise individualized risk prediction of prenatal depression, which contributed innovative evidence to the knowledge body. Moreover, several previously underestimated factors were examined for the association with prenatal depression. Besides, multiple parameters were employed to assess the performance of the nomograms, the agreed positive results improved the validity of the risk prediction models.

Despite its strengths, the findings of this study should be interpreted with the consideration of the following limitations. First, the research participants were recruited from two research sites with a convenience sampling approach, which could introduce selection bias to the study. Second, formal sample size estimation was not performed in this study, and therefore, there is the risk that the sample size is not adequate enough. Third, participants in the training and validation sets were from the same population, making the external validity of the risk prediction models uncertain. Fourth, a large proportion of the variables were measured in a subjective manner, which might lead to reporting bias. Besides, due to the nature of a cross-sectional design, the causal relationship between the associated factors and prenatal depression was not guaranteed.

In view of the relatively high prevalence of prenatal depression among pregnant women and its burdensome negative consequences, regular assessment and early preventive strategies should be included in the routine care of obstetrics clinics. The nomograms constructed in the current study could be made use of for the early detection of pregnant women highly vulnerable to prenatal depression, ideally in an auto-generated approach making use of the health data in the hospital information systems.

To examine the external validity of the novelly constructed nomograms for the risk prediction of prenatal depression, large-scale, multi-centered studies originating from diverse populations are necessary. Due to the very limited evidence regarding the association between the antecedent history of gynecologic diseases and prenatal depression, and the inconsistent findings on the association between trimester and prenatal depression, further studies examining such associations are valuable. In addition, longitudinal studies are desirable to confirm the causal relationship between the candidate risk factors and prenatal depression.