Postpartum depression (PPD) is a serious public health problem that affects 10% to 20% of pregnant women [1-3]. PPD not only adversely affects the physical and mental health of mothers, it is detrimental to the growth and development of infants. In extreme cases even suicide and infanticide may occur [4]. Establishing an effective PPD prediction model that can be used in pregnancy may enable earlier identification, thus, helping health care providers offer more effective management to at-risk patients [5]. Previous studies have explored this possibility and demonstrated its feasibility [6,7].

Machine learning (ML) may be useful in making accurate predictions based on data from multiple sources and has been applied in prediction studies in recent years [8]. There are many predictive factors for PPD including demographics, psychology, and environment [5,9,10]. Assessing risk factors during pregnancy can allow enough time for subsequent interventions. The expert consultation method has often been used to generate guidelines for PPD detection, based on expert opinion and clinical experience. In contrast, ML approaches rely on the use of empirical data to generate prediction models. The key to building good ML models is in the rigorous selection of appropriate features and algorithms. There are two approaches to address the important challenge of feature selection in ML: filter and wrapper [11]. A random forest-based filter feature selection (FFS-RF) algorithm can use the importance score of a so-called random forest (RF) of variables as the evaluation criterion for feature selection, which will identify the subsets of data features that may be most relevant to accurately predict the targeted outcome variable(s) of interest. Such strategies to identify the most relevant data features have proven to be effective ways to explore the risk factors for some diseases [12]. There are two main algorithms used in depression prediction studies, namely, the support vector machine (SVM) and RF algorithms [8]. Depression prediction studies using these two methods have achieved relatively good results [13-15]. SVM is an example of supervised learning. It focuses on minimizing structural risks within the set of available data [16]. It has great advantages in solving high-dimensional modeling problems and performs well in situations that have relatively less available sample data [17]. In contrast, RF models are built using a decision tree as the basic classifier. RF approaches have high classification accuracy, strong inductive capacity, a simple parameter adjustment process, fast calculation speed, relatively low sensitivity to missing data values, and the ability to output feature importance [12,18].

Comparison between those ML methods concerning PPD has not been studied. This study is based on data drawn from a large, ongoing cohort study of pregnant women in the Hunan province of south central China. In this paper we combined the two feature selection methods and the two ML algorithms described above to assess four PPD prediction models using data during pregnancy to compare the effect of PPD prediction models, pick the optimal predictive model, and provide a reference for the development of ML in PPD.

We compared four PPD prediction models and provided a reference for the application of ML in PPD. Compared with the expert consultation method approach, the FFS-RF method identified fewer predictive factors. We found that the F-SVM model was the best model. The strongest predictive factor was the psychological resilience of pregnant women.

Between the expert consultation method and FFS-RF method, the latter selected far fewer predictive factors. Furthermore, there was no significant difference between the two methods in terms of their effects on model performance, indicating that the FFS-RS method could reduce dimensions and improve the efficiency of the algorithmic function without changing model predictive performance. The reduction in the number of predictive factors means that the burden of collecting information is reduced, making the model easier to implement and popularize, especially in busy obstetric clinics.

The SVM was chosen as the better algorithm, as it showed higher sensitivity than the RF algorithm (E-SVM=0.67, F-SVM=0.69, E-RF=0.48, F-RF=0.48). SVM had a clear advantage over RF in processing our research data, and the smaller sample size may be the main reason for this finding. Previous research on depression suggested that sample size is a key factor affecting the performance of ML models. When the sample size is small, SVM can avoid overfitting while providing efficient computing time and produces better prediction results in depression [29,30]. Our results also support this view. Therefore, we believe that when the data set is small, SVM is more practical than RF in prediction research for PPD. Several previous studies used the SVM algorithm to make PPD predictions. Jiménez [13] collected data on postpartum women from seven Spanish hospitals and used the EPDS score as the outcome indicator to train a PPD prediction model based on SVM. Sriraam [15] used social media as a data source and, based on the mental health data of 173 mothers, a SVM-based PPD prediction model was established. De Choudhury [31] developed a SVM model to identify high-risk emotions and behaviors predictive of PPD using the content of Twitter posts. As these studies either target different populations or use different methods to detect the occurrence of PPD, the model prediction effects cannot be easily compared. However, the results of the optimal F-SVM model in our study are within range (sensitivity=0.69, ROC-AUC=0.78) and consistent with the findings of previous studies (sensitivity=0.56-0.78, ROC-AUC=0.63-0.81) [13,15,31]. Due to the negative effects of PPD on mothers and infants [32,33], such as the negative effects on the physical and mental health of mothers, the closeness of the mother-infant bond, and infant development, it is important to have a model with high sensitivity while maintaining a high ROC-AUC value. The selection of indicators in evaluating depression prediction models varies across studies. For example, Sriraam [15] and De Choudhury [31] emphasized the accuracy of the model's prediction of PPD. Jiménez [13] emphasized model sensitivity and specificity. The balance between them is the geometric mean. The ROC-AUC is also widely used to evaluate the comprehensive performance of a model [14,15]. Our evaluation criteria provide a reference for prediction research for screening purposes, but the approach may be different in research studies.

We found that the top 3 most important predictors in the models were psychological resilience, depression during the third trimester, and monthly income level. First, psychological resilience is the most important factor in the prediction of PPD, which can be attributed to the protective effect of psychological elasticity. Pregnancy and childbirth are a challenging time for women emotionally and physiologically, and the mother's body and mind are under greater stress [15]. Previous research has shown that psychological resilience as an important regulatory process can enable people to recover from and adapt to stress and life events, reducing the occurrence of adverse outcomes [34-36]. Our results also support the findings of Lu [37], who found that the level of psychological elasticity was negatively correlated with the occurrence of PPD. Second, the results regarding depression in the third trimester are consistent with most previous studies. Depression in the third trimester is associated with PPD [9,38,39]. A review by Robertson [5] mentioned that “depression and anxiety during pregnancy are the strongest predictors of PPD”. Mora's [40] research suggests that depression in the third trimester may continue to develop into the postpartum period. Third, the monthly income levels remain important factors affecting PPD, which supports Rhonda's [41] findings that mothers with low income levels faced obstacles in using mental health resources and were more likely to be frustrated. Epidemiological studies of PPD worldwide have also found that the incidence in developing countries is higher than that in developed countries [42].

The identification of these predictors also reveals the different aspects of PPD risk factors. A pregnant woman's psychological elasticity may reflect her personality traits. Depression in the third trimester may be a special symptom accompanying pregnancy. The income of a pregnant woman and her partner reflects the stability and coping resources available to them. It indicates that PPD risk should be assessed based on a combination of individual long-term, short-term, and environmental characteristics.

This study has several limitations. First, there was potential selection bias. Women who were not lost to follow-up might have had a greater awareness of mental health services. Second, the 50% loss to follow-up and small sample size may have negatively affected the applicability of the PPD model, indicating that more extensive validation is required. Third, a larger number of potential predictive factors would have been useful. Further studies should develop different PPD models using other ML algorithms and data from different sources as well as incorporating additional cultural factors to expand the application of the PPD models.