Databases: PubMed, Web of Science, China Biomedical Literature Database (CBM), VIP, CNKI, Wanfang Data Knowledge Service Platform; Retrieval time: October 31, 2024. Search terms included artificial intelligence (AI), machine learning (ML), algorithm, postpartum depression (PPD), maternity blues, depression after delivery, postpartum depression, postpartum, depression, postnatal depression, post-partum depression, postnatal depression, predict, prediction, risk, risk prediction, risk score, risk assessment, model, prediction model, predictive model, prognostic model, hazard, danger, and threat. The search strategy was peer-reviewed by a medical librarian using the PRESS checklist to ensure completeness. PRISMA guidelines for a systematic review were followed. The search strategy included databases such as PubMed and Web of Science, and the search string is: (“postpartum depression” or “PPD”) AND (“machine learning” or “AI” or “predictive model”).

The study protocol has not been prospectively registered in PROSPERO or other systematic review registration platforms.

The inclusion criteria were as follows: (1) the purpose of the study was to predict and diagnose patients with PPD early, (2) use of at least one ML algorithm, (3) statistical validation of the results to evaluate the performance of the model, and (4) follow-up time of <1 year postpartum.

The exclusion criteria were as follows: (1) conference abstract, literature review, and class experiment; (2) full text not available; and (3) patients who had previously been diagnosed with PPD or had undergone PPD-related medical interventions; (4) participant level data could not be used for >20% of the variables ML performance metrics were not reported (e.g., no AUROC/ACC/F1).

Two authors reviewed the titles, descriptions, and full texts of the collected studies based on set guidelines for inclusion and exclusion. In cases of disagreement, they decided after discussion, and a third researcher decided whether to include or exclude literature that could not be ruled on. Data extraction included publication time, research objects, research types, and main research results.

The Prediction Model Risk of Bias Assessment Tool (PROBAST) was used to evaluate the risk of bias and the applicability of the included studies. Two independent reviewers conducted an evaluation, and any disagreements were resolved through discussion. If necessary, a third reviewer made a decision. We evaluated the reliability of internal evaluators and found high consistency (Cohen’s Kappa = 0.85).

The studies were merged using narrative synthesis. We reviewed the performance of the algorithm and the risk factors affecting PPD and evaluated the risk of bias in all included studies.

To minimize publication bias and capture region-specific evidence, we also searched the China National Knowledge Infrastructure (CNKI) Dissertations and Theses database, a major source of grey literature in Chinese. Master’s theses and doctoral dissertations were included given their frequent coverage of locally validated AI models in PPD screening.

A total of 129 articles were retrieved, including 16 from CNKI, 4 from the Wanfang Data Knowledge Service Platform, 1 from VIP Journal, 3 from the China Biomedical Literature Database, 67 from PubMed, 37 from Web of Science, and 1 from other sources. Finally, 11 articles were included (14–24) (Figure 1).

The articles’ basic characteristics are presented in Table 1. The studies included in this review were published between 2003 and 2023 in China (14–18), Japan (19), Israel (20), the United Kingdom (21), the United (22), and Spain (23, 24), and the sample sizes ranged from 732 to 266,544. Clear inclusion criteria were set in all 11 articles, clear exclusion criteria were set in 10 articles, and exclusion criteria were not specified in one article (15). The evaluation time points varied among the different studies, including 1 week postpartum (23), 4 weeks postpartum (19), 6 weeks postpartum (14), 4–6 weeks postpartum (17), 42 days postpartum (18), 32 weeks postpartum (24), and 1 year postpartum (20, 21). Other studies (15, 16, 22) did not indicate time points. The outcome was determined by various criteria, mainly based on Edinburgh Postnatal Depression Scale (EPDS) (14, 16–19, 23, 24), in which EPDS≥9 or EPDS≥10 were used as the threshold for determining postnatal depression. Other studies used the ICD-9/10 codes (20), PHQ-9 scale (15), PHQ-2 scale (22), and antidepressant use or non-pharmacological treatment (21) as criteria. Diagnostic interviews (DIGS) were also been used in other studies (23, 24).

The best ML algorithms and their performance indicators for the included studies are listed in Table 2. The frequency and AUROC of each ML algorithm in the included studies are listed in Table 3.

The Random Forest (RF) algorithm has demonstrated excellent performance in multiple studies. Zhong et al. (14) showed that the AUROC values of the RF algorithm in the training and test set 6 weeks after delivery reached 0.925 and 0.943, respectively, and the accuracy rates reached 0.871 and 0.903, respectively. Brier scores were 0.091 and 0.073, respectively, which were lower than those reported by Liu et al. (18). In addition, the RF algorithm was used the most frequently among all algorithms, reaching 14 times, and the AUROC value exceeded 0.9 two times, indicating its stability and high accuracy in predicting PPD and that the RF algorithm can handle large-scale data sets and evaluate the importance of each feature.

Support Vector Machine (SVM) and Logistic Regression (LR) algorithms are commonly used, which have been used 12 times and 11 times, respectively. The maximum AUROC value of the SVM algorithm in predicting depression risk at 6 weeks postpartum was 0.925 (14), which showed high efficiency in dealing with nonlinear problems. The LR algorithm has been used in several studies (14, 15, 17, 18, 22, 23). Compared to other algorithms, the LR algorithm is simple, easy to understand, and has strong interpretability. However, this requires a highly linear assumption for the data, and the prediction results may be biased when there is a nonlinear relationship in the data.

The Artificial Neural Network (ANN) algorithm, as the best algorithm in Tortajada et al. (23), has an AUROC value of 0.84 and an accuracy rate of 0.84, showing its ability to process complex data patterns. However, ANN is rarely used, possibly because of its high requirements for data volume and feature engineering. To train ANN models with good performance, a large amount of high-quality data is usually required as support. However, the acquisition and processing of medical data are often limited by ethics, privacy, resources, and other aspects, which are difficult to achieve. On the other hand, feature engineering is one of the key steps in the construction of ANN model, which involves extracting useful features from the original data to better describe and predict the target variables. However, this process is often complicated and time-consuming, and the medical explanation of the influence of the respective variables on the dependent variables in the model remains unclear.

The Extreme Gradient Boost (XGBoost) algorithm was used five times, and the AUROC value exceeded 0.8 one time (15). In a study by Liu et al. (18), the AUROC value of the training set was 0.789 and its accuracy was 0.702, indicating high specificity and sensitivity. It was also listed as the best algorithm by Hochman et al. (20), with an AUROC value of 0.712, demonstrating its advantage in handling unbalanced datasets. However, many model parameters and meticulous parameter adjustments are required to achieve the best performance.

Other algorithms such as Gradient Boosting Machines (GBM), Bayesian networks (BN), Decision Tree (DT), etc. have also been applied in various studies. However, they are used less frequently, possibly because they are not as stable as RF, SVM, and LR on specific datasets.

The observed difference in AUROC values across different outcome measures could be due to the inherent properties of each tool. The EPDS is specifically designed to assess postpartum depression and has been widely validated, which might contribute to its higher sensitivity and specificity. In contrast, PHQ-2/9 and DIGS, while also valid tools, might not be as specifically tailored to the postpartum period, potentially affecting their performance in this context. Further research is needed to explore the impact of outcome measure selection on the performance of AI-based predictive models for PPD.

Additionally, we conducted a thorough assessment of the model optimizations and validations used in each study to understand their impact on model performance (Table 4). This included hyperparameter settings, hyperparameter selections and feature selections. Furthermore, we analyzed the strategies employed to address class imbalance, such as propensity score matching (PSM), synthetic minority oversampling technique (SMOTE), and the use of class weights, as these techniques can significantly influence the model’s ability to generalize from the training data to unseen data. Model validations were rigorously carried out using methods including 10-Fold CV, random split validation, and temporal validation, ensuring a reliable assessment of model performance.

As shown in Table 5, various social demographic, psychological, obstetric, and clinical-related factors were associated with the occurrence of PPD. More than one-third of the included studies mentioned the same predictors of PPD (>3 times), and the frequency of these factors suggested a high correlation in predicting PPD (Figure 2).

As shown in Table 6, all studies demonstrated a risk of bias, and the main reason for the bias was the utilization of a retrospective design. Specifically, among the predictors, the risk of bias of the predictors in Shin et al. (22)‘s study was not clear, and the predictors in the other studies were all low-risk. Some studies (15, 16, 22) did not specify the outcome evaluation time; therefore, there was a high risk of bias. In the analysis items, some studies (14–20, 22–24) had problems, such as the lack of universality of the selected research objects or insufficient sample size. For example, Deng (16) selected pregnant women from agricultural counties inhabited by ethnic minorities in Yunnan Province, whereas Tortajada et al. (24) selected white women. In addition, Amit et al. (21) did not report scores on the PPD-related scale, so there was a high risk of bias in the overall study. Among the applicability evaluations, eight studies were rated as high risk (14–18, 20, 23, 24) and three as low risk (20–22).

Research shows that the prevalence rate of PPD is approximately 10-20% in the world. China is a populous country with a large number of pregnant women. The incidence rate of PPD in China is 15–20%, and nearly 5–7 million women suffer from depression (25, 26). Traditional screening and diagnostic methods often rely on questionnaires and clinical interviews, which are subjective, time-consuming, and laborious. The introduction of AI technology, particularly the application of ML algorithms, has provided new perspectives for the early identification and prediction of PPD. These algorithms can improve prediction accuracy by building complex models and mining predictors closely related to the occurrence of PPD from massive data. The results of this study show that existing PPD models have good performance indicators. Domestic scholars (14–16) show that the AUROC value of the RF algorithm in the training and test sets reaches the highest values of 0.925 and 0.943, respectively, and the average AUROC value exceeds 0.8. The prediction results were integrated with the classification results of multiple decision trees, and the final results were more reliable (27, 28). In addition, some domestic scholars (17, 18) also used the XGBoost and BN algorithms to rapidly process high-dimensional and multi-classification tasks, and the AUROC values exceeded 0.75, showing good prediction performance (29). These results indicate that AI technology in China can effectively and quickly extract risk factors for PPD from complex medical data and provide support for clinical decision-making.

Globally, the application of AI technology for PPD risk prediction has great potential and application prospects. Researchers in different countries and regions have begun to explore the effectiveness of ML algorithms for predicting PPD risk. These studies not only provide valuable data support for local areas, but also provide possibilities for international comparison and cooperation. ML algorithms, such as the RF and GBM algorithms, have strong adaptability in predicting PPD risk. Shin et al. (22) showed that a model based on the RF algorithm had good predictive performance in American pregnant women, with an AUROC value of 0.884. Amit et al. (21) applied the GBM algorithm with an AUROC value of 0.844, and its performance was better than that of similar algorithms used in domestic studies (18) (XGBoost algorithm). This not only provides guidance for local medical practice, but also provides a reference for other countries.

Compared with other algorithms, the Random Forest (RF) algorithm demonstrated superior performance in identifying key factors influencing postpartum depression and provided quantitative analysis, which is crucial for timely and effective interventions. Its comprehensive evaluation mechanism allows it to handle complex datasets and evaluate the importance of each feature, contributing to its high accuracy and stability in predicting PPD (30). The robustness of the RF algorithm in adapting to large-scale datasets has been widely recognized (31), making it particularly suitable for the analysis of extensive medical data commonly encountered in PPD research. Furthermore, the adjustment or setting of hyperparameters in RF, such as the number of trees and their depth, plays a critical role in optimizing model performance (32, 33). These parameters can significantly influence the model’s ability to generalize from the training data to unseen data, as evidenced by the high AUROC values reported in several studies included in this review (14, 15). Therefore, it is essential to continue strengthening the application of RF in PPD prediction and to explore its full potential through meticulous hyperparameter tuning and feature selection processes.

This study showed that maternal age was the main predictor of PPD. First-time mothers may experience self-blame and anxiety due to their lack of reproductive experience and the need to care for their newborns while recovering from the postpartum period, often experiencing small practical difficulties that are difficult to resolve quickly. These feelings can increase over time and lead to PPD (34). Stress and poor mood during pregnancy are major predictors of PPD. During pregnancy, women experience significant hormonal changes that affect emotional regulation and increase their sensitivity to stress. Furthermore, concerns about childbirth, changes in body image, and anxiety about the future role of motherhood can add to the psychological burden (35). The lack of adequate social support can exacerbate this sense of isolation, making it more difficult for pregnant women to cope with these challenges (36). This study also showed that a history of mental disorders was a predictor of PPD. Studies have speculated that PPD is mostly a continuation of prenatal psychological problems and mood disorders, and that the risk of depression increases after childbirth due to a sharp drop in hormones in the mother’s body (37). In addition, maternal education, marital relationship, sleep status, and other factors were highly correlated with PPD, which is consistent with a number of previous studies (38). It is worth noting that pre-pregnancy BMI is also an important predictor of PPD, and further studies on the mechanism of PPD should be conducted.

This systematic review aimed to synthesize AI-oriented prediction models for the risk of PPD. It not only focuses on the prediction performance and applicability of AI-oriented ML algorithms but also provides valuable insights for continuous challenges and further responses.

This study has several limitations. Owing to the limitations of the retrieved database, our findings may not cover all available evidence, and we cannot explicitly claim that these are the only relevant results. We excluded studies published in languages other than Chinese or English, which may have limited the scope and comprehensiveness of our search. Moreover, we could not quantitatively analyze the predictive performance of all AI-oriented ML algorithms because the performance indicators reported in some studies were inconsistent.

Another limitation of this study is that the study protocol failed to be prospectively registered on a public registration platform (such as PROSPERO) before the start of the study. This means that the research process lacks pre declared and publicly verifiable plans, which may increase the potential risk of selective reporting bias. We acknowledge that this is an important defect of this study, and similar studies in the future will strictly comply with the norms of prospective registration.

In addition, the variability in PPD definitions poses challenges for generalizing AI models. It highlights the need for more standardized approaches to outcome assessment and model development in future research. This could involve the use of common data collection protocols, standardized assessment tools, and agreed-upon definitions of PPD (44). By addressing these issues, future studies may be better positioned to provide more consistent and generalizable findings.