Review article  评论文章

Application of machine learning in predicting the risk of postpartum depression: A systematic review
机器学习在预测产后抑郁风险中的应用：系统评价

We searched eight health science databases: PubMed, EMBASE, Web of Science, the Cochrane Library, China National Knowledge Infrastructure (CNKI), the Wanfang database, the China Science and Technology Journal Database (VIP), and the China Biology Medicine disc (CBM). Searches aimed to identify studies involving the use of ML-based models to predict PPD risk from the date of inception to March 2022. Additionally, we reviewed the reference lists of each study to find more relevant studies. The search strategy was formed using Medical Subject Headings (MeSH) terms and free words, as follows: (“Depression, Postpartum” OR “Postpartum Depression” OR “Post-partum Depression” OR “Postnatal depression” OR “Post-natal Depression”) AND (“Prediction Model” OR “Risk Model” OR “Risk Prediction Model” OR “Risk Stratification Model”) AND (“Machine Learning” OR “Logistic Regression” OR “Decision Trees” OR “Bayesian Network” OR “Naive Bayes” OR “K Nearest Neighbor” OR “Random Forest” OR “Support Vector Machine” OR “Artificial Neural Network” OR “AdaBoost” OR “XGBoost” OR “Gradient Boosting Machine” OR “Deep learning”).
我们检索了八个健康科学数据库：PubMed、EMBASE、Web of Science、Cochrane Library、中国知网（China National Knowledge Infrastructure， CNKI）、万方数据库、中国科技期刊数据库（China Science and Technology Journal Database， VIP）和中国生物医学光盘（China Biology Medicine disc， CBM）。检索旨在确定涉及使用基于 ML 的模型预测从建库之日到 2022 年 3 月的 PPD 风险的研究。此外，我们回顾了每项研究的参考文献列表，以找到更相关的研究。检索策略使用医学主题词 （MeSH） 检索词和自由词形成，如下所示：（“抑郁症，产后” OR “产后抑郁症” OR “产后抑郁症” OR “产后抑郁症” OR “产后抑郁症”） AND （“预测模型” OR “风险模型” OR “风险预测模型” OR “风险分层模型”） AND （“机器学习” OR “逻辑回归” OR “决策树” OR “贝叶斯网络” OR “朴素贝叶斯” OR “K 最近邻” OR “随机森林” OR “支持”向量机“或”人工神经网络“或”AdaBoost“或”XGBoost“或”梯度提升机“或”深度学习”）。

The studies included were those that (1) aimed to identify patients having symptoms of or at the risk of PPD, (2) developed or validated models for predicting PPD based on ML techniques that could be found in the Scikit Learn package or were declared by the authors to belong to the ML domain, (3) reported the performance of the proposed ML-based models, and (4) original studies published through a peer-reviewed process in English and Chinese. We excluded studies that (1) lacked available full text, (2) were conference abstracts, reviews, systematic reviews, or meta-analyses, and (3) involved animal experiments.
纳入的研究是 （1） 旨在识别有 PPD 症状或有 PPD 风险的患者，（2） 开发或验证了基于 ML 技术的预测 PPD 的模型，这些模型可以在 Scikit Learn 包中找到或被作者声明属于 ML 领域，（3） 报告了拟议的基于 ML 的模型的性能， （4） 通过同行评审过程以英文和中文发表的原始研究。我们排除了以下研究：（1）缺乏可用的全文，（2）是会议摘要、综述、系统评价或 meta 分析，以及（3）涉及动物实验。

The initial citations and records from the aforementioned databases were imported into the EndNote 20 software. All duplicate studies were then identified and removed by the software. Titles and abstracts of the resulting articles were independently screened by two investigators based on the eligibility criteria. Any discrepancies regarding study inclusion were resolved through discussion or by referral to a third investigator.
将上述数据库中的初始引文和记录导入 EndNote 20 软件。然后，软件会识别并删除所有重复的研究。结果文章的标题和摘要由两名研究者根据资格标准独立筛选。有关研究纳入的任何差异均通过讨论或转介给第三位研究者来解决。

Data were extracted from all identified studies using the check list for critical appraisal and data extraction for systematic reviews of prediction modeling studies (CHARMS) (Moons et al., 2014). Variables included study, country, language, participants (data source, sample size, and eligible criteria), outcome (classification methods and measurement time), ML techniques, final predictors, data processing (numeric and categorical variables and missing values), data splitting (training, validation, and testing sets), model optimization (feature selection and hyperparameter value selection), model validation status (internal and/or external), model translation, and model performance metrics. Two investigators extracted the data independently and discrepancies were resolved through discussion, with the help of a third investigator, if necessary.
使用预测建模研究 （CHARMS） 系统评价的严格评估和数据提取清单从所有已确定的研究中提取数据（Moons et al.， 2014）。变量包括研究、国家、语言、参与者（数据源、样本量和合格标准）、结果（分类方法和测量时间）、ML 技术、最终预测因子、数据处理（数字和分类变量以及缺失值）、数据拆分（训练、验证和测试集）、模型优化（特征选择和超参数值选择）、模型验证状态（内部和/或外部）、模型翻译、 和模型性能指标。两名研究者独立提取数据，并在必要时在第三名研究者的帮助下通过讨论解决差异。

PROBAST was used to assess the risk of bias of the predictive models presented in the studies (Moons et al., 2019). This tool comprises 20 signaling questions across four domains: (1) participants, (2) predictors, (3) outcome, and (4) analysis. Each domain is rated as “low,” “high,” or “unclear” for the risk of bias. The ratings of the four domains resulted in an overall determination of the risk of bias. Low overall risk of bias was assigned if each domain had low scores. High overall risk of bias was assigned if at least one domain was determined to have a high risk of bias. Unclear overall risk of bias was determined if at least one domain was considered unclear and all other domains had low risk of bias. Similarly, two investigators assessed the quality of the studies involved, and discrepancies were resolved through discussion or through referral to a third investigator.
PROBAST 用于评估研究中提出的预测模型的偏倚风险（Moons et al.， 2019）。该工具包含四个领域的 20 个信号问题：（1） 参与者，（2） 预测变量，（3） 结果和 （4） 分析。每个领域的偏倚风险都被评为 “低”、“ 高 ”或 “不清楚”。四个领域的评级导致了偏倚风险的总体确定。如果每个领域的分数都较低，则分配较低的总体偏倚风险。如果确定至少一个领域具有高偏倚风险，则分配高总体偏倚风险。如果至少有一个领域被认为不明确，而所有其他领域的偏倚风险都较低，则确定不明确的总体偏倚风险。同样，两名研究者评估了所涉研究的质量，并通过讨论或转介给第三名研究者来解决差异。

To synthesize the studies included, a descriptive analysis was performed in the form of a narrative report. It is important to note that quantitative analyses, such as meta-analysis, could not be performed owing to the significantly inconsistent performance metrics reported across the studies.
为了综合纳入的研究，以叙述性报告的形式进行了描述性分析。需要注意的是，由于研究中报告的实施指标明显不一致，因此无法进行定量分析 ，例如荟萃分析。

Discrimination and calibration were two crucial aspects to consider when evaluating the prediction performance of ML-based models. The AUROC score can be regarded as a discrimination index of models for predicting dichotomous outcomes. The median AUROC score in five studies focusing on the internal validation of ML-based models was 0.78 (range: 0.763–0.833), and the best AUROC score of a single study on the external validation of ML-based models was 0.84 (SVM) for six-week postpartum depression. In the case of one-year postpartum depression, the median AUROC score was 0.884 (range: 0.79–0.952) in three studies involving the internal validation of ML-based models and 0.844 (range: 0.712–0.89) in three studies involving the external validation of such models. The Brier score is an effective calibration index for ML-based prediction models, and it was reported only by Zhang et al. (2021) at a score of 0.16 (reference value: 0–0.25) for the LR-based ML model for predicting the risk of PPD within a one-year postpartum period.
在评估基于 ML 的模型的预测性能时，判别和校准是需要考虑的两个关键方面。AUROC 评分可被视为预测二分类结果的模型的鉴别指数。五项专注于基于 ML 的模型内部验证的研究的 AUROC 评分中位数为 0.78 （范围：0.763-0.833），而一项关于基于 ML 的模型外部验证的研究的最佳 AUROC 评分为 0.84 （SVM） 产后 6 周抑郁症。在产后一年抑郁症的情况下，在涉及基于 ML 的模型内部验证的三项研究中，AUROC 评分中位数为 0.884（范围：0.79-0.952），在涉及此类模型外部验证的三项研究中，AUROC 评分中位数为 0.844（范围：0.712-0.89）。Brier 评分是基于 ML 的预测模型的有效校准指标，仅由 Zhang 等人（2021 年） 报道，基于 LR 的 ML 模型预测产后一年内 PPD 风险的评分为 0.16（参考值：0-0.25）。

We identified 62 ML-based models across 17 studies to predict PPD risk. We then systematically reviewed the development and performance of the ML-based models, after which we assessed the bias and applicability risks associated with all the studies. Research interests in ML techniques for predicting PPD risk have grown significantly, with a substantial increase in the number of publications over the past three years. In addition to predicting the maternal risks of PPD, paternal risks of PPD are gradually being considered. Current research generally focuses on PPD risk within six weeks of or one year after the postpartum period. The ML-based models for predicting PPD risk are mainly constructed using supervised learning algorithms, with multiple predictors involving demographic, psychosocial, obstetric, clinical, and biological variables. Most of the studies were based on the development and validation of ML-based models associated with the prediction PPD risk. A few studies focus on the improvement of ML-based models and the innovation of ML algorithms. The best-performing ML-based models were different in different PPD measurement time, such as SVM and FGB for predicting PPD risk at 6 weeks and 1 year postpartum, respectively. Approximately 1/3 of the ML-based models achieved AUROC scores of <0.8, thereby indicating that the prediction performance of existing models was mediocre. Additionally, few studies translated ML-based models into clinical assessment tools. All the studies demonstrated a high risk of bias and applicability risk owing to non-comprehensive development and validation processes or results reports. Therefore, the current studies demonstrated certain potential and feasibility in predicting PPD risk based on ML techniques. However, there is a long way to go before the research results are translated into implementations that can be applied in clinical practice.
我们在 17 项研究中确定了 62 个基于 ML 的模型来预测 PPD 风险。然后，我们系统地回顾了基于 ML 的模型的开发和性能，之后我们评估了与所有研究相关的偏倚和适用性风险。对用于预测 PPD 风险的 ML 技术的研究兴趣显着增长，在过去三年中出版物数量大幅增加。除了预测产妇患 PPD 的风险外，人们还逐渐考虑父亲患 PPD 的风险。目前的研究通常侧重于产后 6 周内或产后 1 年内的 PPD 风险。用于预测 PPD 风险的基于 ML 的模型主要使用监督学习算法构建，具有多个预测因子，涉及人口统计学、社会心理、产科、临床和生物变量。大多数研究基于与预测 PPD 风险相关的基于 ML 的模型的开发和验证。一些研究侧重于基于 ML 的模型的改进和 ML 算法的创新。表现最佳的基于 ML 的模型在不同的 PPD 测量时间内存在差异，例如预测产后 6 周和 1 年 PPD 风险的 SVM 和 FGB。大约 1/3 的基于 ML 的模型实现了 <0.8 的 AUROC 分数，这表明现有模型的预测性能一般。此外，很少有研究将基于 ML 的模型转化为临床评估工具 。所有研究均表明，由于开发和验证过程或结果报告不全面，存在较高的偏倚风险和适用性风险。因此，目前的研究表明，基于 ML 技术预测 PPD 风险具有一定的潜力和可行性。 然而，在将研究结果转化为可应用于临床实践的实施之前，还有很长的路要走。

According to the summary of ML-based models, the most commonly used ML algorithms for predicting PPD risk are those based on SVM, RF, and LR. The SVM is a supervised ML algorithm or approach that uses minimal input data to define the hyperplane that maximizes the margins between two classes (Jimenez-Serrano et al., 2015). It is suitable for handling multiple features and small-sample problems, and has the advantages of anti-noise, high learning efficiency, and good generalization (Zhang et al., 2020). RF is an ensemble classifier that combines the prediction of multiple decision trees and outputs the class voted by a majority of the trees (Xiao et al., 2020). It is also suitable for high-dimensional feature problems and has the advantages of low sensitivity to missing data, no need for feature selection, simple parameter adjustment process, fast calculation speed, and feature importance output (Zhang et al., 2020). LR is a type of generalized linear regression model, which is suitable for low-dimensional feature sampling and is characterized by fast calculation speed, simple form, and high interpretability (Jimenez-Serrano et al., 2015; S. Wang et al., 2019). However, currently, there remains no absolute statement regarding the type of ML-based model with the best performance. This may be due to the applicability of different ML techniques in different research samples. For example, in a comparative study on SVM and RF-based ML models for predicting PPD risk, those constructed using SVM demonstrated enhanced levels of sensitivity. One possible explanation for this result is that SVM-based ML models can avoid overfitting in a small sample compared to those based on RF (Zhang et al., 2020).
根据基于 ML 的模型的总结，预测 PPD 风险最常用的 ML 算法是基于 SVM、RF 和 LR 的算法。SVM 是一种监督式 ML 算法或方法，它使用最少的输入数据来定义最大化两个类之间边距的超平面（Jimenez-Serrano et al.， 2015）。适用于处理多特征和小样本问题，具有抗噪声、学习效率高、泛化性好等优点（Zhang et al.， 2020）。RF 是一个集成分类器，它结合了多个决策树的预测，并输出由大多数决策树投票的类（Xiao et al.， 2020）。它也适用于高维特征问题，具有对缺失数据敏感度低、无需特征选择、参数调整过程简单、计算速度快、特征重要性输出等优点（Zhang et al.， 2020）。LR 是一种广义线性回归模型，适用于低维特征采样，具有计算速度快、形式简单、可解释性高等特点（Jimenez-Serrano et al.， 2015;S. Wang et al.， 2019）。但是，目前，对于具有最佳性能的基于 ML 的模型类型，仍然没有绝对的声明。这可能是由于不同的 ML 技术在不同的研究样本中的适用性。例如，在对用于预测 PPD 风险的 SVM 和基于 RF 的 ML 模型的比较研究中，使用 SVM 构建的模型表现出更高的灵敏度水平。 对这一结果的一种可能的解释是，与基于 RF 的 ML 模型相比，基于 SVM 的 ML 模型可以避免在小样本中过度拟合（Zhang et al.， 2020）。

Additionally, several studies tried to improve or innovate ML techniques for predicting PPD risk. Park et al. constructed PPD risk prediction model using a pre-processing method called reweighing that applied different weights to each group-label combination according to conditional probability of label by race, resulting in a greater reduction in algorithmic bias for PPD between White and Black individuals than simply excluding race from the prediction models (Park et al., 2021). Natarajan et al. adapted an innovation technique FGB that can handle class imbalance in a principled manner, where the gradients are computed as the difference between observed and predicted probabilities of each example and a new regression tree is fitted to these examples at every iteration. This study showed that FGB gained a good performance, with the AUROC, recall and precession of 0.952, 0.84 and 0.92, which was outperform the baseline classifiers (Natarajan et al., 2017).
此外，一些研究试图改进或创新用于预测 PPD 风险的 ML 技术。Park 等人使用一种称为重新加权的预处理方法构建了 PPD 风险预测模型，该方法根据种族标签的条件概率对每个组标签组合应用不同的权重，从而比简单地从预测模型中排除种族更大地减少了白人和黑人个体之间 PPD 的算法偏差（Park 等人， Natarajan 等人采用了一种创新技术 FGB，该技术可以以有原则的方式处理类不平衡，其中梯度计算为每个样本的观察概率和预测概率之间的差异，并在每次迭代时为这些样本拟合新的回归树。这项研究表明，FGB 获得了良好的性能，AUROC、召回率和进动分别为 0.952、0.84 和 0.92，优于基线分类器（Natarajan et al.， 2017）。

Demographic features were the most commonly included predictors, with the highest inclusion frequency being that for maternal age. The predictive value of age on PPD is polarized, whereby women younger than 19 or over 35 years are at greater risk for PPD (Amit et al., 2021; Shin et al., 2020).
人口统计学特征是最常见的预测因素，其中最高的纳入频率是产妇年龄 。年龄对 PPD 的预测价值是两极分化的，即 19 岁以下或 35 岁以上的女性患 PPD 的风险更大（Amit 等人，2021 年;Shin et al.， 2020）。

Psychosocial predictors appear to be the most significant predictors for PPD. EHRs or background data combined with the scores of self-reported psychometric scales could increase the AUROC or sensitivity scores during the screening process (Amit et al., 2021; Andersson et al., 2021). The models' performances decreased when the women were stratified based on their history of depression or other psychiatric disorders before pregnancy (Andersson et al., 2021; Hochman et al., 2021). Although Payne et al. demonstrated that the epigenetic PPD biomarker contained in model was highly accurate among women with and those without a psychiatric history, they also emphasized that the antenatal depression status was also crucial in generating accurate predictions of PPD (Payne et al., 2020).
社会心理预测因子似乎是 PPD 最重要的预测因子。EHR 或背景数据与自我报告的心理测量量表的分数相结合， 可能会在筛选过程中增加 AUROC 或敏感性分数（Amit 等人，2021 年;Andersson 等人，2021 年）。当根据女性怀孕前的抑郁症或其他精神疾病史对女性进行分层时，模型的表现会降低（Andersson 等人，2021 年;Hochman et al.， 2021）。尽管 Payne 等人证明模型中包含的表观遗传 PPD 生物标志物在有和没有精神病史的女性中高度准确，但他们也强调产前抑郁状态对于生成 PPD 的准确预测也至关重要（Payne 等人，2020 年）。

Obstetric predictors related to PPD risk include the gender of newborns, delivery mode, parity, and gestational age at delivery. The association between the gender of newborns and the risk of PPD varies across countries, such as the preference for male infants in Japan (Mori et al., 2018). As reported in previous studies, the role of parity and delivery mode in predicting PPD risk remains controversial (Baba et al., 2021; Deng et al., 2021). Women who give birth before the gestational period is complete are at a higher risk of PPD (Amit et al., 2021).
与 PPD 风险相关的产科预测因素包括新生儿的性别、分娩方式、胎次和分娩胎龄。新生儿性别与 PPD 风险之间的关联因国家而异，例如日本偏爱男婴（Mori 等人，2018 年）。正如以前的研究所报道的那样，胎次和交付方式在预测 PPD 风险中的作用仍然存在争议（Baba 等人，2021 年; 邓等人，2021 年）。在妊娠期完成之前分娩的妇女患 PPD 的风险更高（Amit 等人，2021 年）。

Clinical predictors associated with PPD comprise physical signs, medication use and disease diagnoses, such as pre-pregnancy BMI, thyroid dysfunction, and the use of antidepressants. Previous studies show that disease diagnoses (vs. demographic or medication) demonstrated the best performance (AUROC = 0.72) when it came to predicting PPD risk and that combining disease diagnoses with medication improved the prediction PPD risk (AUROC = 0.76) compared to using disease diagnoses alone (Shin et al., 2020).
与 PPD 相关的临床预测因素包括体征、药物使用和疾病诊断，例如孕前 BMI、 甲状腺功能障碍和抗抑郁药的使用。先前的研究表明，疾病诊断（与人口统计或药物相比）在预测 PPD 风险方面表现出最佳性能 （AUROC = 0.72），并且与单独使用疾病诊断相比，将疾病诊断与药物相结合提高了 PPD 风险的预测 （AUROC = 0.76）（Shin 等人，2020 年）。

Biological predictors include laboratory indicators and epigenetic biomarkers obtained from patients' blood samples, and the latter seem to predict PPD risk better when included in the model. For instance, Hochman et al. considered albumin, ferritin, hematocrit, hemoglobin, and creatinine levels during pregnancy when predicting the risk of PPD, obtained the AUROC, sensitivity and specificity of 0.712, 0.349 and 0.905 at the 90th percentile risk threshold(Hochman et al., 2021). While Payne et al. showed that the evaluation of DNA methylation in HP1BP3 and TTC9B resulted in an accurate prediction of future EPDS scores among patients, with an AUROC of 0.84 (Payne et al., 2020). Tortajada et al. selected combination genotypes (5-HTT-GC) as predictors of PPD risk, obtained the AUROC, sensitivity and specificity of 0.84, 0.78 and 0.85 (Tortajada et al., 2009).
生物预测因子包括从患者血液样本中获得的实验室指标和表观遗传生物标志物，后者在包含在模型中时似乎可以更好地预测 PPD 风险。例如，Hochman 等人在预测 PPD 风险时考虑了怀孕期间的白蛋白、 铁蛋白 、血细胞比容、血红蛋白和肌酐水平，在第 90 个百分位风险阈值处获得了 AUROC、0.349 和 0.905（Hochman 等人，2021 年）。而 Payne 等人表明，对 HP1BP3 和 TTC9B 中 DNA 甲基化的评估导致对患者未来 EPDS 评分的准确预测，AUROC 为 0.84（Payne 等人，2020 年）。Tortajada 等人选择组合基因型 （5-HTT-GC） 作为 PPD 风险的预测因子，获得了 AUROC，敏感性和特异性为 0.84、0.78 和 0.85 （Tortajada et al.， 2009）。

Additionally, unstructured data features also can be used to predict PPD risk, but the prediction effect varies greatly, which may be related to the predicted PPD risk population. For examples, Gabrieli et al. used acoustic features of infants' cries (fundamental frequency, first four formants, and intensity) and demographic information (infants' gender, mothers' age) to predict the risk of PPD, obtained the AUROC, accuracy, recall and precision of 0.969, 0.895, 0.88 and 0.904 (Gabrieli et al., 2020). Fatima et al. employed 20 linguistic features of women's texts posted on Reddit to predict the risk of PPD, obtained the accuracy, recall and precision of 0.869, 0.868 and 0.87, and believed that the linguistic feature “family” played an important predictive role (Fatima et al., 2019). Shatte et al. evaluated the risk of PPD using postpartum changes in fathers' depression symptoms, including behavior, discussion topics, linguistic styles, and emotion. The results showed that demonstrated that the high-risk fathers exhibited lower behavior engagement with the platform, and engaged in fewer discussions categorized as Humor, Lifestyle, and Image related; the low-risk fathers have higher anger and use of swear words across both the prepartum and postpartum periods than the high-risk group. However, the performance metrics were not excellent, with the accuracy, recall and precision of 0.66, 0.68 and 0.67, which might be due to the poor applicability of the selected predictors because they were adapted from mothers for fathers, or neglecting fathers in possible high risk who may only passively observe discussions without actively contributing (Shatte et al., 2020).
此外，非结构化数据特征也可用于预测 PPD 风险，但预测效果差异很大，这可能与预测的 PPD 风险人群有关。例如，Gabrieli 等人使用婴儿哭声的声学特征（基频、前四个共振峰和强度）和人口统计信息（婴儿的性别、母亲的年龄）来预测 PPD 的风险，获得了 AUROC、准确率、召回率和精密度为 0.969、0.895、0.88 和 0.904（Gabrieli 等人，2020 年).Fatima et al. 利用 Reddit 上发布的 20 种女性文本的语言特征来预测 PPD 的风险，获得了 0.869、0.868 和 0.87 的准确率、召回率和精密度，并认为语言特征“家庭”起着重要的预测作用（Fatima et al.， 2019）。Shatte 等人使用父亲抑郁症状的产后变化（包括行为、讨论话题、语言风格和情绪）评估了 PPD 的风险。结果表明，这表明高危父亲对平台的行为参与度较低，并且参与的讨论较少，归类为幽默、生活方式和图像相关;与高危组相比，低风险父亲在产前和产后期间的愤怒和使用脏话的几率更高。然而，性能指标并不出色，准确率、召回率和精密度分别为 0.66、0.68 和 0.67，这可能是由于所选预测变量的适用性差，因为它们是从母亲改编为父亲的，或者忽略了可能处于高风险中的父亲，他们可能只是被动地观察讨论，而没有积极贡献（Shatte 等人， 2020 年）。

Except for Study Fatima et al. that did not define the PPD measurement time (Fatima et al., 2019), we found that the other included studies mainly measured PPD in 6 periods, and the most commonly measurement time were 6-week postpartum and 1-year postpartum. This is an obvious heterogeneity among included studies, leading to the difference of methods used in studies and the difference of results obtained from the studies with different measurement time. Regarding methods, we observed that most of studies that measured PPD risk at 6-week postpartum modeled using prospective clinical study data (Andersson et al., 2021; Fang, 2019; Payne et al., 2020; Xiao et al., 2020; Zhang et al., 2020); while most of studies that measured PPD risk during 1-year postpartum modeled based on retrospective data or database (Amit et al., 2021; Hochman et al., 2021; Shin et al., 2020; S. Wang et al., 2019; Zhang et al., 2021). This phenomenon might be explained by the fact that mothers are usually required to return to hospital at 6-week postpartum to check the health status of themselves and their infants, providing the chance for research to assessing maternal depression symptoms directly. Also, it is hard for research implement prospective monitoring of PPD during 1-year postpartum due to the limitation of manpower, energy and money. Regarding results, we found that the commonly used ML techniques of PPD risk at 6-week postpartum and during 1-year postpartum was SVM (Payne et al., 2020; Shatte et al., 2020; Zhang et al., 2020) and algorithms based on gradient- boosting (Amit et al., 2021; Hochman et al., 2021; Natarajan et al., 2017), respectively. This result was likely related to the data complexity of prospective studies and retrospective databases. As discussed above, SVM is suitable for multiple features and small-sample problems while algorithms based on gradient-boosting are suitable for class imbalance and larger sample.
除了 Fatima et al. 的研究没有定义 PPD 测量时间 （Fatima et al.， 2019） 外，我们发现其他纳入的研究主要测量 6 个时期的 PPD，最常见的测量时间为产后 6 周和产后 1 年。这是纳入研究之间的明显异质性，导致研究中使用的方法不同，以及从不同测量时间的研究中获得的结果不同。关于方法，我们观察到大多数测量产后 6 周 PPD 风险的研究都是使用前瞻性临床研究数据建模的（Andersson 等人，2021 年;Fang， 2019;Payne et al.， 2020;Xiao et al.， 2020;Zhang et al.， 2020）;而大多数测量产后 1 年 PPD 风险的研究都是基于回顾性数据或数据库建模的（Amit et al.， 2021;Hochman 等人，2021 年;Shin et al.， 2020;S. Wang et al.， 2019;Zhang et al.， 2021）。这种现象可能是因为母亲通常需要在产后 6 周返回医院检查自己和婴儿的健康状况，从而为直接评估母亲抑郁症状的研究提供了机会。此外，由于人力、精力和金钱的限制，研究很难在产后 1 年内实施 PPD 的前瞻性监测。 关于结果，我们发现产后 6 周和产后 1 年 PPD 风险的常用 ML 技术是 SVM（Payne 等人，2020 年;Shatte 等人，2020 年;Zhang et al.， 2020）和基于梯度提升的算法（Amit et al.， 2021;Hochman 等人，2021 年;Natarajan et al.， 2017） 分别。这一结果可能与前瞻性研究和回顾性数据库的数据复杂性有关。如上所述，SVM 适用于多特征和小样本问题，而基于梯度提升的算法适用于类不平衡和较大样本。

Multiple studies on predicting PPD risk have been published in recent years. However, the total numbers of the studies published remain insignificant and their overall quality is not adequate. Additionally, the performance of the models proposed in these studies is uneven, and as a result, it cannot be directly applied in clinical practice. This indicates that ML has gained tremendous interest over the recent years, but it remains a relatively novel field. One possible reason is that the PROBAST was published in 2019 (Moons et al., 2019), and the standardized guidelines on the conduct of ML studies was proposed in 2020 (Stevens et al., 2020), thereby resulting in the overall lack of a method-based reference for some studies that were conducted before the publication of these guidelines. Therefore, the studies had many deficiencies in the development and validation processes, involving participants, predictors, outcome and analysis domains.
近年来发表了多项关于预测 PPD 风险的研究。然而，已发表的研究总数仍然微不足道，其总体质量不足。此外，这些研究中提出的模型的性能参差不齐，因此不能直接应用于临床实践。这表明 ML 近年来引起了极大的兴趣，但它仍然是一个相对较新的领域。一个可能的原因是 PROBAST 于 2019 年发布（Moons et al.， 2019），而 2020 年提出了 ML 研究进行的标准化指南（Stevens et al.， 2020），从而导致在这些指南发布之前进行的一些研究总体上缺乏基于方法的参考。因此，这些研究在开发和验证过程中存在许多缺陷，涉及参与者、预测因子、结果和分析领域。

Regarding participants, some studies obtained data from EHRs or medical system database (Amit et al., 2021; Hochman et al., 2021; Park et al., 2021; Shin et al., 2020; S. Wang et al., 2019; Zhang et al., 2021), and as a result, the predictors and outcomes may be missing or defined in different ways, thereby resulting in decreased data integrity and availability. Additionally, the criteria for eligibility in some studies excluded subgroups with the potential risk of PPD, such as women with a history of mental disorders (Fang, 2019), those with thyroid dysfunction (Xiao et al., 2020), and those suffering from chronic diseases. However, previous studies have shown that a history of mental disorders is a strong predictor for the risk of PPD (Shakeel et al., 2018), thyroid dysfunction poses an increased risk for PPD (RR 1.49, 95 % CI 1.11–2.00) (George et al., 2021; Johar et al., 2020), and chronic diseases, such as diabetes, are also associated with the increased risk for PPD (OR 2.23, 95 % CI 1.23–4.05) (Ruohomaki et al., 2018).
关于参与者，一些研究从 EHR 或医疗系统数据库获取数据（Amit 等人，2021 年;Hochman 等人，2021 年;Park 等人，2021 年;Shin et al.， 2020;S. Wang et al.， 2019;Zhang et al.， 2021），因此，预测因子和结果可能缺失或以不同的方式定义，从而导致数据完整性和可用性降低。此外，一些研究的资格标准排除了具有 PPD 潜在风险的亚组，例如有精神障碍史的女性（Fang，2019 年）、甲状腺功能障碍患者（Xiao 等人，2020 年 ）和患有慢性病的女性 。然而，以前的研究表明， 精神障碍史是 PPD 风险的有力预测因素（Shakeel 等人，2018 年），甲状腺功能障碍会增加 PPD 的风险（RR 1.49,95 % CI 1.11-2.00）（George 等人，2021 年;Johar et al.， 2020） 和糖尿病等慢性病也与 PPD 风险增加有关 （OR 2.23， 95 % CI 1.23–4.05） （Ruohomaki et al.， 2018）。

Regarding predictors, some studies developed ML-based models that mainly focused on specific aspects, such as infants' cry vocalizations (Gabrieli et al., 2020) and the linguistic features of social-media posts(Fatima et al., 2019), which may result in a certain rate of missed screening owing to non-comprehensive predictors. Additionally, several studies did not report the definition, classification, measurement, and effectiveness of the included predictors in their final models, thereby resulting in a high risk of bias. Furthermore, the number of predictors in most studies exceeded 20, thereby making it significantly difficult to make assessments and recordings in a busy clinical setting and significantly reducing the applicability of the prediction models. Hochman et al. evaluated and compared the performance levels between a main model using 156 predictors and a Q-based model using 9 predictors, and they established that the latter was simpler and easier to implement, with only a slight reduction in accuracy (Hochman et al., 2021). Model performance metrics, especially accuracy and AUC scores, remain stable even when the number of variables used in the models is reduced from 100 % to 50 % and even when it is reduced to 25 % of all the variables available (5–10 variables) (Andersson et al., 2021). As discussed above, this may be due to redundancy during the evaluation of a large number of predictors, for example, depression and anxiety during pregnancy being highly correlated with some background and medical history variables.
关于预测因子，一些研究开发了基于 ML 的模型，主要关注特定方面，例如婴儿的哭声发声（Gabrieli et al.， 2020）和社交媒体帖子的语言特征（Fatima et al.， 2019），这可能导致一定的筛选漏检率由于预测因子不全面。此外，一些研究在其最终模型中没有报告所包含预测因子的定义、分类、测量和有效性，因此存在高偏倚风险。此外，大多数研究中的预测因子数量超过 20 个，从而使得在繁忙的临床环境中进行评估和记录变得非常困难，并显着降低了预测模型的适用性。Hochman 等人评估并比较了使用 156 个预测变量的主模型和使用 9 个预测变量的基于 Q 的模型之间的性能水平，他们确定后者更简单、更容易实现，准确性仅略有降低（Hochman et al.， 2021）。即使模型中使用的变量数量从 100 % 减少到 50 %，甚至减少到所有可用变量的 25%（5-10 个变量），模型性能指标，尤其是准确性和 AUC 分数，也保持稳定（Andersson 等人，2021 年 ）。如上所述，这可能是由于在评估大量预测因子时存在冗余，例如，怀孕期间的抑郁和焦虑与一些背景和病史变量高度相关。

Regarding outcomes, the risk of bias in a few studies mainly originated from the use of suboptimal classification methods of PPD. Generally, the diagnosis of ICD-9/10 codes, and the results of scales screening or psychiatrist clinical interview are standard or accepted classification methods for PPD symptoms or risk. However, two studies assessed patient's PPD by their linguistic features, which was related to that modeling data were social media textual data. This raised two considerations regarding the risk of bias, one is that the results are not classified reasonably enough, and the other is that the measurement of outcome might utilize predictors, which would reduce the PPD risk model's predictive performance confidence.
关于结局，一些研究的偏倚风险主要源于使用次优的 PPD 分类方法。一般来说，ICD-9/10 代码的诊断以及量表筛查或精神科医生临床访谈的结果是 PPD 症状或风险的标准或公认的分类方法。然而，两项研究通过患者的语言特征评估了患者的 PPD，这与建模数据是社交媒体文本数据有关。这引发了关于偏倚风险的两个考虑因素，一个是结果分类不够合理，另一个是结果的测量可能会使用预测变量，这会降低 PPD 风险模型的预测性能置信度。

The analysis domain has the most problems associated with the risk of bias across all the studies. First, none of the included studies justified the appropriateness of their sample size, given the number of candidate predictors used in model development. The range of events per variable (EPV) in several studies was small and probably resulted in the risk of overfitting. Although the range of EPV for at least 20 has been recognized as no risk of bias in PROBAST, prediction models developed using ML techniques often require the number of EPV to be over 200 (Moons et al., 2019). Therefore, in our future studies, we shall use dimensional reduction techniques before modeling to reduce the number of candidate predictors if a small sample is inevitably used. Next, model internal validation in several studies used random split validation only. However, when the sample size is small, it has a low data utilization rate, resulting in the risk of bias, whereas K-fold cross validation and Bootstrap resampling are better alternatives (Moons et al., 2012b). To our regret, only approximately 20 % of the models underwent complete external verification. This is probably because most researchers pay more attention to model development than model validation, thereby resulting in the continuous emergence of models for predicting the risk of PPD over the recent years, but few of them have been effectively verified. Finally, most of the studies reported discrimination metrics only and lacked calibration metrics, indicating that they cannot achieve consistency between the event probabilities predicted by most ML-based models and the observed event probability. Actually, an excellent prediction model should have high discrimination and calibration levels, and the former should be the basis of the latter. If the calibration is not satisfactory, the model should be recalibrated to enhance performance (Moons et al., 2012a).
在所有研究中，分析领域与偏倚风险相关的问题最多。首先，考虑到模型开发中使用的候选预测因子的数量，纳入的研究都没有证明其样本量的适当性。在几项研究中，每个变量的事件范围 （EPV） 很小，可能导致过拟合的风险。尽管至少 20 的 EPV 范围已被公认为在 PROBAST 中没有偏倚风险，但使用 ML 技术开发的预测模型通常要求 EPV 的数量超过 200（Moons 等人，2019 年）。因此，在我们未来的研究中，如果不可避免地使用小样本，我们将在建模前使用降维技术来减少候选预测因子的数量。接下来，几项研究中的模型内部验证仅使用随机拆分验证。然而，当样本量较小时，数据利用率较低，导致偏倚风险，而 K 折叠交叉验证和 Bootstrap 重采样是更好的选择（Moons et al.， 2012b）。遗憾的是，只有大约 20% 的模型进行了完整的外部验证。这可能是因为大多数研究人员更注重模型开发而不是模型验证，从而导致近年来预测 PPD 风险的模型不断涌现，但很少有得到有效验证的模型。最后，大多数研究仅报告了判别指标，缺乏校准指标，这表明它们无法在大多数基于 ML 的模型预测的事件概率与观察到的事件概率之间实现一致性。 其实，一个优秀的预测模型应该具有较高的判别和校准水平，前者应该是后者的基础。如果校准不令人满意，则应重新校准模型以提高性能（Moons et al.， 2012a）。

Throughout the literature, there is a growing trend of using EHRs to develop ML-based models for predicting the risk of PPD, thereby resulting in higher requirements for data processing technology. In this systematic review, a total of four studies developed ML-based models for predicting the risk of PPD using EHRs (Amit et al., 2021; Hochman et al., 2021; S. Wang et al., 2019; Zhang et al., 2021), including the latest three studies. This tendency exists because EHRs comprise digital health information documented by tracking repeated measurements of patients' conditions over time, and they cover a large population (Aliabadi et al., 2020). Accurate and longitudinal EHRs have the potential to be used in the identification and analysis of new risk factors or outcomes at the individual and population levels. Additionally, ML is a vigorous data mining technique that is suitable for deriving further insight from large sets of multivariate medical data (Nwanosike et al., 2022). Therefore, EHR-based ML presents a potential to develop more comprehensive and usable models for risk stratification. However, EHRs, as sources of temporal data, usually have complex structures and unbalanced distribution of clinical events, thereby posing multiple technical challenges, such as data irregularity, heterogeneity, and sparsity. Given the limitations of general ML algorithms in addressing these challenges, advanced ML approaches, such as deep learning-based methods must be proposed in future studies (Xie et al., 2022).
纵观文献，使用 EHR 开发基于 ML 的模型来预测 PPD 风险的趋势越来越大，从而对数据处理技术提出了更高的要求。在本系统评价中，共有四项研究开发了基于 ML 的模型，用于使用 EHR 预测 PPD 风险（Amit 等人，2021 年;Hochman 等人，2021 年;S. Wang et al.， 2019;Zhang et al.， 2021），包括最新的三项研究。之所以存在这种趋势，是因为 EHR 包含通过跟踪患者病情随时间的重复测量来记录的数字健康信息，并且它们涵盖了大量人群（Aliabadi 等人，2020 年 ）。准确和纵向的 EHR 有可能用于识别和分析个人和人群水平的新风险因素或结果。此外，ML 是一种强大的数据挖掘技术，适用于从大量多变量医疗数据中获得进一步的见解（Nwanosike 等人，2022 年）。因此，基于 EHR 的 ML 有可能开发更全面和可用的风险分层模型。然而，EHR 作为时间数据的来源，通常具有复杂的结构和临床事件分布不平衡，从而带来了数据不规则性、异质性和稀疏性等多重技术挑战。鉴于通用 ML 算法在应对这些挑战方面的局限性，未来的研究必须提出先进的 ML 方法，例如基于深度学习的方法（Xie et al.， 2022）。

Owing to advancements in ML technology, a series of ensemble and reinforcement learning algorithms show high performance, but the interpretability of such model decreases, and this is called the black-box effect. However, when using clinical data to make predictions and decisions, it is necessary to know both the model output results and the risk factors, their weight, correct interpretation, and specific guidance for clinical practice. Therefore, researchers should be mindful not to pursue high accuracy at the cost of the significance for clinical practice. However, in this systematic review, few models were translated into clinical tools and tested in clinical environments. While some studies have interpreted which features have a greater impact on ML models through feature importance, we could not explain the specific association between features and the prediction results (S. Wang et al., 2019). To improve the interpretability of ML models, recent studies have replaced feature importance with the SHAP mean to evaluate the contribution of each feature to the model output, especially for ensemble tree models, such as XGBoost (Hochman et al., 2021). In our opinion, rationally interpreting ML models or translating ML model into clinical tools are still the focus of researchers and clinical workers, and they are also the key to promoting the continuous development and improvement of ML models in clinical practice.
由于 ML 技术的进步，一系列集成和强化学习算法表现出高性能，但此类模型的可解释性降低，这称为黑盒效应。但是，在使用临床数据进行预测和决策时，有必要了解模型输出结果和风险因素、它们的权重、正确的解释以及对临床实践的具体指导。因此，研究人员应注意不要以牺牲临床实践的意义为代价来追求高准确性。然而，在这项系统评价中，很少有模型被转化为临床工具并在临床环境中进行测试。虽然一些研究通过特征重要性解释了哪些特征对 ML 模型的影响更大，但我们无法解释特征与预测结果之间的具体关联（S. Wang et al.， 2019）。为了提高 ML 模型的可解释性，最近的研究用 SHAP 均值取代了特征重要性，以评估每个特征对模型输出的贡献，特别是对于集成树模型，例如 XGBoost（Hochman 等人，2021 年）。在我们看来，理性解读 ML 模型或将 ML 模型转化为临床工具仍然是研究人员和临床工作者关注的重点，也是推动 ML 模型在临床实践中不断发展和改进的关键。