During the perinatal period, the risk of developing depression is high and it is estimated that approximately 10%–15% of mothers experience perinatal depression.¹ In recent years, machine learning has been widely used in the research of mental health, and it has been suggested that machine learning could be useful in the clinical management of mental disorders by providing accurate predictions for the diagnosis, prognosis. If an effective postpartum depression (PPD) prediction model can be established, it will enable early identification of high-risk individuals and early intervention by healthcare providers in high-risk individuals.²

In this study, we used machine learning methods to construct a prediction model for depression in the first postpartum month using demographic information and subjective ratings of pregnant women collected from the time of pregnancy to the fifth postpartum day after delivery. A verbal and written explanation of the study was given to all participants, and written informed consent was obtained from all those who agreed to participate. The study protocol was approved by the Ethics Committee of the Nagoya University Graduate School of Medicine. Detailed methods are described in the Supplementary materials Appendix S1. The flowchart of the study procedures is shown in Fig. S1. 1559 women participated in the study and 1416 women responded to all 10 Edinburgh Postnatal Depression Scale items 1 month after delivery. In this study, we included these 1416 women (mean age 32.4 years, standard deviation ±4.6 years) in our machine learning. The flowchart of the recruitment process is shown in Fig. S2. We show the method details in Appendix S1 and comparison of predictors between the PPD group and the non PPD group in Table S1. We also show missing percentages for each item in Table S3.

We used a machine learning approach, logistic regression, decision tree, gradient-boosting decision tree (GBDT), and balanced GBDT to develop the PPD prediction model. GBDT gives a predictive model as an ensemble of decision tree and achieves high predictive ability with a differentiable loss function. Early prediction models for PPD need to be sensitive so as not to overlook women at high risk for PPD. Therefore, we set sensitivity to 80% and built the model using the machine learning approach. We calculated a 95% confidence interval for the AUC to confirm the predictive power of the model used. The results are shown in Table S2. The high accuracy (73.11%), high specificity (71.3%) and high positive predictive value (42.2%) was obtained in the balanced GBDT, and the high AUC value (0.8285) was obtained from the logistic regression model. Therefore, we thought that the balanced GBDT model would be the best. Figure 1 shows the feature value with high importance and AUC when using the balanced GBDT. In addition, we used the 28 features shown in Fig. 1 as predictors and built a predictive model again using the balanced GBDT. From the results, we were able to build a model with high accuracy and good sensitivity and specificity (accuracy 78.6%, AUC 0.83, sensitivity 71.1%, specificity 80.6%, positive predictive value 48.9%, negative predictive value 91.4%).

In this study, we used machine learning technique and successfully constructed a highly accurate and convenient model for predicting depression in the first month after childbirth using data from pregnant women. In our model, we found that depression during pregnancy and immediately after delivery, maternity blues, age, parity, temperament, and other factors were important predictors. The prevalence of PPD is about 10%–15%, and the proportion of positives is smaller than the proportion of negatives. Therefore, the positive predictive value (PPV) tends to be low, as in previous studies^{3, 4} and this study. Future studies will need to develop a prediction tool with high PPV. In addition to the data used as predictors in this study, previous studies using machine learning reported that any history of depression is an important factor.³ In our previous studies, we also confirmed that a history of depression⁵ is a risk factor for PPD. However, since the number of responses to this item of our study was still small, we could not add it to this analysis. We think that it is important to try to build a model that includes this item to the predictor in the future.

We were able to build a simple and useful model for predicting PPD by using machine learning techniques. In the future, it will be important to build a more accurate model by adding data on the history of depression to the predictor.

围产期患抑郁症的风险很高，估计大约 10%–15% 的母亲患有围产期抑郁症。¹近年来，机器学习已广泛应用于心理健康研究，有人认为机器学习可以通过为诊断和预后提供准确的预测来应用于精神障碍的临床管理。如果能够建立有效的产后抑郁症（PPD）预测模型，将有助于早期识别高危人群，并让医疗保健提供者对高危人群进行早期干预。²

在这项研究中，我们利用机器学习方法，利用从怀孕时到产后第五天收集的孕妇的人口统计信息和主观评分，构建了产后第一个月抑郁症的预测模型。向所有参与者提供了该研究的口头和书面解释，并获得了所有同意参与的人的书面知情同意书。该研究方案得到了名古屋大学医学研究生院伦理委员会的批准。详细方法在补充材料附录 S1 中描述。研究程序的流程图如图S1所示。 1559 名女性参与了这项研究，其中 1416 名女性在产后 1 个月对爱丁堡产后抑郁量表的所有 10 个项目做出了回应。在这项研究中，我们的机器学习纳入了这 1416 名女性（平均年龄 32.4 岁，标准差 ±4.6 岁）。招聘流程如图S2所示。我们在附录 S1 中显示了方法细节，并在表 S1 中显示了 PPD 组和非 PPD 组之间的预测因子比较。我们还在表 S3 中显示了每个项目的缺失百分比。

我们使用机器学习方法、逻辑回归、决策树、梯度提升决策树 (GBDT) 和平衡 GBDT 来开发 PPD 预测模型。 GBDT 给出了作为决策树集成的预测模型，并通过可微的损失函数实现了高预测能力。 PPD 的早期预测模型需要敏感，以免忽视 PPD 高风险女性。因此，我们将灵敏度设置为80%，并使用机器学习方法构建模型。我们计算了 AUC 的 95% 置信区间，以确认所用模型的预测能力。结果如表S2所示。平衡GBDT获得了高准确度（73.11%）、高特异度（71.3%）和高阳性预测值（42.2%），逻辑回归模型获得了高AUC值（0.8285）。因此，我们认为平衡的 GBDT 模型是最好的。图1显示了使用平衡GBDT时具有高重要性的特征值和AUC。此外，我们使用图1所示的28个特征作为预测器，并使用平衡GBDT再次构建了预测模型。从结果来看，我们能够建立一个具有高精度、良好敏感性和特异性的模型（准确性 78.6%，AUC 0.83，敏感性 71.1%，特异性 80.6%，阳性预测值 48.9%，阴性预测值 91.4%）。

在这项研究中，我们利用机器学习技术，利用孕妇的数据，成功构建了一个高度准确且方便的模型来预测产后第一个月的抑郁症。在我们的模型中，我们发现怀孕期间和分娩后的抑郁症、产妇忧郁症、年龄、胎次、气质和其他因素是重要的预测因素。 PPD的患病率约为10%~15%，阳性比例小于阴性比例。因此，阳性预测值 (PPV) 往往较低，如之前的研究^3、4和本研究一样。未来的研究需要开发一种具有高 PPV 的预测工具。除了本研究中用作预测因素的数据之外，之前使用机器学习的研究报告称，任何抑郁症史都是一个重要因素。³在我们之前的研究中，我们也证实抑郁症病史⁵是 PPD 的危险因素。然而，由于对我们研究的此项的回复数量仍然很少，因此我们无法将其添加到此分析中。我们认为，尝试构建一个将此项纳入未来预测器的模型非常重要。

我们能够使用机器学习技术构建一个简单且有用的模型来预测 PPD。未来，通过将抑郁症病史数据添加到预测器中来构建更准确的模型将非常重要。