Objective To construct a predictive model for assisted diagnosis of maternal perinatal depression (PND) based on machine learning (ML) algorithms and to evaluate the performance of different ML algorithm models.

Methods A total of 5 814 pregnant women (4 665 prenatal study subjects and 1 149 postnatal study subjects) were evaluated using the 9-item Patient Health Questionnaire Depression Scale (PHQ-9). A total of 19 scale dimension variables of 7 Scales and demographic characteristics were collected as observation variables. Prenatal and postnatal subjects were matched at a 1∶1 ratio propensity score according to age. The feature selection variables were determined by single factor analysis and Pearson correlation coefficient. A diagnostic model for prenatal and postnatal depression was constructed based on five ML algorithm, including Logistic regression model, Random Forest (RF), support vector machine (SVM), Limit Gradient Lift Tree (XGBoost) and Backpropagation (BP) neural network. A 5-fold cross-validation method was used to evaluate thepredictive performance of the model, including sensitivity, specificity and area under the curve (AUC).

Results When different variables were included, the sensitivity, specificity and AUC of the prediction model constructed by five ML algorithms based on prenatal and postnatal subjects were all within the range of 0.600 to 0.900. RF algorithm was the optimal algorithm in the construction of both prenatal prediction model(when all variables were included, the AUC was 0.834; when the feature selection variable set was included, the AUC was 0.849) and postnatal prediction model(when all variables were included, the AUC was 0.873; when the feature selection variable set was included, the AUC was 0.864).

Conclusion The prediction model based on five ML algorithms can effectively predict the risk of PND in pregnant women, and the performance of RF algorithm is the best, which provides a reference for the development of auxiliary tools for rapid screening and diagnosis of PND.