Objective To construct a model for predicting the risk of postpartum depression (PPD) in parturients within 6 to 8 weeks after childbirth based on machine learning (ML) model and Shapley additive explanations (SHAP).
Methods The Edinburgh postnatal depression scale was used to evaluate the occurrence of PPD in 201 parturients within 6 to 8 weeks after delivery. The Boruta algorithm was used to screen the risk factors of PPD. Four ML models, namely support vector machine, naive Bayesian model, linear discriminant analysis and mixture discriminant analysis (MDA), were constructed to predict the risk of PPD. The prediction accuracy of four ML models was evaluated using the calibration curve, precision-recall curve, precision-recall-gain curve and receiver operating characteristic (ROC) curve, and the optimal ML model was interpreted and visualized through SHAP.
Results The incidence of PPD in 201 parturients was 15.92% (32/201). The Boruta algorithm screened out the Pittsburgh sleep quality index (PSQI) score, feeding method, social support rating scale (SSRS) score, educational attainment and per capita monthly income of family were the risk variables of PPD. Among the four ML models, the consistency index, recall rate, recall gain and area under the ROC curve of the MDA model were superior to those of the other three ML models. The SHAP interpretation and visualization of the MDA model could accurately predict the risk of PPD in parturients.
Conclusions The SSRS score, PSQI score, educational attainment, feeding style and per capita monthly income of family are associated with the PPD risk. The MDA model based on the interpretation and visualization of SHAP values can effectively predict the risk of PPD.
DownLoad: