Screening for Late-Onset Fetal Growth Restriction in Antepartum Fetal Monitoring Using Deep Forest and SHAP

Abstract

Fetal growth restriction (FGR) is the second leading cause of perinatal death, of which late-onset fetal growth restriction (LFGR) accounts for 70%–80% and has a low detection rate. Cardiotocography (CTG) is a routine tool for antepartum fetal monitoring, continuously recording fetal heart rate (FHR) to assess the development of the fetus. Therefore, in this paper, we proposed a hybrid DF-SHAP model that screens LFGR in routine CTG monitoring using deep forest (DF) and Shapley Additive Explanation (SHAP). Firstly, principal component analysis, spearman correlation analysis and logistic regression analysis were implemented to explore significant FHR features for LFGR. After data preprocessing, deep forest multi-granularity scanning was introduced to probe the connection among the features. Then the cascade forest phase, which was designed to integrate random forest, extra trees, logistic regression and extreme gradient boosting as the basic classifiers, iteratively generated new layers and finally got the best performance model. Finally, SHAP was introduced to enhance the interpretability of DF and to interpret the impact of each feature on the predicted value. The experimental results showed that the proposed DF-SHAP model outperformed the state-of-the-art LFGR recognition models using CTG data and had good interpretability. This indicates that the DF-SHAP model is feasible for screening LFGR in antepartum fetal monitoring.