Deep Gaussian Mixture Model on Multiple Interpretable Features of Fetal Heart Rate for Pregnancy Wellness

Abstract

About 1% of the newborns have disorders of cardiac rhythm and conduction that result in heart failure, brain damage and even sudden newborn death. To ensure the safety of pregnancy, the electronic fetal monitoring (EFM) technique is widely used in obstetrics. However, the existing automatic diagnosis methods suffer from two main problems: insufficient features and low interpretability. In order to improve the interpretability and effect of the method, we propose a novel fully interpretable method. We first propose an iterative local linear regression (ILLR) method of linear complexity, which calculates over all local ranges of the fetal heart rate (FHR) and generates local gradients and coefficients of determination, that are used as indicators of intensity and typicality of fetal heart activity. Then, we elaborate the methodology of extraction of dozens of features by interpretable methods. Finally, we propose an interpretable deep Gaussian mixture model that can automatically select multiple features, which is composed of a mixture model based on Gaussian model weighted by features and a regression model. We conduct cross validation experiments on the full benchmark intrapartum database CTU-UHB, which shows that our method obtains significant improvements of 5.61% accuracy over state-of-the-art baselines.