FHRGAN: Generative adversarial networks for synthetic fetal heart rate signal generation in low-resource settings

Abstract

Fetal heart rate (FHR) monitoring is an important medical-assisted diagnostic technique widely used by clinicians to assess fetal well-being. However, one challenge is that the capabilities of such diagnostic algorithms often rely on an enormous quantity of labeled clinical data to train a model, which do not preserve patient privacy. The high performance of such diagnostic algorithms is further hindered by category imbalance problems. Therefore, the general objective of this study is to develop a small-sample generation method that generates FHR signals of different physiological/pathological categories and arbitrary lengths. This study focuses on two significant impediments to the existing generation methods: the instability of generative adversarial networks (GAN) during model training, the mode collapse problem and the subsequent training of different specific models for different data categories, which contributes to high model training costs. To address these problems, we propose a novel generative adversarial architecture, referred to as CCWGAN-GP, based on a deep neural network optimized by the Wasserstein distance with gradient penalty, and incorporate an auxiliary classifier as a category constraint to enrich the diversity of generated data. The proposed method is comprehensively evaluated using 200 real FHR recordings from four aspects: training performance, the fidelity and diversity of generated data, and the potential improvement in the classification model. Compared with training on small-sample datasets and category-imbalanced datasets, training on augmented datasets improves the accuracy by approximately 12% and 8%, respectively. The developed architecture provides a reference value for a practical solution to the FHR data imbalance and insufficient sample problems.