CTGGAN:A modified generative adversarial network for imbalanced CTG signal classification during labor

Cardiotocography (CTG) is a primary tool for real-time monitoring fetal heart rate (FHR) and uterine contraction (UC) signals during labor. Nowadays CTG signal classification models based on deep learning have facilitated end-to-end CTG interpretation through deep neural networks for automatic feature extraction. However, the extreme rarity of abnormal cases in clinical settings has posed great challenges for CTG deep learning. Meanwhile, most Generative Adversarial Network (GAN)-based studies focused on FHR signal generation and ignored the importance of UC signals during labor. Hence, we introduced CTGGAN, a modified GAN specifically designed to simulate both FHR and UC intrapartum signals simultaneously. Our approach incorporated Wasserstein Distance to prevent gradient vanishing during training, alongside a set of latent code vectors that enhance the generator controllability. We also integrated an auxiliary classifier to impose categorical constraints, thereby enriching the diversity of the generated samples. The experimental results demonstrated that CTGGAN effectively generated CTG signals that closely resembled real CTG signals. In addition, the incorporation of CTGGAN significantly improved performance, with a 19.18% increase in sensitivity over the traditional classification without GAN. Furthermore, the comparison with several state-of-the-art CTG classification models underscored the discriminative performance of our CTGGAN. In conclusion, this study effectively enhanced the detection of abnormal cases for intrapartum fetal monitoring.