Exploring Effective Approaches on Transformer-Based Neural Models for Multi-clinical Large-Scale Cardiotocogram Data

Cardiotocogram (CTG) is a method that is widely used to detect the abnormality of fetuses to reduce their medical risks and to promote positive prognosis in newborns in clinical practice. Automatic classification and processing of CTG is still a challenging task although there has been a lot of research, some of which applied CNNs as classification tasks. It is difficult even for obstetricians or experienced medical practitioners to interpret a pair of time series of fetal heart rates (FHR) and uterine contractions (UC) and to predict newborn prognosis. Recent rapid development in deep neural networks (DNNs) has been applied to medical diagnosis in many circumstances. In this study, we use large-scale real clinical data which is collected from multiple medical institutions. Each clinical sample has real prognoses labels such as Apgar scores and cord hydronium ions concentration (pH) values. We compare recently developed DNN models such as EfficientNet and Transformer variants (Vision Transformer (ViT), MetaFormer, and Swin Transformer). In addition, to enhance and stabilize prediction accuracy (ROC-AUC), we explore techniques to enhance prediction such as model architectures, preprocessings, and data-augmentations through the experiments. Particularly, the proposed data augmentation methods specialized for CTG data are shown effective to improve the accuracy.