Fetal Distress Classification from CTG Data using Deep Learning

Abstract

Clinical medical data, during pre-labour and labour of pregnancy, consist of multivariate time series of observations. Cardiotocography (CTG) is the graphical representation of such time series data, which is a constant tracing of newborn's heart rate (FHR) and uterine contractions (UC) . In the current situation, clinicians have to manually analyse the data to get insights and make decisions. The decisions vary among various clinicians. There is a need to have single inference that can be trusted. The further diagnostic decisions on the baby depends on such conclusions. This paper presents the study to evaluate the data automatically from the received images and make accurate differences among normal and abnormal cases using Deep Learning algorithms. Convolution neural network is used to classify images with the proposed architecture and further compared with other standard architectures like AlexNet and VGGNet for performance. The signals are converted to two dimensional grayscale images of respective size, which are applied to the CNN models after dividing into images of 800 timestamps each. The noise and zero values from the data are not discarded for accurate diagnosis. The FHR value of actual data has been smoothed using moving average keeping UC unchanged. Image augmentations are applied which increased the overall performance of the applied solution.