Acquiring Domain Knowledge for Cardiotocography: A Deep Learning Approach

Abstract

Infant cardiac distress is the leading cause of neonatal deaths in the world. Cardiotocography (CTG) is a diagnostic tool used for recording fetal heartbeat and uterine contractions during pregnancy to determine cardiac distress. To avoid the need of continuous monitoring by on-site medical personnel, researchers have been working on several machine learning tools to automate the process. Most of these approaches discover statistical trends in data to predict target variables. However, being reliant on these trends makes them prone to overfitting and other statistical perils. In this paper, we demonstrate the usage of a modified deep neural network to learn about 2 seemingly disjointed tasks in the field of cardiotocography. The proposed model acquires predictive power in one task whilst being trained on a separate yet related task in the same field. Further, it establishes that regularization facilitates the sharing of knowledge across tasks. The resulting model mimics the human learning process by demonstrating the ability to acquire domain knowledge.