Siamese Neural Networks for IUGR Identification in Cardiotocographic Recordings

The Fetal Heart Rate (FHR) is a valuable source of information which is routinely monitored to assess fetal well-being via cardiotocography (CTG).Recently, Siamese Neural Networks (SNN) have been introduced as an effective strategy for metric learning, due to their capability to learn complex embeddings useful for classification with little training data.In this paper, we present preliminary results on the use of SNNs for intra-uterine growth restriction (IUGR) identification in CTG recordings.The developed model comprises three identical residual branches and computes the Euclidean distance among the embeddings of the inputs (i.e., the triplet).The dataset employed is extracted from a wider CTG database and is balanced with respect to the class and gestational age.In particular, 466 recordings are used for training and validation and 92 as a hold-out test set.The trained SNN is then used to project the data into a lower dimensional space where they are classified using a k-Nearest Neighbors classifier.The proposed approach allows to discriminate between IUGR and physiological pregnancies with an accuracy score on the test set ranging from 63% to 72.6% when varying the number of neighbors.The results, which are still far from being satisfactory for clinical applications, align with previous studies, showing the validity of the proposed approach.