Article’s

Fetal electrocardiogram (FECG) analysis is a valuable method for evaluating fetal cardiac activity and diagnosing abnormal cardiac rhythms in pregnancy. Reliable fetal arrhythmia classification models are, however, hampered by the lack of labeled FECG data because obtaining such data is challenging and time-consuming. Existing deep learning (DL) approaches predominantly rely on supervised learning and may suffer from overfitting and limited generalisation when trained on small datasets. Moreover, fetal arrhythmia is also reflected in a morphological variation of the FECG waveform and temporal abnormalities of cardiac rhythm. Most of the classification methods used, however, only learn these features from a small amount of labelled data, and they fail to explicitly make use of the relationship between the morphology of the waveforms and the rhythm pattern to which they belong. This article introduces a Morphology–Rhythm Contrastive Representation Learning framework (MRCL-FECG) for label-efficient fetal arrhythmia classification to overcome these drawbacks. Pre-processing and segmentation of FECG are performed first to get fixed-length windows, in which cardiac rhythm information is represented by the corresponding RR-interval and Fetal Heart Rate (FHR) sequences. The FECG waveforms are used to extract the morphological representation using lightweight one-dimensional convolutional neural networks (1D-CNNs) and the temporal rhythm using the RR-interval and FHR sequence. The resulting morphology and rhythm representations are projected in a shared feature space that encourages agreement among complementary representations derived from the same cardiac segment, and differentiates the representation of different segments, while being unrelated to one another. The learned representations are then fine-tuned for fetal arrhythmia classification with a few-shot learning method with a small number of labelled fetal data.