MFConvTr: Multi-Frequency Convolutional Transformer for Fetal Arrhythmia Detection in Non-Invasive fECG

TL;DR

This paper introduces MFConvTr, a novel deep learning model that effectively detects fetal arrhythmia from non-invasive fetal ECG by capturing multi-frequency information and modeling long-term dependencies, achieving state-of-the-art results.

Contribution

The paper proposes a Multi-Frequency Convolutional Transformer architecture that efficiently learns multi-frequency features and long-term dependencies in fetal ECG signals, outperforming existing methods.

Findings

Achieves state-of-the-art detection accuracy.

Uses fewer parameters than existing models.

Extensive ablation studies validate the design choices.

Abstract

NI-fECG have emerged as alternative for fetal arrhythmia monitoring. But due to multi-signal waveform they are tough to understand and due to highly varying and complex nature traditional fiducial methods cannot be applied. Further, it has also been observed that the fetal arrhythmia can be differentiated from the normal signals in both spectral and temporal scales. To this end, we propose Multi-Frequency Convolutional Transformer, a novel deep learning architecture that learns information in contexts with multiple-frequency and can model long-term dependencies. The proposed model utilizes a convolutional-backbone consisting of model Multi-Frequency Convolutions (MF-Conv) and residual connections. MF-Conv in-turn captures multi-frequency contexts in an efficient manner by splitting the input channel and then convoluting each of the splits individually with different kernel size. Accredited to these properties, the proposed model attains state-of-the-art results and that too utilizing very low number of parameters. To evaluate the proposed we also perform extensive ablation studies.