FAViTNet: An Effective Deep Learning Framework for Non-Invasive Fetal Arrhythmia Diagnosis

Analyzing fetal electrocardiograms (fECG) to classify fetal arrhythmia is a challenging task; still, it is indispensable for evaluating fetal cardiac health status. This study intends to develop a framework for the effective discernment of fetal arrhythmia that assists obstetricians in diagnosing whether the fetuses have any abnormality in cardiac rhythm. In this work, authors aim to develop a three-stage fetal arrhythmia detection framework comprised of 1) fECG extraction from abdominal Electrocardiogram (aECG) signal based on functional link neural network adaptive filter, 2) denoising of extracted fECG exploiting a generative adversarial network where the generator and the discriminator follow robust architectures, and 3) fetal arrhythmia detection based on FAViTNet architecture. The proposed work deeply focuses on the last stage, where we have unfolded a deep convolutional architecture that detects fetal arrhythmia utilizing high-resolution spectral images acquired from the Stockwell transform. The classification stage demonstrates an architecture involving spectral image tokenization and transformer encoder blocks. The spectral image tokenization module involves ghost bottleneck blocks, which are utilized to generate feature maps from the spectral images, subsequently transformed into 1D token embeddings. Channel-wise calibration is utilized to attain more attention to the deep feature maps acquired from the ghost network. The transformer encoder block effectively attains the long-term dependencies from the 1D embedded features with gated linear unit (GLU)-based multi-head self-attention where deep features are learned to discern fetal arrhythmia effectively. The proposed algorithm shows an accuracy of 96.85%, sensitivity of 96.69%, specificity of 96.98%, and precision of 96.48%.