A Novel Approach for Multi-Class Fetal ECG Arrhythmia Classification by ResNet-50 Combined with Transformer

Abstract

Accurate classification of fetal arrhythmias from non-invasive fetal electrocardiogram (NI-fECG) signals is essential for the early diagnosis and prenatal management of congenital cardiac abnormalities. However, reliable automated fetal arrhythmia classification remains challenging due to the low signal-to-noise ratio of fetal ECGs, maternal ECG interference, class imbalance, and the limited availability of annotated fetal ECG datasets. To address these challenges, this study proposes a patient-wise deep transfer learning framework for multiclass fetal ECG arrhythmia classification using adult ECG pretraining and a hybrid ResNet50–Transformer network. The proposed framework first pretrains the hybrid model on the MIT-BIH Arrhythmia Database to learn generalized cardiac electrophysiological representations from a large-scale adult ECG dataset. The pretrained model is then fine-tuned on the Non-Invasive Fetal ECG Arrhythmia Database (NIFEADB) to adapt the learned representations for fetal cardiac rhythm analysis. Prior to classification, NI-fECG signals are preprocessed and transformed into time-frequency spectrograms using the Short-Time Fourier Transform (STFT). ResNet50 is employed to extract discriminative spatial features from the spectrograms, while the Transformer encoder models global contextual relationships through self-attention to improve the discrimination of subtle arrhythmic patterns. To eliminate data leakage and provide a clinically realistic evaluation, the proposed framework adopts a strict patient-wise validation protocol, ensuring that recordings from the same subject are never simultaneously included in both the training and testing sets. The framework performs multiclass classification of four clinically significant fetal cardiac conditions, namely Normal Rhythm, Premature Atrial Contractions (PACs), Tachyarrhythmia, and Bradyarrhythmia/Heart Block (HB). Experimental results demonstrate that the proposed patient-wise transfer learning framework effectively leverages knowledge learned from adult ECGs to improve fetal arrhythmia classification despite the limited availability of fetal training data, providing a robust and clinically reliable solution for automated prenatal cardiac screening.