Fetal ECG Arrhythmia Detection Based on DensNet Transfer Learning

Abstract

Purpose: The mortality rate of fetuses due to heart defects is a major concern for clinicians. The fetus's heart is monitored non-invasively using the abdominal Electrocardiogram (ECG) of the mother. Most of the methods in literature diagnose fetal arrhythmia based on fetal heart rate. However, there are various challenges in fetal heart rate monitoring and arrhythmia detection. Therefore, very few methods are explored for fetal arrhythmia classification and have not achieved promising results. Materials and Methods: In this article, a fetal arrhythmia classification method is investigated. The method has exploited the transfer learning principle where DenseNet architecture is utilized to learn fetal ECG patterns. Fetal ECG (fECG) signal extracted from the mothers abdominal has been processed for denoising and heartbeats are segmented using signal processing techniques. The extracted heartbeats have transformed into 2D fECG images to re-train the pre-trained DenseNet architecture. Results: The proposed method has been evaluated on the publicly available Non-Invasive Fetal Arrhythmia Database (NIFADB) of Physionet and achieved 98.56% classification accuracy, thus outperforming other existing methods. Conclusion: The arrhythmia in a fetus can be detected using a non-invasive fetal ECG. Due to the faster convergence of the learning algorithm, the proposed method offers better fetal diagnosis in real-time.