Real-time fetal arrhythmia detection using deep learning on an embedded fECG monitoring system

Abstract

• A deep learning method is proposed for fetal arrhythmia detection via fECG. • ICA and filtering are used to extract fetal ECG from abdominal and thoracic signals. • A CNN-based system detects fetal P, QRS, and T waves from fECG. • A lightweight CNN classifies 3-second segments as normal or arrhythmic. • The system runs in real time on Raspberry Pi with high classification accuracy • Arrhythmia detection is performed and visualized on an OLED display. • Results are sent to a mobile app for continuous monitoring and alerting. • Waveform segmentation achieves high sensitivity and precision metrics. • The system enables compact, low-cost, and non-invasive fetal monitoring. Fetal arrhythmia is a critical condition that can cause severe complications and even sudden infant death if not detected in time. Early and accurate diagnosis is therefore essential to enable timely medical intervention and improve neonatal outcomes. Traditional detection methods are limited by low signal quality, due to the weak amplitude of the fetal ECG (fECG) and interference from the maternal ECG, which hinders reliable real-time, non-invasive monitoring. This work presents the development of a real-time, non-invasive, and embedded system for accurate fetal arrhythmia detection using deep learning applied to fECG signals. Abdominal and thoracic ECG recordings are preprocessed with Independent Component Analysis (ICA) to isolate the fetal component, followed by band-pass filtering and segmentation into 3-second windows. Each segment is classified by a lightweight one-dimensional convolutional neural network (CNN) trained on annotated data from the Fetal Health Classification Dataset (Kaggle) and PhysioNet NIFECGDB and optimized for real-time inference on Raspberry Pi. Classification outcomes are transmitted to a mobile application for continuous monitoring and alerting. Experimental results demonstrate high performance, achieving 98.2% accuracy, 98.8% sensitivity, 97.4% specificity, and a 98.1% F1-score. The embedded implementation confirms the system’s suitability for portable, real-time fetal monitoring, offering a robust and accessible alternative to conventional medical analysis.