Enhanced Fetal Arrhythmia Classification by Non-Invasive ECG Using Cross Domain Feature and Spatial Differences Windows Information

Fetal arrhythmia refers to an abnormal heart rhythm in a fetus, characterized by irregular, too fast (tachycardia), or too slow (bradycardia) heartbeats. Diagnosis is typically made using ultrasound or non-invasive fetal electrocardiography (NI-FECG), which monitors the electrical activity of the fetal heart. Accurate detection and management are crucial, as severe arrhythmias may lead to complications that affect the fetus’s health during pregnancy and delivery. This study aims to develop an effective screening method for detecting arrhythmia (ARR) to aid physicians in diagnosing potential heart disease in fetal during pregnancy. This study presents a new cross-domain feature extraction method that incorporates temporal relationships between consecutive windows, improving feature representation by examining the correlation between neighboring windows. First, the original waveform signals from six sensors were transformed into a multi-level decomposition using the HAAR wavelet. Subsequently, a sample expansion was applied using a various-sized window sliding approach to each ARR and normal signal. Second, feature selection was implemented to reduce data dimensionality by selecting features highly relevant to the class labels. Finally, we applied oversampling techniques to address the issue of imbalanced data. Based on the results of an indepth experimental analysis, it was found that the application of window sampling for data expansion produced favorable outcomes, particularly at a window size of 500. The combination of dimensionality reduction using Mutual Information and oversampling with the Radius-SMOTE method achieved the highest performance, yielding an accuracy of 96.5%, precision of 95.2%, recall of 96.3%, and an F1-measure of 96.2%, while reducing the feature dimensionality to 100. Moreover, the proposed method demonstrated improved performance compared to state-of-the-art approaches. With further development, this method is expected to serve as a foundational model for the advancement of diagnostic tools for physicians.