Scalogram-Based Binary Fetal Arrhythmia Classification Using ML and DL

Objectives: Fetal electrocardiography (FECG) is a valuable tool for assessing fetal health and identifying cardiac rhythm abnormalities. This study aimed to determine the convolutional neural network (CNN)-based deep learning (DL) model that provides the highest performance for binary classification of fetal arrhythmia and normal rhythm (ARR vs NR) in FECG signals. Methods: FECG signals were converted to time-frequency images using the Continuous Wave Transform (CWT), and these images were used as input to DL-based CNN architectures such as AlexNet, GoogleNet, and ResNet50. The effect of data length was investigated to achieve high accuracy with transfer learning. In addition, these DL models were compared with traditional machine learning methods. Results: Among the evaluated DL models, AlexNet achieved the highest performance on the 10-second segments. It outperformed the traditional machine learning approaches across all evaluated metrics. Among the conventional methods, the Trilayered Neural Network showed the best performance, with an accuracy of 87.50%. Conclusions: The findings suggest that scalogram-based CNN models have strong potential for classifying binary fetal arrhythmias from FECG signals. In particular, the AlexNet model showed promising performance under the experimental conditions of this study. However, the results should be interpreted with caution due to the limited dataset size, and further validation on larger, independent datasets is needed.