Abstract
Congenital heart defects are the leading cause of mortality related to birth defects, making early monitoring of fetal electrocardiogram (FECG) vital for the detection of abnormal fetal heart rate (FHR) patterns. Accurate fetal QRS complex detection in FECG is crucial for assessing fetal health, including heart rate and the early identification of congenital conditions. This study presents a novel automated framework using a one-dimensional Convolutional Neural Network (1D-CNN) to detect fetal QRS complexes from abdominal electrocardiogram (AECG) signals, sourced from the PhysioNet Non-Invasive FECG Database (NI-FECGDB). The proposed CNN architecture comprises five convolutional layers, seven batch normalization layers, three dropout layers, and three dense layers. In the initial phase to generate annotations tailored for the proposed network, each AECG signal was partitioned into overlapping 1-second windows, facilitating data augmentation. Binary annotations were generated for each 100-millisecond segment based on the presence or absence of fetal QRS complexes. Subsequently, after performing preprocessing steps, the developed algorithm was applied, and its performance was evaluated using metrics such as accuracy, mean squared error, F1-score, sensitivity, specificity, and precision. The results were compared with those from previous studies utilizing the NI-FECGDB database. The proposed lightweight 1D-CNN architecture demonstrated exceptional performance, achieving 96.79% accuracy, 97.91% sensitivity, 92.79% specificity and 97.88% precision while requiring only 20 AECG signals for training - a significant improvement over existing methods that typically demanded larger datasets. This framework's innovative design eliminated the need for maternal ECG component extraction, thereby reducing computational complexity and potential signal decomposition artifacts. The combination of a simplified five-layer architecture with our novel 100 ms resolution labeling strategy enabled high-precision fetal QRS detection while maintaining minimal preprocessing requirements. These advances position our method as a robust and efficient solution for clinical fetal monitoring applications.