Associate Professor Wichita State University, United States
Introduction: Early detection of fetal abnormalities during pregnancy and delivery is critical to reduce fetal mortality, and the most effective way to accomplish it will be ambulatory fetal electrocardiogram (fECG) monitoring. According to the American Heart Association (AHA) and Centers for Disease Control and Prevention (CDC), congenital heart defects (CHDs) are not only the most common birth defects but also the leading cause of birth deaths [1]. Approx. 1% of infants born in the U.S. have a CHD, but the cause of CHDs is unknown in most cases. In addition to CHDs, abnormal cardiac activities of fetus can be an indicator of severe pregnancy complications such as fetal hypoxia and intrauterine growth restriction [2]. Therefore, there is a great interest in accurate long-term fetal heart activity monitoring to provide proper drug and/or surgical therapies in a timely manner to prevent mortality. However, current gold standard techniques such as a Doppler ultrasound and invasive fetal electrocardiogram (I-fECG) are not suitable for the ambulatory fECG monitoring since they are bulk, not portable, and expensive. Furthermore, they require expertise and skills for safe and successful operations. Wearable-based non-invasive fECG monitoring can be safe and easily accepted by users for daily use, but accurate fECG extraction from abdominal ECG (aECG) signals is very challenging. fECG and maternal ECG (mECG) signals often overlap both in time and frequency domain, and the fetal position changes constantly across pregnant women and the gestation age.
Materials and
Methods: Since there are limited public database of aECG and fECG signals, we have utilized the FECGSYN toolbox to synthesize aECG signals [3]. It helped to generate one minute-long aECG signals at the sampling rate of 1 kHz as mixing stimulated mECG and fECG signals and various noises as shown in Fig. A. As noises, different decibel levels of baseline wander (BW) and muscular artifact (MA) were added to the mixture of mECG and fECG signals. Total 160 of stimulated aECG signals in various conditions of pre-selected 2, 4, and 8 channels went through a number of existing extraction algorithms (e.g., blind sources separation (BSS), template subtraction (TS), and adaptive filtering (AF)) to extract fECG signals from them as shown in Fig. B. R-peak detection was implemented on original fECG and extracted fECG signals to understand the effect of channel number and performance of fECG extraction methods.
Results, Conclusions, and Discussions: Fig. C shows an example of fECG signal extraction from aECG signals. The black curve (upper plot) is an aECG signal, the mixture of mECG and fECG signals. The red curve (middle plot) is an original fECG signal, and the blue curve (bottom plot) is an extracted fECG signal from the eECG signal the by BSS. The green dot lines indicate the success of fECG signal extraction. Fig. D show F1 values of fECG extraction in different BSS variants (FID, FIS, JADE, and PCA) and number of aECG channels. The result indicates the extraction accuracy increases as the number of aECG channels increases. With a higher number of aECG channels, BSS offers higher extraction accuracy regardless of the noise level. Fig. E shows the comparison of performance of extraction algorithms. For BSS, 8 channel’s results were used for the comparison. As shown in Fig. D, the performance of BSS is close to 100% accuracy for all noise levels. For MA noises, the performances of TS and AF algorithms decrease as the noise level increases. For BW noises, the performances of TS and AF algorithms are similar for all noise levels. It may be because the signal distortion by BW noises is moderate. In this study, we were able to simulate aECG signals in various conditions, which can be easily observed in the practical world, and successfully examined the performances of fECG extraction algorithms. Overall, the quality of extracted fECG decreases as the noise level increases. The performance of BSS is less sensitive to the noise level, but it requires more channels, which increases the complexity of a fECG monitoring device. As a future work, aECG signals in different conditions such as fetal movement, heart rate change, and abnormal heartbeat will be examined.
Acknowledgements (Optional): This work was supported by NIH (NICHD R15HD107526) and the Kansas INBRE (P20 GM103418).
