Separating fetal ECG from transabdominal electrical signal: An application of AE-UNet3+

Inhalt:
In the clinical fetal monitoring area, it is difficult to separate fetal electrocardiogram (FECG) from the mixed signals of pregnant women due to the influence of electrode displacement, instrument power frequency interference, maternal electrocardiogram (MECG), and electromyogram (EMG). The MECG overlaps greatly with FECG in the frequency domain, while in the time domain, the QRS groups of MECG and FECG are often aliased, which increases the difficulty of FECG extraction. Through the experiment and improvement of the UNet model, we combine deep learning theory and the template subtraction method to separate FECG signals. A cascade model extraction method based on the AE-UNet3+ model is proposed to successfully isolate FECG components, and the effect of the model is evaluated using simulated and real abdominal electric data. In order to extract FECG from mixed signals, the cascade AE-UNet3+ model is used to suppress MECG to obtain residual error, and then the cascade AE-Unet3+ model is used to separate FECG from residual error. The F1 index of R peak detection accuracy on simulated abdominal electrical data and open real data is 97.24% and 94.74%, respectively. Compared with traditional methods (TS, TS-PCA, EKF, and UKF), we verify the effect of our AE-UNet3+ model by comparing the accuracy of R peak detection, signal-to-noise ratio, and fetal heart rate calculation as evaluation indexes, and objectively evaluating the final output prediction FECG.