An End-to-End RoI-Based Encoder-Decoder for Fetal ECG Recovery and QRS Complex Detection

Although pregnancy monitoring techniques are widespread, the still high number of stillbirths highlights the necessity for more precise and secure approaches. Non-invasive fetal monitoring techniques - such as those based on maternal electrocardiogram (ECG) - allow assessing the fetus's health during pregnancy safely. Many computational methods propose to analyze those signals and recover fetal ECG (fECG) and other important features such as fetal QRS (fQRS) and fetal heart rate. In this work, we propose a multi-channel encoder-decoder-based method coupled with a region of interest (RoI) mask to recover fECG from maternal ECG. Our approach is trained end-to-end in a multi-task manner. Using the RoIs, our method retrieves the fECG with precise QRS complexes and allows direct detection of fQRS peaks. Our method achieves a mean L1-norm error not larger than 0.1223 ± 0.0183 in fECG recovery and up to 96.18 ± 2.09% F1-score for peak QRS detection in popular benchmarks, including cross-dataset assessment. Our results are comparable with state-of-the-art techniques that preprocess the input signals and apply off-the-shelf peak detectors. We evaluate the model in three datasets: ADFECG, NI-FECG, and NiNFEA, showing a good response to a slight domain shift. Also, our model is lightweight and might apply to real-world scenarios without taking special care with preprocessing.