A Deep Learning Approach for the Assessment of Signal Quality of Non-Invasive Foetal Electrocardiography

Figure A1. Confusion matrix of the CNN-FSQI cross-validation results. It shows the total number of predictions summed over all repetitions of the outer 10-fold CV.

Figure A2. Examples of good- and bad-quality segments. Blue circle: maternal R peak, red cross: foetal R peak.

Figure A3. The full neural network architecture.

Figure 1. Example of NI-FECG. The abdominal mixture (a) contains both a maternal (b) and a foetal (c) component. The foetal component has a lower SNR than the maternal component. The figures depict an ideal case. In practice, the SNR is typically worse.

Figure 2. The NI-FECG electrode configuration used in this work.

Figure 3. Examples of good- and bad-quality segments. Only the first abdominal channel is shown. Blue circle: maternal R peak, red cross: foetal R peak.

Figure 4. Time–frequency representation of an abdominal ECG segment (single-channel) with a good ground truth quality label.

Figure 5. Neural network architecture for NI-FECG signal quality prediction.

Figure 6. Detailed architecture of the CNN path for one channel. The network contains four of these in parallel.

Figure 7. Illustration of the nested cross-validation procedure.

Figure 8. Examples of true and false predictions. Only the first abdominal channel is shown. The ground truth foetal R peaks are labelled with a red cross.

Figure 9. Training graphs for the best (a) and worst (b) repetitions of the CV procedure.