Transformer-based deep learning for estimating bidirectional maternal-fetal cardiac coupling

Figures

• 

  Figure 1. Illustration of the proposed transformer-based pipeline for bidirectional maternal-fetal cardiac coupling. The right pipeline represents maternal-to-fetal (MTF) coupling, where 10 s of maternal ECG and respiratory signals are processed through convolutional layers and a transformer to estimate fetal HRV features. The left pipeline depicts fetal-to-maternal (FTM) coupling, using 10 s of fetal ECG to predict maternal HRV features through a similar architecture. Here, C denotes channel-wise feature concatenation between upsampled decoder features and the corresponding transformer skip connections.

• 

  Figure 2. Coupling Strength Over Delay: Mean accuracy of MTF and FTM coupling strengths across varying time delays.

• 

  Figure 3. Comparison of ground truth and predicted HRV features for MTF (fetal HRV; a) and FTM (maternal HRV; b) coupling analysis in a representative normal subject. Top panels show predicted and actual HRV features over time, with calculated coupling strength indicating periods of strong and weak coupling. Bland-Altman plots below each figure assess agreement between predicted and actual mean HRV values.

• 

  Figure 4. Comparison of ground truth and predicted HRV features for MTF (fetal HRV; a) and FTM (maternal HRV; b) coupling analysis in an IUGR case. Top panels show predicted and actual HRV features over time, with calculated coupling strength indicating periods of strong and weak coupling. Bland-Altman plots below each figure assess agreement between predicted and actual mean HRV values.

• 

  Figure 5. Mean MTF coupling strength from gestational week 17 to 40. Red circles denote normal pregnancies, while black crosses indicate abnormal cases. Patient IDs for abnormal samples are labeled above each respective point.

• 

  Figure 6. Mean FTM coupling strength from gestational week 17 to 40. Normal cases are marked in red, and abnormal cases are shown as black crosses with corresponding patient IDs annotated.

• 

  Figure 7. Cohort-level regression performance for individual HRV features using mean absolute error (MAE). Each boxplot summarizes subject-level prediction errors for maternal-to-fetal (MTF) and fetal-to-maternal (FTM) coupling directions for both normal and abnormal pregnancies. (a) Mean R-R interval, (b) SDNN, (c) RMSSD, (d) SD1, and (e) SD2.

• 

  Figure 8. Subject-level Bland–Altman analysis for the mean features prediction in the maternal-to-fetal (left) and fetal-to-maternal (right) directions. Each point represents one subject. Abnormal cases are highlighted in red.

• 

  Figure 9. Feature visualizations (top row), corresponding receiver operating characteristic (ROC) curves (middle row), and precision–recall (PR) curves (bottom row) for four classification models under leave-one-subject-out cross-validation. ROC-AUC and PR-AUC values are reported as mean ± standard deviation.

• 

  Figure 10. Negative-control validation of the proposed deep coupling strength (DCS) metric. The distributions of DCS obtained from the true paired data are compared with two negative-control settings: mismatched mother–fetus pairs across different subjects and temporally scrambled surrogate signals, for both the maternal-to-fetal (MTF) and fetal-to-maternal (FTM) directions.