A robust deep learning framework for automated fetal behavioral state classification: leveraging multi-center datasets to improve antepartum heart rate monitoring

Figures

• 

  Fig. A.1.. FHR signal before and after noise augmentation. Linear interpolations are highlighted in magenta and random outliers with red dots.

• 

  Fig. A.2.. Macro F1-Score in 5 fold cross validation with and without class weighting and batch normalization freezing.

• 

  Fig. A.3.. Scatterplot of Gestational Age versus Macro F1-Score stratified by dataset. No significant correlation can be observed

• 

  Fig. 1.. A fetal heart rate trace recorded via CTG. The overlaid line indicates clinician-scored fetal behavioral states: Active (red), Quiet (blue), and Indeterminate (gray).

• 

  Fig. 2.. Schematic representation of the Residual Attention U-Net architecture with a depth D of 5.

• 

  Fig. 3.. Macro F1-Score (mean and standard error) in the pre-training of the 15 best-performing combinations. Note that at this step the scores are relative to the pseudo-labels generated by the HMM.

• 

  Fig. 4.. Performance evaluation on the hold-out test sets. Error bars show 95% bootstrap percentile confidence intervals (10,000 repetitions), computed by resampling individual traces.

• 

  Fig. 5.. Trace by trace performance evaluation on the hold-out test sets.

• 

  Fig. 6.. Fetal heart rate (FHR) trace alongside annotated behavioral states (top) and the network's predicted probability of the active state (bottom). The annotated states indicate periods of “Quiet” (blue), “Active” (red), and “Indeterminate” (gray). In Panel A) the network's predictions follow the expected trend, with low probabilities of the active state during the quiet phase and a gradual increase leading up to the active phase, reaching a perfect Macro F1-Score of 100%. In panel B), a steep transition from “Active” to “Quiet” can be observed around the 20th minute, characterized by a reduction in FHR variability and a decreasing baseline. The probability of the “Active” state then increases gradually. The zone of disagreement occurs during this transitional phase and the Macro F1-Score is 94%. Panel C) reports the trace with the lowest macro F1 score (i.e., 38%), likely because the trace presents unclear characteristics: frequent small accelerations despite low variability. The FHR signal in panel A and C were sampled from the test set of NAPAMI, while B from PASS.