Concordance analysis of intrapartum cardiotocography between physicians and artificial intelligence-based technique using modified one-dimensional fully convolutional networks

Figures

• 

  Fig. 1. Study flow chart. EFM = electronic fetal monitoring.

• 

  Fig. 2. Supervised learning on the signal segmentation task. XH (fetal heartbeats sequence) and XU (uterine contraction signal sequence) (darker means higher numbers of units) are input to the neural network model, then the model outputs predicted sequences and (colors represented prediction types and locations) for further comparison with the labeled sequences YH and YU (colors represented event types and locations). The loss is designed to measure the discrepancy between predicted and labeled sequences, and the model parameters are tuned if the loss is large. FHR = fetal heart rate; UC = uterine contractions.

• 

  Fig. 3. Distributions of FHR patterns recognized by physicians and artificial intelligence machine in our study data. Classes 1–6 represent moderate/minimal variability, acceleration, and late/variable/early deceleration, respectively. AI = artificial intelligence.

• 

  Fig. 4. Performance of the artificial intelligence model on recognizing six fetal heart rate patterns. Classes 1–6 represent moderate/minimal variability, acceleration, and late/variable/early deceleration, respectively. AUC = area under the receiver operating characteristic curve; ROC = receiver operating characteristic.

• 

  Fig. 5. Comparisons between clinical practice and FCN on nonreassuring fetal status evaluation. FCN = fully convolutional network.

• 

  Fig. 6. Performance of the artificial intelligence model on nonreassuring fetal status evaluation was positively related to training dataset size. The AUC increased from (A) one-third, (B) two-thirds, and (C) whole training dataset. AUC = area under the receiver operating characteristic curve; ROC = receiver operating characteristic.