A semi-Supervised Deep Learning Approach to Automate the Identification of Fetal Behavioral States in Fetal Heart Rate Tracings

Computerized Cardiotocography (cCTG) facilitates a thorough and objective examination of the Fetal Heart Rate (FHR), providing valuable insights into the fetal condition and its well-being. A crucial aspect within this context pertains to the automatic identification of periods of fetal activity and quiescence, which are associated with different FHR patterns. The accurate discrimination of these patterns holds the potential to improve the interpretability and diagnostic capabilities of FHR quantitative analysis. Indeed, disruptions in the cycling between active and quiet periods are associated with the development of pathological conditions. This study introduces a deep learning based methodology for the identification of fetal behavioral heart rate patterns. Specifically, the implemented deep neural network (DNN) adopts a 1D encoder-decoder architecture, which is trained to recognize and automatically segment the FHR recordings into active and quiet periods. The proposed framework includes a semi-supervised training process, based on two steps: a) DNN pre-training based on pseudo-labels generated by a Hidden Markov Model (HMM), b) DNN fine-tuning integrating the annotations of an expert Ob-Gyn clinician. The trained DNN exhibits promising results: Balanced Accuracy of 88.37%, Macro F1-Score of 87.87% and Matthews Correlation Coefficient (MCC) of 75.80% on a distinct hold-out test set, encompassing 45 FHR traces annotated by an expert Ob-Gyn clinician.