Intelligent classification of antenatal cardiotocography signals via multimodal bidirectional gated recurrent units

作者：Fei, Yue; Chen, Fan; He, Lifang; Chen, Jiamin; Hao, Yuexing; Li, Xia; Liu, Guiqing; Chen, Qinqun; Li, Li; Wei, Hang^*

来源：Biomedical Signal Processing and Control, 2022, 78: 104008.

DOI：10.1016/j.bspc.2022.104008

• Multimodal fusion
• Cardiotocography signal classification
• Bidirectional gated recurrent unit
• Antepartum fetal monitoring

Computerized Cardiotocography (cCTG), which involves the continuous recording of the fetal heart rate (FHR) and uterine contraction (UC) signals, plays a critical role in the evaluation of antepartum fetal wellbeing. However, traditional cCTG usually performs the processes of CTG feature extraction and classification separately, which has a certain degree of calibration bias and cannot fully use the CTG information. In this paper, we develop a multimodal bidirectional gated recurrent units (MBiGRU) network for end-to-end CTG feature extraction and classification. Specifically, data preprocessing was first conducted on raw CTG data, including missing values and outliers processing, FHR signal normalization, signal segmentation and enhancement. Afterward, the synchronous FHR, UC, and fetal movement (FetMov) signal fragments were converted into the corresponding two-dimensional fragments through the embedding layers, and then the multimodal fusion of signals was performed in the concatenating layers. Furthermore, classification results were obtained by bidirectional gated recurrent unit (BiGRU) with a fully connected layer and sigmoid function. The effectiveness of the proposed MBiGRU model was tested on 16,355 antenatal CTG records collected from collaborating hospitals with consistent case interpretation by three expert obstetricians. The experimental results of ten-fold cross validation showed that the average accuracy, F1-score, and area under the curve (AUC) values of MBiGRU were 86.45%, 86.14%, and 0.9327, respectively. The MBiGRU improved upon the performance of the baseline BiGRU network with no UC or FetMov signal inputs and outperformed other deep learning methods. In conclusion, the proposed MBiGRU network is promising for intelligent antepartum fetal monitoring.

• 单位
  广州医学院; 广州中医药大学; 1