FHR signal analysis using attention-based 1DCNN-BiLSTM neural network for intrapartum fetal monitoring

Abstract

The accurate prediction of fetal hypoxia is crucial in reducing fetal mortality rates. The Cardiotocography (CTG) signal is a widely used tool in fetal monitoring, especially for identifying fetal hypoxia. However, manual CTG analysis presents challenges, leading to a reduced diagnostic rate influenced by subjective factors. Automated CTG analysis emerges as a promising solution to these challenges. Numerous studies have been done on fetal hypoxia detection, but data imbalance poses a hurdle in obtaining the desired results. In response, we propose a novel approach integrating signal denoising through Discrete Wavelet Transform (DWT) based techniques, data balancing using Synthetic Minority Over-sampling Technique (SMOTE), and sliding window-based signal segmentation. Subsequently, an attention-based hybrid 1DCNN-BiLSTM model is employed for fetal hypoxia classification. Our proposed approach achieves impressive results with accuracy, sensitivity, specificity, F1 score, and quality index reaching 93.13%, 93.12%, 94.14%, 93.12%, and 93.53%, respectively. The proposed approach advances fetal hypoxia detection by addressing challenges associated with manual interpretation and data imbalance.