ETCNN: An ensemble transformer-convolutional neural network for automatic analysis of fetal heart rate

Objective: Traditional methods face challenges in accurately analyzing fetal heart rate (FHR) signals due to the complexity of accelerations and decelerations (Acc/Dec) and their cyclic definition relationship with baseline. We aim to develop a deep learning model, Ensemble Transformer-Convolutional Neural Network (ETCNN), to improve baseline/Acc/Dec determination accuracy and validate its generalization across multi-center and multi- device test datasets. Methods: We proposed ETCNN as a solution, treating FHR analysis as a one-dimensional signal segmentation problem. ETCNN consists of four subnetworks (TCNNs), , each equipped with convolutional kernel size of 21, 31, 61, and 81, respectively. Each subnetwork integrates Channel-Residual (C-Res) modules and Channel Cross fusion with Transformer (CCT) modules. C-Res modules dynamically prune irrelevant channels, focusing on critical FHR episodes, while CCT modules harness multi-scale features to narrow semantic gaps. Results: Trained on Lille Catholic University's open-access database (LCU-DB), ETCNN's performance surpassed twelve traditional methods and three deep learning models across four independent multi-center and multi- device test datasets. Ablation experiments demonstrated the effectiveness of ensemble learning, multi-scale convolution, residual channel attention, channel cross fusion attention, and multi-head attention in improving performance. Conclusion and significance: ETCNN shows promise for accurate and efficient FHR analysis, with successful generalization across various datasets. Its advancements hold potential for clinical applications in fetal monitoring.