Abstract
INTRODUCTION: Fetal heart monitoring (FHR) is a critical tool for assessing fetal health, but traditional methods rely on subjective physician interpretation, exhibiting significant variability that can lead to misdiagnosis and overtreatment. Artificial intelligence (AI) technology offers a novel approach to address this issue, yet existing research predominantly utilizes unimodal (FHR-only) data, failing to align with clinical guidelines emphasizing "bimodality analysis of fetal heart rate and uterine contractions (UC)." This study aims to develop a deep learning-based bimodal intelligent monitoring system to enhance the accuracy and clinical utility of fetal health assessment. METHODS: The research team constructed the first fetal heart-contraction bimodal clinical dataset for Chinese pregnant women (n = 326). Based on the DenseNet121 architecture, a selective attention mechanism (SK module) was introduced, proposing the DenseNet121-SK model. Standardized FHR and UC signals were extracted using image processing techniques. Dense connections and the SK module dynamically fused multi-scale features (e.g., transient fluctuations and contraction cycle associations). The model employed lightweight design during training to enhance physician usability. RESULTS: (1) Dual-modality input significantly outperformed single-modality input, achieving a classification AUC of 0.944 (vs. 0.812 for single-modality), validating the clinical value of multi-parameter collaborative interpretation; (2) The SK module simulated obstetricians' multi-scale cognition, achieving 95.88% accuracy with 100% recall for abnormal cases; (3) The system effectively reduced subjective interpretation variability, providing technical support for minimizing overtreatment. DISCUSSION: This study achieves a balance between clinical interpretability and high performance through lightweight AI design (only 8.3 million parameters) and dual-modality data fusion, making it particularly suitable for resource-constrained primary care settings. Future work should further optimize generalization capabilities through multicenter validation and explore integration with large language models to generate standardized reports. These findings provide important references for optimizing perinatal healthcare resources and AI-assisted decision-making.