MIRF-Net: A Multimodal Data Fusion Framework for Intrapartum Fetal Risk Assessment.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Obstetrics & Gynecology

Background:

• Accurate intrapartum fetal risk assessment is crucial for optimizing perinatal outcomes and minimizing unnecessary interventions.
• Current deep learning methods often rely solely on cardiotocography (CTG), which has limitations like high false-positive rates and lack of clinical context.

Purpose of the Study:

• To develop and evaluate MIRF-Net, a multimodal deep learning framework for enhanced intrapartum fetal risk assessment.
• To integrate CTG signals, fetal heart rate correlation images (GADF), and maternal metadata for improved risk prediction.

Main Methods:

• MIRF-Net employs a PatchTST encoder for CTG, a ResNet101 for GADF images, and an autoencoder for maternal metadata.
• A fusion Transformer facilitates cross-modal interaction learning for final risk prediction.
• The model was trained and validated on 552 CTG recordings from the CTU-UHB database.

Main Results:

• MIRF-Net achieved a quality index (QI) of 74.76%, AUC of 0.7413, and Brier score of 0.2537 on the test set.
• Performance surpassed representative baseline models, indicating improved fetal risk discrimination and calibrated probabilities.
• Ablation studies confirmed the value of each data modality and the effectiveness of Transformer-based fusion.

Conclusions:

• MIRF-Net offers a robust multimodal framework for intrapartum fetal risk assessment.
• The model demonstrates potential for reliable decision support in intelligent intrapartum monitoring.
• Integrating diverse data sources significantly enhances the accuracy of fetal risk prediction during labor.