A deep feature fusion network for fetal state assessment

Area of Science:

• Obstetrics and Gynecology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Cardiotocography (CTG) is a standard tool for diagnosing fetal hypoxia but has limitations in specificity for prenatal acidosis and neurological impairment.
• Early detection of fetal hypoxia during labor is crucial for preventing severe organ damage.
• Current methods may not fully capture complex spatio-temporal patterns in fetal heart rate (FHR) signals.

Purpose of the Study:

• To propose a novel Deep Feature Fusion Network (DFFN) for enhanced fetal state assessment using CTG data.
• To improve the accuracy and specificity of identifying fetal acidosis and neurological impairment.
• To combine the strengths of deep learning and traditional feature engineering for robust fetal monitoring.

Main Methods:

• Extraction of spatial and temporal features from FHR signals using a multiscale Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) network.
• Integration of multiscale CNN-BiLSTM features with frequently used features within a deep learning framework.
• Development of the DFFN model to fuse diverse features for improved classification accuracy.

Main Results:

• The DFFN achieved a sensitivity of 61.97%, specificity of 73.82%, and quality index (QI) of 66.93% on the public CTU-UHB database.
• The proposed method demonstrated superior performance, achieving the highest QI on a private database.
• The DFFN model showed competitive accuracy compared to existing methods in fetal state assessment.

Conclusions:

• The DFFN effectively combines feature engineering and deep learning, enhancing fetal state assessment accuracy.
• The model's ability to extract diverse spatial and temporal information from FHR signals contributes to its improved performance.
• The DFFN shows promise for more accurate and reliable intrapartum fetal monitoring, potentially reducing adverse outcomes.