Multimodal Deep Learning for Predicting Adverse Birth Outcomes Based on Early Labour Data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Obstetrics

Background:

• Cardiotocography (CTG) monitors fetal heart rate (FHR) during labor, but visual interpretation is subjective and error-prone.
• Automated CTG analysis methods have not yet improved the detection of abnormal fetal heart rate patterns.
• Accurate fetal health assessment during labor is critical for optimizing birth outcomes.

Purpose of the Study:

• To develop and compare deep learning models for classifying cardiotocography (CTG) signals to identify fetuses with severe compromise at birth.
• To evaluate the effectiveness of 1D-CNN, LSTM, and 2D-CNN based architectures, including a multi-modal approach, using routinely collected FHR data.
• To assess model performance using partial area under the curve (PAUC) and sensitivity at 95% specificity.

Main Methods:

• Utilized a dataset of 51,449 term birth CTG recordings, focusing on the initial 20 minutes of FHR data.
• Developed and compared three 1D-CNN and LSTM-based architectures.
• Transformed FHR signals into 2D time-frequency images (spectrograms, scalograms) for analysis with 2D-CNNs.
• Proposed and evaluated a multi-modal architecture combining 1D-CNN-LSTM and 2D-CNN models in parallel.

Main Results:

• The 1D-CNN-LSTM parallel architecture achieved the best performance among the evaluated models.
• The top-performing model yielded a partial area under the curve (PAUC) of 0.20.
• The model demonstrated a sensitivity of 20% at 95% specificity in detecting severe fetal compromise.

Conclusions:

• The 1D-CNN-LSTM parallel architecture shows promise for automated CTG analysis in identifying fetal compromise.
• Further improvements may be achieved by utilizing larger datasets, analyzing longer FHR traces, and integrating clinical risk factors.
• Advanced AI models can potentially enhance the accuracy and objectivity of fetal well-being assessments during labor.