Multimodal Model for Predicting Fetal Acidosis in Delivery Room

Area of Science:

• Perinatal medicine
• Artificial intelligence in healthcare
• Machine learning for medical diagnosis

Background:

• Fetal well-being assessment in the delivery room relies on laboratory tests, cardiotocography (CTG) biosignals, and fetal echocardiography.
• Accurate prediction of fetal acidosis is crucial for timely intervention and improved neonatal outcomes.
• Current methods may have limitations in comprehensively evaluating fetal status.

Purpose of the Study:

• To develop and evaluate a multimodal machine learning model for predicting post-delivery fetal acidosis.
• To integrate diverse data sources including electronic health records, biosignals, and medical imaging.
• To assess the model's predictive performance using a real-world clinical dataset.

Main Methods:

• Feature extraction from unstructured data (biosignals, imaging) and integration with structured medical record data.
• Development of a machine learning classifier using concatenated feature vectors.
• Validation of the model on a dataset of 2,266 deliveries from a tertiary hospital delivery room.

Main Results:

• The multimodal machine learning model achieved an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.752 on the test dataset.
• Demonstrated successful integration of heterogeneous data sources for predictive modeling.
• Indicated the potential for enhanced accuracy in predicting fetal acidosis compared to single-modality approaches.

Conclusions:

• Multimodal machine learning models offer a promising approach for predicting fetal acidosis.
• Integration of medical records, biosignals, and imaging data can improve fetal outcome prediction.
• Further research can explore the application of such models for other critical fetal and neonatal conditions.