Fusing Tabular Features and Deep Learning for Fetal Heart Rate Analysis: A Clinically Interpretable Model for Fetal...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Obstetrics and Gynecology

Background:

• Cardiotocography (CTG) is a standard tool for monitoring fetal well-being during labor, assessing fetal heart rate (FHR).
• Current CTG interpretation faces limitations due to low sensitivity and high false-positive rates, impacting its effectiveness in reducing adverse fetal outcomes.
• Accurate prediction of fetal compromise is crucial for timely intervention during labor.

Purpose of the Study:

• To develop an interpretable deep learning model integrating FHR time series and tabular clinical data.
• To enhance the prediction accuracy of fetal compromise, defined by umbilical artery pH < 7.05.
• To improve the reliability and transparency of CTG interpretation in clinical practice.

Main Methods:

• Introduction of Fusion ResNet, a novel deep learning architecture combining convolutional neural networks for FHR analysis and a parallel network for tabular clinical features.
• Training and internal validation on a private dataset of 9,887 FHR recordings.
• External validation on the open-access CTU-UHB dataset (552 recordings) and model interpretability assessment using SHAP and Grad-CAM.

Main Results:

• Fusion ResNet achieved an AUC of 0.77 in internal cross-validation and a state-of-the-art AUC of 0.84 on the external CTU-UHB dataset.
• The model outperformed existing deep learning approaches for fetal compromise prediction.
• SHAP analysis identified key clinical predictors, and Grad-CAM highlighted significant FHR patterns associated with fetal compromise.

Conclusions:

• The developed Fusion ResNet model significantly improves the accuracy of predicting fetal compromise using multimodal data.
• The model's interpretability features provide clinically meaningful insights, enhancing the transparency and trustworthiness of CTG analysis.
• This work highlights the potential of interpretable AI in optimizing fetal monitoring and supporting clinical decision-making in obstetrics.