Active Machine Learning for Pre-procedural Prediction of Time-Varying Boundary Condition After Fontan Procedure Using...

Area of Science:

• Cardiovascular Surgery
• Medical Imaging
• Machine Learning

Background:

• The Fontan procedure is crucial for single ventricle palliation.
• Accurate surgical planning is vital for optimizing Fontan procedure outcomes.
• Pre-operative measurements may not reflect post-operative conditions, impacting planning accuracy.

Purpose of the Study:

• To apply machine learning for predicting post-operative vena caval flow conditions in Fontan patients.
• To develop a framework for accurate surgical planning using predictive models.
• To address the challenge of limited data in cardiovascular studies.

Main Methods:

• Developed a diversity-aware generative adversarial active learning framework.
• Synthesized a virtual cohort using lumped-parameter models.
• Trained deep neural networks on limited pre-operative and post-operative data.

Main Results:

• The proposed framework achieved high prediction accuracy and coefficient of determination.
• Demonstrated success in training predictive models with limited case data.
• Outperformed other methods across 14 experimental combinations of strategies, metrics, and augmentation.

Conclusions:

• The framework enables accurate prediction of post-operative boundary conditions for Fontan surgical planning.
• Represents a significant advancement in deep learning for cardiovascular flow prediction.
• Reduces labeling requirements and expands the learning space for cardiovascular studies.