Model-Based Feature Augmentation for Cardiac Ablation Target Learning From Images.

Area of Science:

• Biomedical Engineering
• Computational Biology
• Medical Imaging

Background:

• Cardiac radio-frequency ablation (RFA) relies on accurate identification of target tissues.
• Current machine learning approaches for RFA target detection primarily use imaging data alone.
• Improving the performance of these learning algorithms is crucial for effective RFA procedures.

Purpose of the Study:

• To develop and evaluate a model-based feature augmentation scheme to enhance machine learning for cardiac RFA target detection.
• To compare the performance of a learning algorithm using only image features versus one augmented with model-based features.
• To investigate the impact of patient-specific data on the model's predictive capabilities.

Main Methods:

• Computed image features from delayed-enhanced magnetic resonance imaging (DE-MRI) for heterogeneity description.
• Developed a patient-specific image-based model integrated with electrophysiology models to simulate intracardiac electrograms.
• Extracted features from simulated electrograms for model-based feature augmentation of the learning algorithm.

Main Results:

• Achieved average classification scores of 97.2% accuracy, 82.4% sensitivity, and 95.0% positive predictive value with model-based feature augmentation.
• Demonstrated that training the algorithm on the closest patient data improved classification results compared to using all patients.
• Confirmed that model-based feature augmentation significantly strengthens a purely image-driven learning scheme.

Conclusions:

• A feature augmentation scheme based on biophysical cardiac electrophysiology modeling effectively increases prediction scores for RFA target identification.
• Model-based feature augmentation offers a significant performance boost over image-only learning for cardiac ablation target prediction.
• This approach serves as a proof of concept for enhancing diagnostic learning schemes through biophysical modeling.