Transfer Learning From Simulations on a Reference Anatomy for ECGI in Personalized Cardiac Resynchronization Therapy

Area of Science:

• Cardiovascular Medicine
• Biophysics
• Computational Biology

Background:

• Noninvasive cardiac electrophysiology (EP) model personalization is crucial for predicting cardiac resynchronization therapy (CRT) response.
• Current methods face limitations in clinical applicability due to restricted complexity handling and high computational costs.

Purpose of the Study:

• To develop a novel approach for noninvasive EP model personalization that addresses limitations of existing methods.
• To enable accurate prediction of CRT response in complex cardiac conditions.
• To achieve fast, patient-specific EP predictions with reduced computational burden.

Main Methods:

• Utilized relevance vector regression and shape dimensionality reduction on a large simulated database to analyze complex propagation patterns (multiple onsets, scar tissue).
• Implemented an offline learning strategy on a reference anatomy, transferable to patient-specific anatomies for rapid online predictions.
• Evaluated the method on 20 dyssynchrony patients (120 cardiac cycles) and compared it with electrocardiographic imaging (ECGI).

Main Results:

• The proposed method demonstrated good identification of cardiac activation patterns compared to ECGI.
• Predictions of five different paced patterns for each patient showed good predictive power when compared with body surface potential mappings (BSPM) and ECGI.
• Offline learning from simulated data effectively captured key cardiac EP characteristics for fast, personalized predictions.

Conclusions:

• Offline learning on a reference anatomy from a large simulated database enables accurate, noninvasive cardiac EP characteristic capture.
• This approach facilitates fast, patient-specific predictions, overcoming limitations of current EP modeling techniques.
• The developed method represents a significant advancement for noninvasive CRT patient selection and therapy optimization.