This AI Learned From Previous Heart Procedures — And It Just Slashed Pacing Sites by 67%
New continual learning model localizes arrhythmia targets with 81% accuracy using just 4.5 sites
Ventricular tachycardia (VT) is a life-threatening arrhythmia often treated by catheter ablation, which requires clinicians to pace multiple sites in the heart and interpret ECGs to locate the target. Traditional active learning AI models help guide pacing but must retrain from scratch for each VT target, wasting prior data. Researchers led by Dylan O'Hara and Linwei Wang have developed cAPM (Continual AI-Assisted Pace-Mapping), a system that combines a task-agnostic surrogate neural network, active learning for site selection, and continual learning to retain knowledge from past targets across both same-patient and cross-patient scenarios.
Evaluated on an in-silico testbed with sequentially presented localization tasks under varying physiological conditions and ventricular geometries, cAPM achieved an 81% probability of localizing within clinical tolerance (5 mm accuracy) using only 4.5 pacing sites. This dramatically outperforms the previous state-of-the-art active learning method, which required 13.7 pacing sites for only 38% probability. The system's ability to transfer knowledge across multiple VTs could significantly reduce ablation procedure times and improve patient outcomes, and the authors note these results provide a strong basis for moving toward in-vivo preclinical and clinical studies.
- cAPM uses a surrogate neural network that learns pacing site to 12-lead ECG mapping across multiple VT targets
- Achieves 81% localization accuracy (within 5mm) using 4.5 pacing sites vs 38% with 13.7 sites for prior methods
- Continual learning strategy enables knowledge transfer across different VTs in the same patient or across patients without retraining
Why It Matters
Could reduce VT ablation procedure time and improve success rates by requiring fewer pacing sites per target.